-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathindex.html
More file actions
1041 lines (989 loc) · 81.8 KB
/
Copy pathindex.html
File metadata and controls
1041 lines (989 loc) · 81.8 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<title>LinStereo: Linear-Complexity Global Attention for Multi-Scale Iterative Stereo Matching</title>
<meta name="description" content="LinStereo — a linear-complexity global-attention (PALA) stereo decoder on a frozen Depth Anything V3 backbone. Competitive on standard benchmarks and zero-shot to new domains, including underwater. ECCV 2026." />
<!-- Open Graph / social -->
<meta property="og:type" content="website" />
<meta property="og:title" content="LinStereo: Linear-Complexity Global Attention for Stereo Matching" />
<meta property="og:description" content="Linear-complexity global attention for iterative stereo on a frozen Depth Anything V3 backbone — competitive on standard benchmarks and strong zero-shot generalization (28% lower AbsRel on TartanAir-UW, 26% on SQUID). ECCV 2026." />
<meta property="og:image" content="static/images/og_card.jpg" />
<meta property="og:image:width" content="1200" />
<meta property="og:image:height" content="630" />
<meta name="twitter:card" content="summary_large_image" />
<!-- Google Scholar / citation metadata -->
<meta name="citation_title" content="LinStereo: Linear-Complexity Global Attention for Multi-Scale Iterative Stereo Matching" />
<meta name="citation_author" content="Wang, Yiran" />
<meta name="citation_author" content="Turner, Oliver" />
<meta name="citation_author" content="Ila, Viorela" />
<meta name="citation_publication_date" content="2026" />
<meta name="citation_conference_title" content="European Conference on Computer Vision (ECCV)" />
<meta name="citation_arxiv_id" content="2606.25437" />
<meta name="citation_pdf_url" content="https://arxiv.org/pdf/2606.25437" />
<meta name="citation_abstract_html_url" content="https://arxiv.org/abs/2606.25437" />
<link rel="icon" type="image/svg+xml" href="static/favicon.svg" />
<link rel="preconnect" href="https://fonts.googleapis.com" />
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin />
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&family=Noto+Sans:ital,wght@0,400;0,500;0,600;1,400&family=Spline+Sans+Mono:wght@400;500&display=swap" rel="stylesheet" />
<link rel="stylesheet" href="static/css/style.css" />
</head>
<body>
<!-- ================= NAVBAR ================= -->
<header class="nav" id="nav">
<div class="nav__inner">
<a class="nav__brand" href="#top" aria-label="LinStereo home">
<svg class="nav__glyph" viewBox="0 0 32 32" aria-hidden="true"><path d="M2 20c4 0 4-4 8-4s4 4 8 4 4-4 8-4M2 12c4 0 4-4 8-4s4 4 8 4 4-4 8-4M2 27c4 0 4-3 8-3s4 3 8 3 4-3 8-3" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round"/></svg>
<span>LinStereo</span>
</a>
<button class="nav__burger" id="navBurger" aria-label="Toggle menu" aria-expanded="false">
<span></span><span></span><span></span>
</button>
<nav class="nav__links" id="navLinks">
<a href="#abstract">Abstract</a>
<a href="#method">Method</a>
<a href="#comparisons">Comparisons</a>
<a href="#dataset">Dataset</a>
<a href="#supplementary">More Results</a>
<a href="#results">Benchmarks</a>
<a href="#bibtex">BibTeX</a>
</nav>
</div>
</header>
<!-- ================= HERO ================= -->
<section class="hero" id="top">
<div class="hero__caustics" aria-hidden="true"></div>
<div class="hero__fade" aria-hidden="true"></div>
<div class="hero__content">
<p class="hero__badge">ECCV 2026</p>
<h1 class="hero__title">LinStereo</h1>
<p class="hero__subtitle">Linear-Complexity Global Attention for Multi-Scale Iterative Stereo Matching</p>
<p class="hero__tagline">Linear-complexity global attention for iterative stereo on a <strong>frozen Depth Anything V3</strong> backbone — strong on standard benchmarks and zero-shot to new domains, including underwater.</p>
<div class="authors">
<span class="author"><a href="#" >Yiran Wang</a><sup>†</sup></span>
<span class="author"><a href="#">Oliver Turner</a></span>
<span class="author"><a href="#">Viorela Ila</a><sup>‡</sup></span>
</div>
<p class="affil">Australian Centre for Robotics, The University of Sydney · <sup>†</sup>Corresponding author · <sup>‡</sup>Project lead</p>
<div class="pills">
<!-- TODO: SeaStereo dataset link still a placeholder (href="#") until release -->
<a class="pill" href="https://arxiv.org/abs/2606.25437" target="_blank" rel="noopener" title="Paper on arXiv (2606.25437)">
<svg viewBox="0 0 24 24" aria-hidden="true"><path d="M6 2h8l4 4v16H6z" fill="none" stroke="currentColor" stroke-width="1.6"/><path d="M14 2v4h4M9 13h6M9 17h6M9 9h2" fill="none" stroke="currentColor" stroke-width="1.6"/></svg>
<span>Paper</span>
</a>
<a class="pill pill--cite" href="#bibtex" title="Jump to the copyable BibTeX citation">
<svg viewBox="0 0 24 24" aria-hidden="true"><path d="M6 3h10a2 2 0 0 1 2 2v16l-6-3-6 3z" fill="none" stroke="currentColor" stroke-width="1.6" stroke-linejoin="round"/><path d="M9 8h6M9 12h5" fill="none" stroke="currentColor" stroke-width="1.6" stroke-linecap="round"/></svg>
<span>BibTeX</span>
</a>
<a class="pill" href="https://github.com/u7079256/LinStereo" target="_blank" rel="noopener" title="Code (GitHub)">
<svg viewBox="0 0 24 24" aria-hidden="true"><path d="M9 19c-5 1.5-5-2.5-7-3m14 6v-3.9a3.4 3.4 0 0 0-.9-2.6c3-.3 6.2-1.5 6.2-6.7A5.2 5.2 0 0 0 19.9 5 4.9 4.9 0 0 0 19.8 1.4S18.7 1 16 2.9a13.4 13.4 0 0 0-7 0C6.3 1 5.2 1.4 5.2 1.4A4.9 4.9 0 0 0 5.1 5a5.2 5.2 0 0 0-1.4 3.6c0 5.2 3.2 6.4 6.2 6.7a3.4 3.4 0 0 0-.9 2.6V22" fill="none" stroke="currentColor" stroke-width="1.6" stroke-linecap="round" stroke-linejoin="round"/></svg>
<span>Code</span>
</a>
<!-- SeaStereo release button hidden until the dataset is public.
<a class="pill" href="#" title="SeaStereo dataset (coming soon)">
<svg viewBox="0 0 24 24" aria-hidden="true"><ellipse cx="12" cy="5" rx="8" ry="3" fill="none" stroke="currentColor" stroke-width="1.6"/><path d="M4 5v6c0 1.7 3.6 3 8 3s8-1.3 8-3V5M4 11v6c0 1.7 3.6 3 8 3s8-1.3 8-3v-6" fill="none" stroke="currentColor" stroke-width="1.6"/></svg>
<span>SeaStereo</span><span class="pill__soon">soon</span>
</a>
-->
</div>
</div>
<div class="hero__wave" aria-hidden="true">
<svg viewBox="0 0 1440 120" preserveAspectRatio="none"><path d="M0 60c180-50 360-50 540 0s360 50 540 0 220-40 360-30v100H0z" fill="#ffffff"/></svg>
</div>
</section>
<!-- ================= TEASER SLIDER ================= -->
<section class="section section--tight" id="teaser">
<div class="container">
<div class="reveal">
<h2 class="eyebrow">Drag to reveal</h2>
<p class="lead">One frozen model, many domains, zero fine-tuning. Drag the handle: the <strong>stereo input</strong>
on one side, LinStereo's <strong>dense disparity</strong> on the other — the <em>same</em> network across
street scenes, indoor and outdoor benchmarks, and underwater footage it has never seen.</p>
<div class="cmp" id="cmpTeaser" data-scene="booster">
<div class="cmp__frame" style="aspect-ratio:1.37">
<img class="cmp__img cmp__img--back" id="cmpBack" src="static/images/hero/booster_depth.jpg" alt="LinStereo predicted disparity (turbo colormap)" />
<div class="cmp__clip" id="cmpClip">
<img class="cmp__img cmp__img--front" id="cmpFront" src="static/images/hero/booster_input.jpg" alt="Stereo input (left view)" />
</div>
<div class="cmp__label cmp__label--left">Input</div>
<div class="cmp__label cmp__label--right">LinStereo disparity</div>
<div class="cmp__handle" id="cmpHandle" role="slider" aria-label="Reveal slider" aria-valuemin="0" aria-valuemax="100" aria-valuenow="50" tabindex="0">
<span class="cmp__handle-grip"><svg viewBox="0 0 24 24" aria-hidden="true"><path d="M9 6l-4 6 4 6M15 6l4 6-4 6" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/></svg></span>
</div>
</div>
</div>
<div class="tabs" id="teaserTabs" role="tablist" aria-label="Dataset">
<button class="tab is-active" data-scene="booster" role="tab" aria-selected="true">Booster</button>
<button class="tab" data-scene="kitti" role="tab" aria-selected="false">KITTI</button>
<button class="tab" data-scene="eth3d" role="tab" aria-selected="false">ETH3D</button>
<button class="tab" data-scene="tartanair" role="tab" aria-selected="false">TartanAir-UW</button>
<button class="tab" data-scene="squid" role="tab" aria-selected="false">SQUID</button>
<button class="tab" data-scene="seastereo" role="tab" aria-selected="false">SeaStereo</button>
</div>
<p class="caption">A single LinStereo model — frozen <strong>Depth Anything V3</strong> backbone, trained on
<strong>SceneFlow only</strong> — runs zero-shot across all six datasets, from KITTI street scenes to real
underwater SQUID footage it was never trained on. Black = no ground truth.</p>
</div>
</div>
</section>
<!-- ================= TL;DR ================= -->
<section class="tldr">
<div class="container">
<div class="tldr__card reveal">
<svg class="tldr__bolt" viewBox="0 0 24 24" aria-hidden="true"><path d="M13 2L4 14h6l-1 8 9-12h-6z" fill="currentColor"/></svg>
<p><strong>TL;DR.</strong> LinStereo redesigns the iterative stereo update loop around <strong>Position-Aware Linear Attention (PALA)</strong> — a global, <em>O(N)</em> operator on a frozen Depth Anything V3 backbone — so reliable depth propagates across the whole image in a single step. It generalizes zero-shot to underwater scenes where local photometric cues collapse, while staying linear per iteration.</p>
</div>
</div>
</section>
<!-- ================= NEWS ================= -->
<section class="news" id="news">
<div class="container">
<div class="news__card reveal">
<h2 class="news__heading">News</h2>
<ul class="news__list">
<!-- The arXiv news item is intentionally hidden.
<li>
<span class="news__date">Jun 2026</span>
<span>📄 Preprint is live on arXiv — <a href="https://arxiv.org/abs/2606.25437" target="_blank" rel="noopener">arXiv:2606.25437</a>.</span>
</li>
-->
<li>
<span class="news__date">Jun 2026</span>
<span>🎉 LinStereo is accepted to <strong>ECCV 2026</strong>!</span>
</li>
</ul>
</div>
</div>
</section>
<!-- ================= HIGHLIGHTS ================= -->
<section class="section section--soft" id="highlights">
<div class="container">
<div class="stats">
<div class="stat reveal">
<div class="stat__num">28%<span>↓</span></div>
<div class="stat__label">AbsRel — TartanAir-UW ocean</div>
</div>
<div class="stat reveal">
<div class="stat__num">26%<span>↓</span></div>
<div class="stat__label">AbsRel — SQUID (real underwater)</div>
</div>
<div class="stat reveal">
<div class="stat__num">31%<span>↓</span></div>
<div class="stat__label">RMSE vs FoundationStereo (TartanAir-UW)</div>
</div>
<div class="stat reveal">
<div class="stat__num">37%<span>↓</span></div>
<div class="stat__label">Occluded EPE — Middlebury(H)</div>
</div>
</div>
<p class="stats__note reveal">Competitive with ViT-L systems on standard benchmarks and best on underwater generalization — all from a
<strong>ViT-B</strong> backbone trained on <strong>SceneFlow only</strong>, with no underwater, real-world, or domain-specific data at any stage.</p>
</div>
</section>
<!-- ================= ABSTRACT ================= -->
<section class="section" id="abstract">
<div class="container container--narrow">
<h2 class="section__title">Abstract</h2>
<p class="abstract reveal">
Existing Vision Foundation Model (VFM)-based iterative stereo pipelines under-exploit three information
pathways: multi-scale backbone features are collapsed into single-level correlations, geometric priors
remain untapped at initialization, and context propagates only locally. These gaps widen under degraded
photometric cues, making underwater scenes a stringent generalization test. To address this, we propose
<strong>LinStereo</strong>, built upon Depth Anything V3, whose core is a
<strong>Position-Aware Linear Attention (PALA)</strong> module that replaces local recurrence with global
aggregation at linear cost, propagating reliable estimates from well-matched regions into degraded areas
while preserving disparity structure. PALA is made effective by two enabling components:
<strong>Hierarchical Semantic Cost Volumes (HSCV)</strong>, which supply scale-aligned correlations from
the VFM feature hierarchy, and a <strong>Depth Prior Initialization (DPI)</strong> that converts monocular
depth into a metrically calibrated warm start. LinStereo achieves state-of-the-art performance on standard
benchmarks and strong cross-domain generalization, particularly on underwater scenes where severe
photometric degradation makes stereo matching particularly challenging, surpassing all compared methods with
consistent gains (<strong>28%</strong> AbsRel on TartanAir-UW, <strong>26%</strong> on SQUID, a
real-world underwater dataset).
</p>
</div>
</section>
<!-- ================= METHOD ================= -->
<section class="section section--soft" id="method">
<div class="container">
<h2 class="section__title">Method</h2>
<p class="lead lead--center reveal">
Modern VFM backbones encode rich multi-scale semantics and implicit geometry, but the iterative update loop
still consumes them through a narrow interface inherited from lightweight encoders. LinStereo closes this
gap with a single <strong>PALA-centred decoder</strong>, enabled by HSCV and DPI — not three independent modules.
</p>
<figure class="figure reveal">
<img src="static/images/architecture.png" alt="LinStereo architecture: frozen Depth Anything V3 backbone feeding Hierarchical Semantic Cost Volumes, Depth Prior Initialization, and the Position-Aware Linear Attention updater." loading="lazy" />
<figcaption>
<strong>Architecture.</strong> A frozen Depth Anything V3 backbone
produces, in a single forward pass, multi-scale stereo features and a monocular depth map. The PALA updater
replaces the ConvGRU with global linear attention over cost-volume features, using 2D rotary position
encoding and an adaptive gating branch.
</figcaption>
</figure>
<div class="pathways">
<div class="pathway reveal">
<span class="pathway__tag">WHAT to match</span>
<h3>HSCV</h3>
<p>Hierarchical Semantic Cost Volumes build a per-scale correlation volume from each VFM scale
(s ∈ {4,8,16}), kept simultaneously queryable, each with its own 4-level disparity pyramid.</p>
</div>
<div class="pathway reveal">
<span class="pathway__tag">WHERE to start</span>
<h3>DPI</h3>
<p>Depth Prior Initialization turns DA3's affine-invariant monocular depth into a metric disparity warm start
via SIFT-based scale–shift alignment — refinement no longer starts from zero.</p>
</div>
<div class="pathway reveal">
<span class="pathway__tag">HOW to refine</span>
<h3>PALA</h3>
<p>Position-Aware Linear Attention aggregates global context at <em>O(N)</em> per iteration, propagating
reliable disparity from textured regions into degraded, textureless areas.</p>
</div>
</div>
</div>
</section>
<!-- ================= COMPONENTS DEEP DIVE ================= -->
<section class="section" id="components">
<div class="container">
<article class="component reveal">
<div class="component__body">
<span class="component__index">01</span>
<h3 class="component__name">Position-Aware Linear Attention <span>(PALA)</span></h3>
<p>Standard linear attention reaches <em>O(N)</em> by precomputing <span class="mono">KᵀV</span> and sharing
it across queries — but that collapses all spatial relationships into one global summary, making attention
<em>position-agnostic</em>. PALA restores spatial structure with 2D rotary position encoding applied
<strong>asymmetrically</strong>: only the numerator is position-augmented, while the denominator keeps the
plain kernels for stable normalization.</p>
<div class="eq">
O<sub>attn</sub> = Q̃ (K̃<sup>⊤</sup>V) · ( Q̂ K̂<sup>⊤</sup> + ε )<sup>−1</sup>
</div>
<p>A local spatial encoding on the value branch and an adaptive gate let each location absorb new evidence where
matching is confident and preserve its estimate where it is not — so reliable disparity gradually fills
uncertain regions. Cost stays <span class="mono">O(N·C²)</span>; one PALA iteration is as cheap as a
ConvGRU step (3.50 ms vs 3.63 ms).</p>
</div>
</article>
<article class="component component--alt reveal">
<div class="component__body">
<span class="component__index">02</span>
<h3 class="component__name">Hierarchical Semantic Cost Volumes <span>(HSCV)</span></h3>
<p>Instead of one fixed-resolution correlation volume, HSCV builds one per VFM scale and keeps the whole set
accessible, so PALA can re-query any scale at every iteration. Each volume carries its own disparity-axis
pyramid for coarse-to-fine matching while preserving spatial resolution — a two-level hierarchy across
scales and within each scale.</p>
<p class="component__hook">Replacing the hierarchy with a single pooled feature degrades EPE
<span class="delta">1.01→1.07</span>.</p>
</div>
</article>
<article class="component reveal">
<div class="component__body">
<span class="component__index">03</span>
<h3 class="component__name">Depth Prior Initialization <span>(DPI)</span></h3>
<p>Disparity is inversely proportional to depth, so DPI converts the monocular prediction via
<span class="mono">d = α/D<sub>mono</sub> + β</span>, solving the scale–shift
<span class="mono">(α,β)</span> by least squares from sparse SIFT correspondences along epipolar
lines. Refinement starts from a geometrically plausible state rather than zero.</p>
<p class="component__hook">Honest fallback: when inliers < 20 it reverts to zero-disparity init —
firing on <span class="delta">0%</span> of TartanAir-UW and <span class="delta">3.7%</span> of SQUID frames,
costing only <span class="delta">+0.08 px</span> EPE.</p>
</div>
</article>
</div>
</section>
<!-- ================= COMPARISONS (qualitative) ================= -->
<section class="section section--soft" id="comparisons">
<div class="container">
<h2 class="section__title">Qualitative comparison</h2>
<p class="lead lead--center reveal">On well-conditioned scenes recent methods look near-identical; underwater,
the differences explode. The win parses in under a second.</p>
<figure class="figure reveal">
<img src="static/images/qual_underwater.jpg" alt="Underwater qualitative comparison grid: left/right images, ground truth, FoundationStereo, MGStereo, DeFOM-Stereo, Stereo Anywhere, and Ours, on TartanAir-UW and SQUID." loading="lazy" />
<figcaption><strong>Underwater (TartanAir-UW & SQUID).</strong> Columns: input, ground truth, FoundationStereo,
MGStereo, DeFOM-Stereo, Stereo Anywhere, and <strong>Ours</strong>. Baselines over-estimate depth in distant,
backscatter-heavy regions and waver near-field; LinStereo keeps consistent scale at both ranges. Black = no GT.</figcaption>
</figure>
<figure class="figure reveal">
<img src="static/images/qual_standard.jpg" alt="Standard-benchmark qualitative comparison on Booster and ETH3D." loading="lazy" />
<figcaption><strong>Standard benchmarks (Booster Q & ETH3D).</strong> LinStereo recovers sharper thin structures
and handles non-Lambertian surfaces (transparent bottles, car windows).</figcaption>
</figure>
</div>
</section>
<!-- ================= SUPPLEMENTARY ================= -->
<section class="section section--soft" id="supplementary">
<div class="container container--wide">
<h2 class="section__title">More qualitative results</h2>
<p class="lead lead--center reveal">Per-benchmark comparisons, <strong>one scene per row</strong>: the input
alongside recent baselines and <strong>LinStereo (Ours, highlighted)</strong> — across standard,
underwater, real-world, and synthetic domains.</p>
<div class="gallery__group reveal">
<h3 class="subhead">Standard benchmarks</h3>
<figure class="figure panelblock reveal">
<h4 class="panelblock__name">Booster</h4>
<div class="panelset" style="--cols:6">
<div class="panelset__head">
<span>Input</span>
<span>IGEV++</span>
<span>MGStereo</span>
<span>DeFOM-Stereo</span>
<span>Stereo Anywhere</span>
<span class="is-ours">Ours</span>
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/booster_s0_p0.jpg" alt="Booster Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/booster_s0_p1.jpg" alt="Booster IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/booster_s0_p2.jpg" alt="Booster MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/booster_s0_p3.jpg" alt="Booster DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/booster_s0_p4.jpg" alt="Booster Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/booster_s0_p5.jpg" alt="Booster Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/booster_s1_p0.jpg" alt="Booster Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/booster_s1_p1.jpg" alt="Booster IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/booster_s1_p2.jpg" alt="Booster MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/booster_s1_p3.jpg" alt="Booster DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/booster_s1_p4.jpg" alt="Booster Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/booster_s1_p5.jpg" alt="Booster Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/booster_s2_p0.jpg" alt="Booster Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/booster_s2_p1.jpg" alt="Booster IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/booster_s2_p2.jpg" alt="Booster MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/booster_s2_p3.jpg" alt="Booster DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/booster_s2_p4.jpg" alt="Booster Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/booster_s2_p5.jpg" alt="Booster Ours" loading="lazy">
</div>
</div>
<figcaption><strong>Booster (Q).</strong> Transparent glass bottles, a specular metallic sink, and reflective appliances — the target depth is the physical surface, not what is seen through the glass. Baselines bleed depth around glass and leave noise on specular metal; LinStereo recovers clean contours and consistent surface depth (red boxes).</figcaption>
</figure>
<figure class="figure panelblock reveal">
<h4 class="panelblock__name">ETH3D</h4>
<div class="panelset" style="--cols:6">
<div class="panelset__head">
<span>Input</span>
<span>IGEV++</span>
<span>MGStereo</span>
<span>DeFOM-Stereo</span>
<span>Stereo Anywhere</span>
<span class="is-ours">Ours</span>
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/eth3d_s0_p0.jpg" alt="ETH3D Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/eth3d_s0_p1.jpg" alt="ETH3D IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/eth3d_s0_p2.jpg" alt="ETH3D MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/eth3d_s0_p3.jpg" alt="ETH3D DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/eth3d_s0_p4.jpg" alt="ETH3D Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/eth3d_s0_p5.jpg" alt="ETH3D Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/eth3d_s1_p0.jpg" alt="ETH3D Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/eth3d_s1_p1.jpg" alt="ETH3D IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/eth3d_s1_p2.jpg" alt="ETH3D MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/eth3d_s1_p3.jpg" alt="ETH3D DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/eth3d_s1_p4.jpg" alt="ETH3D Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/eth3d_s1_p5.jpg" alt="ETH3D Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/eth3d_s2_p0.jpg" alt="ETH3D Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/eth3d_s2_p1.jpg" alt="ETH3D IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/eth3d_s2_p2.jpg" alt="ETH3D MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/eth3d_s2_p3.jpg" alt="ETH3D DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/eth3d_s2_p4.jpg" alt="ETH3D Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/eth3d_s2_p5.jpg" alt="ETH3D Ours" loading="lazy">
</div>
</div>
<figcaption><strong>ETH3D.</strong> Industrial interiors with thin pipes, low-light building exteriors, and dense outdoor vegetation. Baselines lose thin-structure continuity and produce large-area errors in dark regions; LinStereo preserves fine detail and stays stable under low light.</figcaption>
</figure>
<figure class="figure panelblock reveal">
<h4 class="panelblock__name">KITTI’12</h4>
<div class="panelset" style="--cols:6">
<div class="panelset__head">
<span>Input</span>
<span>FoundationStereo</span>
<span>MGStereo</span>
<span>DeFOM-Stereo</span>
<span>Stereo Anywhere</span>
<span class="is-ours">Ours</span>
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/kitti2012_s0_p0.jpg" alt="KITTI 2012 Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s0_p1.jpg" alt="KITTI 2012 FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s0_p2.jpg" alt="KITTI 2012 MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s0_p3.jpg" alt="KITTI 2012 DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s0_p4.jpg" alt="KITTI 2012 Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/kitti2012_s0_p5.jpg" alt="KITTI 2012 Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/kitti2012_s1_p0.jpg" alt="KITTI 2012 Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s1_p1.jpg" alt="KITTI 2012 FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s1_p2.jpg" alt="KITTI 2012 MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s1_p3.jpg" alt="KITTI 2012 DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s1_p4.jpg" alt="KITTI 2012 Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/kitti2012_s1_p5.jpg" alt="KITTI 2012 Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/kitti2012_s2_p0.jpg" alt="KITTI 2012 Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s2_p1.jpg" alt="KITTI 2012 FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s2_p2.jpg" alt="KITTI 2012 MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s2_p3.jpg" alt="KITTI 2012 DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s2_p4.jpg" alt="KITTI 2012 Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/kitti2012_s2_p5.jpg" alt="KITTI 2012 Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/kitti2012_s3_p0.jpg" alt="KITTI 2012 Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s3_p1.jpg" alt="KITTI 2012 FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s3_p2.jpg" alt="KITTI 2012 MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s3_p3.jpg" alt="KITTI 2012 DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2012_s3_p4.jpg" alt="KITTI 2012 Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/kitti2012_s3_p5.jpg" alt="KITTI 2012 Ours" loading="lazy">
</div>
</div>
<figcaption><strong>KITTI 2012.</strong> Four street scenes. LinStereo yields sharper vehicle boundaries and preserves thin structures — poles, traffic signs — that competing methods blur or fragment (red boxes).</figcaption>
</figure>
<figure class="figure panelblock reveal">
<h4 class="panelblock__name">KITTI’15</h4>
<div class="panelset" style="--cols:6">
<div class="panelset__head">
<span>Input</span>
<span>IGEV++</span>
<span>MGStereo</span>
<span>DeFOM-Stereo</span>
<span>Stereo Anywhere</span>
<span class="is-ours">Ours</span>
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/kitti2015_s0_p0.jpg" alt="KITTI 2015 Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s0_p1.jpg" alt="KITTI 2015 IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s0_p2.jpg" alt="KITTI 2015 MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s0_p3.jpg" alt="KITTI 2015 DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s0_p4.jpg" alt="KITTI 2015 Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/kitti2015_s0_p5.jpg" alt="KITTI 2015 Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/kitti2015_s1_p0.jpg" alt="KITTI 2015 Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s1_p1.jpg" alt="KITTI 2015 IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s1_p2.jpg" alt="KITTI 2015 MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s1_p3.jpg" alt="KITTI 2015 DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s1_p4.jpg" alt="KITTI 2015 Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/kitti2015_s1_p5.jpg" alt="KITTI 2015 Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/kitti2015_s2_p0.jpg" alt="KITTI 2015 Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s2_p1.jpg" alt="KITTI 2015 IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s2_p2.jpg" alt="KITTI 2015 MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s2_p3.jpg" alt="KITTI 2015 DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s2_p4.jpg" alt="KITTI 2015 Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/kitti2015_s2_p5.jpg" alt="KITTI 2015 Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/kitti2015_s3_p0.jpg" alt="KITTI 2015 Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s3_p1.jpg" alt="KITTI 2015 IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s3_p2.jpg" alt="KITTI 2015 MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s3_p3.jpg" alt="KITTI 2015 DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/kitti2015_s3_p4.jpg" alt="KITTI 2015 Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/kitti2015_s3_p5.jpg" alt="KITTI 2015 Ours" loading="lazy">
</div>
</div>
<figcaption><strong>KITTI 2015.</strong> LinStereo recovers small distant objects and road-level structure such as guard rails that the baselines over-smooth or miss (red boxes).</figcaption>
</figure>
</div>
<div class="gallery__group reveal">
<h3 class="subhead">Underwater, real-world & synthetic</h3>
<figure class="figure panelblock reveal">
<h4 class="panelblock__name">TartanAir-UW ①</h4>
<div class="panelset" style="--cols:6">
<div class="panelset__head">
<span>Input</span>
<span>FoundationStereo</span>
<span>MGStereo</span>
<span>DeFOM-Stereo</span>
<span>Stereo Anywhere</span>
<span class="is-ours">Ours</span>
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/tartanair1_s0_p0.jpg" alt="TartanAir-UW Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair1_s0_p1.jpg" alt="TartanAir-UW FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair1_s0_p2.jpg" alt="TartanAir-UW MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair1_s0_p3.jpg" alt="TartanAir-UW DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair1_s0_p4.jpg" alt="TartanAir-UW Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/tartanair1_s0_p5.jpg" alt="TartanAir-UW Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/tartanair1_s1_p0.jpg" alt="TartanAir-UW Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair1_s1_p1.jpg" alt="TartanAir-UW FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair1_s1_p2.jpg" alt="TartanAir-UW MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair1_s1_p3.jpg" alt="TartanAir-UW DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair1_s1_p4.jpg" alt="TartanAir-UW Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/tartanair1_s1_p5.jpg" alt="TartanAir-UW Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/tartanair1_s2_p0.jpg" alt="TartanAir-UW Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair1_s2_p1.jpg" alt="TartanAir-UW FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair1_s2_p2.jpg" alt="TartanAir-UW MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair1_s2_p3.jpg" alt="TartanAir-UW DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair1_s2_p4.jpg" alt="TartanAir-UW Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/tartanair1_s2_p5.jpg" alt="TartanAir-UW Ours" loading="lazy">
</div>
</div>
<figcaption><strong>TartanAir-UW ①.</strong> Simulated underwater scenes (rocky seafloor, an underwater cliff, a coral reef). Baselines overestimate depth in backscatter-affected distant regions and lose small foreground objects; LinStereo preserves accurate near-to-far transitions and resolves individual object depth.</figcaption>
</figure>
<figure class="figure panelblock reveal">
<h4 class="panelblock__name">TartanAir-UW ②</h4>
<div class="panelset" style="--cols:6">
<div class="panelset__head">
<span>Input</span>
<span>FoundationStereo</span>
<span>MGStereo</span>
<span>DeFOM-Stereo</span>
<span>Stereo Anywhere</span>
<span class="is-ours">Ours</span>
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/tartanair2_s0_p0.jpg" alt="TartanAir-UW Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair2_s0_p1.jpg" alt="TartanAir-UW FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair2_s0_p2.jpg" alt="TartanAir-UW MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair2_s0_p3.jpg" alt="TartanAir-UW DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair2_s0_p4.jpg" alt="TartanAir-UW Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/tartanair2_s0_p5.jpg" alt="TartanAir-UW Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/tartanair2_s1_p0.jpg" alt="TartanAir-UW Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair2_s1_p1.jpg" alt="TartanAir-UW FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair2_s1_p2.jpg" alt="TartanAir-UW MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair2_s1_p3.jpg" alt="TartanAir-UW DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair2_s1_p4.jpg" alt="TartanAir-UW Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/tartanair2_s1_p5.jpg" alt="TartanAir-UW Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/tartanair2_s2_p0.jpg" alt="TartanAir-UW Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair2_s2_p1.jpg" alt="TartanAir-UW FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair2_s2_p2.jpg" alt="TartanAir-UW MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair2_s2_p3.jpg" alt="TartanAir-UW DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/tartanair2_s2_p4.jpg" alt="TartanAir-UW Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/tartanair2_s2_p5.jpg" alt="TartanAir-UW Ours" loading="lazy">
</div>
</div>
<figcaption><strong>TartanAir-UW ②.</strong> Severe scattering and low visibility. Baselines show depth collapse, blocky artifacts, or over-smoothing that erases terrain; LinStereo recovers coherent depth gradients and detects small foreground objects in extremely degraded conditions.</figcaption>
</figure>
<figure class="figure panelblock reveal">
<h4 class="panelblock__name">SQUID</h4>
<div class="panelset" style="--cols:6">
<div class="panelset__head">
<span>Input</span>
<span>FoundationStereo</span>
<span>MGStereo</span>
<span>DeFOM-Stereo</span>
<span>Stereo Anywhere</span>
<span class="is-ours">Ours</span>
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/squid_s0_p0.jpg" alt="SQUID Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/squid_s0_p1.jpg" alt="SQUID FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/squid_s0_p2.jpg" alt="SQUID MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/squid_s0_p3.jpg" alt="SQUID DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/squid_s0_p4.jpg" alt="SQUID Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/squid_s0_p5.jpg" alt="SQUID Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/squid_s1_p0.jpg" alt="SQUID Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/squid_s1_p1.jpg" alt="SQUID FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/squid_s1_p2.jpg" alt="SQUID MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/squid_s1_p3.jpg" alt="SQUID DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/squid_s1_p4.jpg" alt="SQUID Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/squid_s1_p5.jpg" alt="SQUID Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/squid_s2_p0.jpg" alt="SQUID Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/squid_s2_p1.jpg" alt="SQUID FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/squid_s2_p2.jpg" alt="SQUID MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/squid_s2_p3.jpg" alt="SQUID DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/squid_s2_p4.jpg" alt="SQUID Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/squid_s2_p5.jpg" alt="SQUID Ours" loading="lazy">
</div>
</div>
<figcaption><strong>SQUID (real-world).</strong> Shallow sandy seafloor, near-field coral and rock, man-made equipment. Under strong color attenuation the baselines show depth-scale inconsistencies and bleeding; LinStereo keeps stable depth and recovers fine detail. Black = unavailable depth.</figcaption>
</figure>
<figure class="figure panelblock reveal">
<h4 class="panelblock__name">Real-world tank</h4>
<div class="panelset" style="--cols:9">
<div class="panelset__head">
<span>Input</span>
<span>Right</span>
<span>GT</span>
<span>IGEV++</span>
<span>MGStereo</span>
<span>FoundationStereo</span>
<span>DeFOM-Stereo</span>
<span>Stereo Anywhere</span>
<span class="is-ours">Ours</span>
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/realworld_s0_p0.jpg" alt="Real-world tank Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s0_p1.jpg" alt="Real-world tank Right" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s0_p2.jpg" alt="Real-world tank GT" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s0_p3.jpg" alt="Real-world tank IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s0_p4.jpg" alt="Real-world tank MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s0_p5.jpg" alt="Real-world tank FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s0_p6.jpg" alt="Real-world tank DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s0_p7.jpg" alt="Real-world tank Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/realworld_s0_p8.jpg" alt="Real-world tank Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/realworld_s1_p0.jpg" alt="Real-world tank Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s1_p1.jpg" alt="Real-world tank Right" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s1_p2.jpg" alt="Real-world tank GT" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s1_p3.jpg" alt="Real-world tank IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s1_p4.jpg" alt="Real-world tank MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s1_p5.jpg" alt="Real-world tank FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s1_p6.jpg" alt="Real-world tank DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/realworld_s1_p7.jpg" alt="Real-world tank Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/realworld_s1_p8.jpg" alt="Real-world tank Ours" loading="lazy">
</div>
</div>
<figcaption><strong>Real-world lab tank — close-up.</strong> A controlled water tank imaged at close range (< 2 m) with AprilTag + CAD-model ground truth down to sub-millimetre, including ~3 mm taut ropes as a fine-structure stress test. Columns: left, right, GT, then five baselines and <strong>Ours</strong>. LinStereo resolves the thin ropes and tank geometry that baselines blur, fragment, or miss. Black = no valid GT.</figcaption>
</figure>
<figure class="figure panelblock reveal">
<h4 class="panelblock__name">SeaStereo</h4>
<div class="panelset" style="--cols:9">
<div class="panelset__head">
<span>Input</span>
<span>Right</span>
<span>GT</span>
<span>IGEV++</span>
<span>MGStereo</span>
<span>FoundationStereo</span>
<span>DeFOM-Stereo</span>
<span>Stereo Anywhere</span>
<span class="is-ours">Ours</span>
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/seastereo_s0_p0.jpg" alt="SeaStereo Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s0_p1.jpg" alt="SeaStereo Right" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s0_p2.jpg" alt="SeaStereo GT" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s0_p3.jpg" alt="SeaStereo IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s0_p4.jpg" alt="SeaStereo MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s0_p5.jpg" alt="SeaStereo FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s0_p6.jpg" alt="SeaStereo DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s0_p7.jpg" alt="SeaStereo Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/seastereo_s0_p8.jpg" alt="SeaStereo Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/seastereo_s1_p0.jpg" alt="SeaStereo Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s1_p1.jpg" alt="SeaStereo Right" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s1_p2.jpg" alt="SeaStereo GT" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s1_p3.jpg" alt="SeaStereo IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s1_p4.jpg" alt="SeaStereo MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s1_p5.jpg" alt="SeaStereo FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s1_p6.jpg" alt="SeaStereo DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo_s1_p7.jpg" alt="SeaStereo Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/seastereo_s1_p8.jpg" alt="SeaStereo Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/seastereo2_s0_p0.jpg" alt="SeaStereo Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s0_p1.jpg" alt="SeaStereo Right" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s0_p2.jpg" alt="SeaStereo GT" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s0_p3.jpg" alt="SeaStereo IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s0_p4.jpg" alt="SeaStereo MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s0_p5.jpg" alt="SeaStereo FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s0_p6.jpg" alt="SeaStereo DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s0_p7.jpg" alt="SeaStereo Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/seastereo2_s0_p8.jpg" alt="SeaStereo Ours" loading="lazy">
</div>
<div class="panelrow">
<img class="panelrow__img" src="static/images/panels/seastereo2_s1_p0.jpg" alt="SeaStereo Input" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s1_p1.jpg" alt="SeaStereo Right" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s1_p2.jpg" alt="SeaStereo GT" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s1_p3.jpg" alt="SeaStereo IGEV++" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s1_p4.jpg" alt="SeaStereo MGStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s1_p5.jpg" alt="SeaStereo FoundationStereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s1_p6.jpg" alt="SeaStereo DeFOM-Stereo" loading="lazy">
<img class="panelrow__img" src="static/images/panels/seastereo2_s1_p7.jpg" alt="SeaStereo Stereo Anywhere" loading="lazy">
<img class="panelrow__img is-ours" src="static/images/panels/seastereo2_s1_p8.jpg" alt="SeaStereo Ours" loading="lazy">
</div>
</div>
<figcaption><strong>SeaStereo (synthetic).</strong> Left, right, ground truth, then five baselines and <strong>Ours</strong>, across Jerlov water types. Baselines show edge noise or shape distortion on elongated structures and over-smooth vegetation, losing depth layering; LinStereo preserves sharp object contours and foreground–background separation.</figcaption>
</figure>
</div>
<p class="caption reveal">Reproduced from the paper's supplementary material. See the paper for full discussion.</p>
</div>
</section>
<!-- ================= DATASET ================= -->
<section class="section section--soft" id="dataset">
<div class="container">
<h2 class="section__title">SeaStereo Dataset</h2>
<p class="lead lead--center reveal">A physically-rendered underwater stereo corpus with dense disparity ground truth,
released with the paper.</p>
<div class="dstats reveal">
<div><span>~40K</span>stereo pairs</div>
<div><span>7</span>Jerlov water types</div>
<div><span>1000+</span>configurations</div>
<div><span>dense</span>GT disparity</div>
</div>
<figure class="figure reveal">
<img src="static/images/dataset_pipeline.png" alt="SeaStereo rendering pipeline: ShapeNet objects composited over marine backgrounds in Blender under Jerlov water models." loading="lazy" />
<figcaption><strong>Rendering pipeline.</strong> ShapeNetCore foreground objects composited over real marine backgrounds
(coral, fish, shipwrecks) and rendered in Blender under varying Jerlov water types — spanning 1000+ configurations
with varied trajectories, focal lengths, inter-ocular distances, and seafloor depths.</figcaption>
</figure>
<h3 class="subhead reveal">More dataset samples</h3>
<p class="caption reveal">Each card shows the <strong>left</strong> & <strong>right</strong> stereo views and the dense <strong>disparity</strong>
ground truth, spanning Jerlov water types I–3C, seafloor depths, and camera configurations.</p>
<div class="montage reveal">
<div class="montage__card"><div class="montage__hdr"><span>Left</span><span>Right</span><span>Disp.</span></div><div class="montage__trip"><img src="static/images/montage/m01_L.jpg" alt="" loading="lazy"><img src="static/images/montage/m01_R.jpg" alt="" loading="lazy"><img src="static/images/montage/m01_D.jpg" alt="" loading="lazy"></div><div class="montage__cap">Jerlov IA · shallow</div></div>
<div class="montage__card"><div class="montage__hdr"><span>Left</span><span>Right</span><span>Disp.</span></div><div class="montage__trip"><img src="static/images/montage/m02_L.jpg" alt="" loading="lazy"><img src="static/images/montage/m02_R.jpg" alt="" loading="lazy"><img src="static/images/montage/m02_D.jpg" alt="" loading="lazy"></div><div class="montage__cap">Jerlov IB · shallow</div></div>
<div class="montage__card"><div class="montage__hdr"><span>Left</span><span>Right</span><span>Disp.</span></div><div class="montage__trip"><img src="static/images/montage/m03_L.jpg" alt="" loading="lazy"><img src="static/images/montage/m03_R.jpg" alt="" loading="lazy"><img src="static/images/montage/m03_D.jpg" alt="" loading="lazy"></div><div class="montage__cap">Jerlov IC · shallow</div></div>
<div class="montage__card"><div class="montage__hdr"><span>Left</span><span>Right</span><span>Disp.</span></div><div class="montage__trip"><img src="static/images/montage/m04_L.jpg" alt="" loading="lazy"><img src="static/images/montage/m04_R.jpg" alt="" loading="lazy"><img src="static/images/montage/m04_D.jpg" alt="" loading="lazy"></div><div class="montage__cap">Jerlov 3C · shallow</div></div>
<div class="montage__card"><div class="montage__hdr"><span>Left</span><span>Right</span><span>Disp.</span></div><div class="montage__trip"><img src="static/images/montage/m05_L.jpg" alt="" loading="lazy"><img src="static/images/montage/m05_R.jpg" alt="" loading="lazy"><img src="static/images/montage/m05_D.jpg" alt="" loading="lazy"></div><div class="montage__cap">Jerlov IB · shallow</div></div>
<div class="montage__card"><div class="montage__hdr"><span>Left</span><span>Right</span><span>Disp.</span></div><div class="montage__trip"><img src="static/images/montage/m06_L.jpg" alt="" loading="lazy"><img src="static/images/montage/m06_R.jpg" alt="" loading="lazy"><img src="static/images/montage/m06_D.jpg" alt="" loading="lazy"></div><div class="montage__cap">Jerlov II · shallow</div></div>
<div class="montage__card"><div class="montage__hdr"><span>Left</span><span>Right</span><span>Disp.</span></div><div class="montage__trip"><img src="static/images/montage/m07_L.jpg" alt="" loading="lazy"><img src="static/images/montage/m07_R.jpg" alt="" loading="lazy"><img src="static/images/montage/m07_D.jpg" alt="" loading="lazy"></div><div class="montage__cap">Jerlov III · shallow</div></div>
<div class="montage__card"><div class="montage__hdr"><span>Left</span><span>Right</span><span>Disp.</span></div><div class="montage__trip"><img src="static/images/montage/m08_L.jpg" alt="" loading="lazy"><img src="static/images/montage/m08_R.jpg" alt="" loading="lazy"><img src="static/images/montage/m08_D.jpg" alt="" loading="lazy"></div><div class="montage__cap">Jerlov I · deep</div></div>
<div class="montage__card"><div class="montage__hdr"><span>Left</span><span>Right</span><span>Disp.</span></div><div class="montage__trip"><img src="static/images/montage/m09_L.jpg" alt="" loading="lazy"><img src="static/images/montage/m09_R.jpg" alt="" loading="lazy"><img src="static/images/montage/m09_D.jpg" alt="" loading="lazy"></div><div class="montage__cap">Jerlov II · deep</div></div>
<div class="montage__card"><div class="montage__hdr"><span>Left</span><span>Right</span><span>Disp.</span></div><div class="montage__trip"><img src="static/images/montage/m10_L.jpg" alt="" loading="lazy"><img src="static/images/montage/m10_R.jpg" alt="" loading="lazy"><img src="static/images/montage/m10_D.jpg" alt="" loading="lazy"></div><div class="montage__cap">Jerlov I · deep</div></div>
<div class="montage__card"><div class="montage__hdr"><span>Left</span><span>Right</span><span>Disp.</span></div><div class="montage__trip"><img src="static/images/montage/m11_L.jpg" alt="" loading="lazy"><img src="static/images/montage/m11_R.jpg" alt="" loading="lazy"><img src="static/images/montage/m11_D.jpg" alt="" loading="lazy"></div><div class="montage__cap">Jerlov IA · deep</div></div>
<div class="montage__card"><div class="montage__hdr"><span>Left</span><span>Right</span><span>Disp.</span></div><div class="montage__trip"><img src="static/images/montage/m12_L.jpg" alt="" loading="lazy"><img src="static/images/montage/m12_R.jpg" alt="" loading="lazy"><img src="static/images/montage/m12_D.jpg" alt="" loading="lazy"></div><div class="montage__cap">Jerlov IB · shallow</div></div>
</div>
<!-- SeaStereo download button hidden until the dataset is public.
<div class="center reveal">
<a class="btn" href="#" title="Dataset download (coming soon)">
<svg viewBox="0 0 24 24" aria-hidden="true"><path d="M12 3v12m0 0l-4-4m4 4l4-4M4 21h16" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" stroke-linejoin="round"/></svg>
Download SeaStereo <span class="pill__soon">soon</span>
</a>
</div>
-->
<h3 class="subhead reveal">Water-condition explorer</h3>
<p class="lead lead--center reveal">Hold the method fixed, change the water. As turbidity and depth climb, LinStereo holds.</p>
<div class="tabs" id="condTabs" role="tablist" aria-label="Water condition">
<button class="tab is-active" data-cond="clear_shallow" role="tab" aria-selected="true">Clear · shallow</button>
<button class="tab" data-cond="clear_deep" role="tab" aria-selected="false">Clear · deep</button>
<button class="tab" data-cond="murky_shallow" role="tab" aria-selected="false">Murky · shallow</button>
</div>
<figure class="figure figure--inset reveal">
<img id="condImg" src="static/images/cond_clear_shallow.jpg" alt="LinStereo result under varying water conditions" loading="lazy" />
</figure>
</div>
</section>
<!-- ================= STANDARD BENCHMARKS TABLE ================= -->
<section class="section" id="results">
<div class="container">
<h2 class="section__title">Standard benchmarks</h2>
<p class="lead lead--center reveal">Zero-shot cross-domain, official weights. We keep LinStereo a
<strong>ViT-B, SceneFlow-only</strong> model to isolate the decoder's contribution — competitive with ViT-L
systems and extra-data methods, and <strong>best among comparable models</strong> on occluded Middlebury.</p>
<div class="tablewrap reveal">
<table class="tbl">
<thead>
<tr>
<th class="tl" rowspan="2">Method</th><th rowspan="2">Extra<br>data</th>
<th colspan="2">KITTI 2015</th><th colspan="2">KITTI 2012</th>
<th colspan="6">Middlebury (H)</th>
<th colspan="2">ETH3D</th><th colspan="2">Booster (Q)</th>
</tr>
<tr>
<th>EPE↓</th><th>bad3↓</th><th>EPE↓</th><th>bad3↓</th>
<th>All EPE↓</th><th>All bad2↓</th><th>Noc EPE↓</th><th>Noc bad2↓</th><th>Occ EPE↓</th><th>Occ bad2↓</th>
<th>EPE↓</th><th>bad1↓</th><th>EPE↓</th><th>bad2↓</th>
</tr>
</thead>
<tbody>
<tr><td class="tl">PSMNet</td><td>–</td><td>4.05</td><td>28.51</td><td>3.77</td><td>27.34</td><td>11.48</td><td>36.06</td><td>10.56</td><td>32.11</td><td>17.69</td><td>62.58</td><td>2.15</td><td>15.13</td><td>14.17</td><td>49.79</td></tr>
<tr><td class="tl">RAFT-Stereo</td><td>–</td><td>1.13</td><td>5.69</td><td>0.90</td><td>4.35</td><td>1.92</td><td>12.6</td><td>1.09</td><td>8.65</td><td>3.31</td><td>26.39</td><td>0.36</td><td>3.3</td><td>4.18</td><td>17.64</td></tr>
<tr class="extra"><td class="tl">IGEV-Stereo</td><td>✓</td><td>1.21</td><td>6.03</td><td>1.03</td><td>5.13</td><td>2.63</td><td>11.93</td><td>2.27</td><td>9.49</td><td>5.02</td><td>26.04</td><td>0.33</td><td>4.0</td><td>4.26</td><td>17.58</td></tr>
<tr><td class="tl">UniMatch</td><td>–</td><td>1.63</td><td>11.01</td><td>1.71</td><td>12.04</td><td>4.21</td><td>34.72</td><td>3.96</td><td>31.52</td><td>5.70</td><td>56.51</td><td>0.69</td><td>18.61</td><td>8.57</td><td>47.30</td></tr>
<tr class="extra"><td class="tl">Selective-RAFT</td><td>✓</td><td>1.27</td><td>6.68</td><td>1.08</td><td>5.19</td><td>2.34</td><td>12.04</td><td>2.05</td><td>9.45</td><td>4.17</td><td>27.4</td><td>0.34</td><td>4.36</td><td>4.14</td><td>19.52</td></tr>
<tr class="extra"><td class="tl">Selective-IGEV</td><td>✓</td><td>1.25</td><td>6.06</td><td>1.08</td><td>5.64</td><td>2.59</td><td>11.79</td><td>2.31</td><td>9.22</td><td>4.35</td><td>28.10</td><td>0.33</td><td>4.05</td><td>4.62</td><td>19.28</td></tr>
<tr><td class="tl">MoCha-Stereo</td><td>–</td><td>1.29</td><td>5.97</td><td>1.02</td><td>4.83</td><td>2.66</td><td>10.18</td><td>2.49</td><td>7.96</td><td>3.84</td><td>24.16</td><td>0.28</td><td>3.47</td><td>3.88</td><td>16.82</td></tr>
<tr><td class="tl">NMRF</td><td>–</td><td>1.17</td><td>5.31</td><td>0.92</td><td>4.63</td><td>2.91</td><td>13.36</td><td>2.73</td><td>10.90</td><td>4.22</td><td>29.27</td><td>0.31</td><td>3.8</td><td>5.05</td><td>26.22</td></tr>
<tr class="extra"><td class="tl">IGEV++</td><td>✓</td><td>1.27</td><td>6.22</td><td>1.20</td><td>6.81</td><td>3.21</td><td>10.85</td><td>2.41</td><td>7.86</td><td>6.83</td><td>28.68</td><td>0.35</td><td>4.60</td><td>5.00</td><td>18.62</td></tr>
<tr><td class="tl">MGStereo</td><td>–</td><td>1.13</td><td>5.62</td><td>0.87</td><td>4.16</td><td>1.15</td><td>8.39</td><td><u>0.85</u></td><td><u>5.67</u></td><td>2.89</td><td>26.50</td><td>0.25</td><td>1.88</td><td>2.26</td><td>11.02</td></tr>
<tr><td class="tl">Stereo Anywhere <span class="vit">ViT-L</span></td><td>–</td><td>1.07</td><td>4.93</td><td>0.83</td><td>3.90</td><td>0.94</td><td>6.96</td><td><b>0.79</b></td><td><b>4.75</b></td><td>2.67</td><td>20.34</td><td><u>0.24</u></td><td>2.66</td><td><u>2.21</u></td><td><b>9.91</b></td></tr>
<tr><td class="tl">DEFOM-Stereo <span class="vit">ViT-L</span></td><td>–</td><td>1.06</td><td>4.99</td><td>0.84</td><td>3.76</td><td><u>0.84</u></td><td><b>5.91</b></td><td>0.90</td><td>5.76</td><td><u>2.11</u></td><td><u>19.34</u></td><td>0.35</td><td>2.35</td><td>3.52</td><td>13.20</td></tr>
<tr><td class="tl">MonSter <span class="vit">ViT-L</span></td><td>–</td><td><b>0.89</b></td><td><b>3.58</b></td><td><b>0.75</b></td><td><u>3.58</u></td><td>1.98</td><td>10.41</td><td>2.03</td><td>8.79</td><td>2.62</td><td>21.56</td><td><b>0.23</b></td><td><u>1.41</u></td><td>3.08</td><td>17.45</td></tr>
<tr><td class="tl">BridgeDepth <span class="vit">ViT-L</span></td><td>–</td><td>1.13</td><td>4.90</td><td>0.86</td><td>4.39</td><td>2.21</td><td>10.27</td><td>2.25</td><td>8.66</td><td>2.68</td><td>20.42</td><td><u>0.24</u></td><td><b>1.33</b></td><td>4.71</td><td>18.79</td></tr>
<tr class="extra"><td class="tl">FoundationStereo <span class="vit">ViT-L</span></td><td>✓</td><td>0.95</td><td>3.65</td><td>0.71</td><td>3.18</td><td>0.71</td><td>3.47</td><td>0.42</td><td>1.46</td><td>2.67</td><td>16.44</td><td>0.21</td><td>1.14</td><td>1.77</td><td>6.89</td></tr>
<tr class="ours"><td class="tl">LinStereo <span class="vit">ViT-B (Ours)</span></td><td>–</td><td><u>1.01</u></td><td><u>4.54</u></td><td><u>0.76</u></td><td><b>3.49</b></td><td><b>0.83</b></td><td><u>6.01</u></td><td>0.89</td><td><u>5.67</u></td><td><b>1.33</b></td><td><b>8.69</b></td><td><u>0.24</u></td><td>2.41</td><td><b>2.14</b></td><td><u>9.93</u></td></tr>
</tbody>
</table>
</div>
<p class="caption reveal"><b>Bold</b> = best, <u>underline</u> = second-best among comparable methods (no extra data).
FoundationStereo (ViT-L + extra data, grey) is the strongest baseline overall and shown for reference.
LinStereo's <strong>1.33</strong> occluded-EPE on Middlebury is <strong>37% below</strong> the previous best (DEFOM, 2.11) —
direct evidence of PALA propagating reliable disparity into occluded pixels. EPE and badX (% pixels past the error
threshold) reported per benchmark; scroll the table sideways for all columns.</p>
</div>
</section>
<!-- ================= UNDERWATER ================= -->
<section class="section section--ocean" id="underwater">
<div class="container">
<h2 class="section__title section__title--light">Underwater generalization</h2>
<p class="lead lead--center lead--light reveal">The headline domain. Trained on SceneFlow only, LinStereo is
<strong>best on every metric on both benchmarks</strong> — even against methods trained with real-world or
domain-specific data.</p>
<div class="tablewrap tablewrap--dark reveal">
<table class="tbl tbl--dark">
<thead><tr><th class="tl">Method · TartanAir-UW (simulated, long-range backscatter)</th><th>Extra<br>data</th><th>AbsRel↓</th><th>SqRel↓</th><th>RMSE↓</th><th>LogRMSE↓</th><th>A1↑</th><th>A2↑</th><th>A3↑</th></tr></thead>
<tbody>
<tr><td class="tl">UniMatch</td><td>–</td><td>0.17</td><td>2.59</td><td>6.35</td><td>0.26</td><td>0.759</td><td>0.830</td><td>0.837</td></tr>
<tr><td class="tl">MoCha-Stereo</td><td>–</td><td>0.15</td><td>2.30</td><td>6.02</td><td>0.25</td><td>0.769</td><td>0.835</td><td>0.843</td></tr>
<tr><td class="tl">NMRF</td><td>–</td><td>0.11</td><td>1.08</td><td>4.63</td><td>0.21</td><td>0.811</td><td>0.856</td><td>0.870</td></tr>
<tr><td class="tl">PSMNet</td><td>–</td><td>0.10</td><td>1.02</td><td>4.56</td><td>0.21</td><td>0.813</td><td>0.857</td><td>0.871</td></tr>
<tr><td class="tl">RAFT-Stereo</td><td>–</td><td>0.08</td><td>0.65</td><td>4.36</td><td>0.20</td><td>0.902</td><td>0.963</td><td>0.984</td></tr>
<tr><td class="tl">MGStereo</td><td>–</td><td>0.08</td><td>0.55</td><td>3.69</td><td>0.19</td><td>0.911</td><td>0.968</td><td>0.987</td></tr>
<tr><td class="tl">Stereo Anywhere</td><td>–</td><td>0.06</td><td>0.41</td><td>3.24</td><td>0.18</td><td>0.946</td><td>0.980</td><td>0.989</td></tr>
<tr><td class="tl">DEFOM-Stereo</td><td>–</td><td>0.05</td><td>0.41</td><td>3.13</td><td>0.17</td><td>0.953</td><td>0.982</td><td>0.990</td></tr>
<tr class="extra"><td class="tl">Selective-Stereo (RAFT)</td><td>✓</td><td>0.09</td><td>0.67</td><td>4.31</td><td>0.21</td><td>0.902</td><td>0.962</td><td>0.981</td></tr>
<tr class="extra"><td class="tl">Selective-Stereo (IGEV)</td><td>✓</td><td>0.11</td><td>1.02</td><td>5.01</td><td>0.23</td><td>0.856</td><td>0.942</td><td>0.976</td></tr>
<tr class="extra"><td class="tl">IGEV-Stereo</td><td>✓</td><td>0.10</td><td>0.90</td><td>4.68</td><td>0.21</td><td>0.891</td><td>0.955</td><td>0.979</td></tr>
<tr class="extra"><td class="tl">IGEV++</td><td>✓</td><td>0.09</td><td>0.81</td><td>4.37</td><td>0.20</td><td>0.905</td><td>0.962</td><td>0.983</td></tr>
<tr class="extra"><td class="tl">FoundationStereo</td><td>✓</td><td>0.05</td><td>0.40</td><td>3.01</td><td>0.16</td><td><b>0.959</b></td><td>0.984</td><td>0.991</td></tr>
<tr class="ours"><td class="tl">LinStereo <span class="vit">ViT-B (Ours)</span></td><td>–</td><td><b>0.04</b></td><td><b>0.19</b></td><td><b>2.08</b></td><td><b>0.09</b></td><td><b>0.959</b></td><td><b>0.992</b></td><td><b>0.998</b></td></tr>
</tbody>
</table>
</div>
<div style="height:18px"></div>
<div class="tablewrap tablewrap--dark reveal">
<table class="tbl tbl--dark">
<thead><tr><th class="tl">Method · SQUID (real-world, color attenuation)</th><th>Extra<br>data</th><th>AbsRel↓</th><th>SqRel↓</th><th>RMSE↓</th><th>LogRMSE↓</th><th>A1↑</th><th>A2↑</th><th>A3↑</th></tr></thead>
<tbody>
<tr><td class="tl">UniMatch</td><td>–</td><td>1.38</td><td>20.91</td><td>6.42</td><td>0.66</td><td>0.561</td><td>0.660</td><td>0.734</td></tr>
<tr><td class="tl">MoCha-Stereo</td><td>–</td><td>0.15</td><td>0.61</td><td>1.55</td><td>0.20</td><td>0.875</td><td>0.931</td><td>0.958</td></tr>
<tr><td class="tl">NMRF</td><td>–</td><td>0.43</td><td>5.27</td><td>2.57</td><td>0.33</td><td>0.844</td><td>0.904</td><td>0.935</td></tr>
<tr><td class="tl">PSMNet</td><td>–</td><td>0.46</td><td>4.31</td><td>3.43</td><td>0.51</td><td>0.736</td><td>0.813</td><td>0.853</td></tr>
<tr><td class="tl">RAFT-Stereo</td><td>–</td><td>0.07</td><td>0.27</td><td>1.25</td><td>0.12</td><td>0.937</td><td>0.971</td><td>0.987</td></tr>
<tr><td class="tl">MGStereo</td><td>–</td><td>0.09</td><td>0.71</td><td>1.99</td><td>0.16</td><td>0.925</td><td>0.958</td><td>0.976</td></tr>
<tr><td class="tl">Stereo Anywhere</td><td>–</td><td>0.07</td><td>0.40</td><td>1.46</td><td>0.13</td><td>0.937</td><td>0.976</td><td>0.985</td></tr>
<tr><td class="tl">DEFOM-Stereo</td><td>–</td><td>0.09</td><td>0.63</td><td>2.00</td><td>0.16</td><td>0.915</td><td>0.954</td><td>0.977</td></tr>
<tr class="extra"><td class="tl">Selective-Stereo (RAFT)</td><td>✓</td><td>0.11</td><td>0.22</td><td>1.16</td><td>0.16</td><td>0.877</td><td>0.940</td><td>0.968</td></tr>
<tr class="extra"><td class="tl">Selective-Stereo (IGEV)</td><td>✓</td><td>0.08</td><td>0.21</td><td>1.05</td><td>0.14</td><td>0.932</td><td>0.966</td><td>0.980</td></tr>
<tr class="extra"><td class="tl">IGEV-Stereo</td><td>✓</td><td>0.20</td><td>1.35</td><td>2.68</td><td>0.46</td><td>0.760</td><td>0.821</td><td>0.863</td></tr>
<tr class="extra"><td class="tl">IGEV++</td><td>✓</td><td>0.06</td><td>0.22</td><td>1.11</td><td>0.12</td><td>0.950</td><td>0.981</td><td>0.990</td></tr>
<tr class="extra"><td class="tl">FoundationStereo</td><td>✓</td><td>0.07</td><td>0.30</td><td>1.36</td><td>0.13</td><td>0.940</td><td>0.976</td><td>0.987</td></tr>
<tr class="ours"><td class="tl">LinStereo <span class="vit">ViT-B (Ours)</span></td><td>–</td><td><b>0.04</b></td><td><b>0.12</b></td><td><b>0.90</b></td><td><b>0.08</b></td><td><b>0.970</b></td><td><b>0.990</b></td><td><b>0.996</b></td></tr>
</tbody>
</table>
</div>
<p class="caption caption--light reveal">The largest gains track each benchmark's dominant degradation:
<strong>31% lower RMSE</strong> than FoundationStereo on TartanAir-UW (long-range backscatter) and
<strong>26% lower AbsRel</strong> than IGEV++ on SQUID (color attenuation). All baselines shown;
AbsRel/SqRel/RMSE/LogRMSE lower is better, A1/A2/A3 higher is better.</p>
</div>
</section>
<!-- ================= REAL-WORLD ================= -->
<section class="section" id="realworld">
<div class="container">
<h2 class="section__title">Real-world deployment</h2>
<p class="lead lead--center reveal">Beyond simulation: a controlled laboratory water tank at close range
(< 2 m), with AprilTag + CAD-model ground truth down to sub-millimetre — including
~3 mm taut ropes as a fine-structure stress test.</p>
<div class="tablewrap reveal">
<table class="tbl">
<thead>
<tr>
<th class="tl">Method</th><th>Extra<br>data</th>
<th>AbsRel↓</th><th>SqRel↓</th><th>RMSE↓</th><th>LogRMSE↓</th>
<th>A1↑</th><th>A2↑</th><th>A3↑</th>
</tr>
</thead>
<tbody>
<tr><td class="tl">PSMNet</td><td>–</td><td>0.18</td><td>0.16</td><td>0.34</td><td>0.28</td><td>0.87</td><td>0.90</td><td>0.93</td></tr>
<tr><td class="tl">UniMatch</td><td>–</td><td>0.08</td><td>0.04</td><td>0.17</td><td>0.16</td><td>0.93</td><td>0.96</td><td>0.98</td></tr>
<tr><td class="tl">Stereo Anywhere</td><td>–</td><td>0.09</td><td>0.06</td><td>0.20</td><td>0.19</td><td>0.92</td><td>0.93</td><td>0.97</td></tr>
<tr><td class="tl">RAFT-Stereo</td><td>–</td><td>0.06</td><td>0.03</td><td>0.15</td><td>0.14</td><td>0.94</td><td>0.96</td><td>0.99</td></tr>
<tr><td class="tl">MGStereo</td><td>–</td><td>0.08</td><td>0.05</td><td>0.18</td><td>0.17</td><td>0.93</td><td>0.94</td><td>0.98</td></tr>
<tr class="extra"><td class="tl">Selective-Stereo (RAFT)</td><td>✓</td><td>0.07</td><td>0.03</td><td>0.14</td><td>0.14</td><td>0.94</td><td>0.97</td><td>0.99</td></tr>
<tr class="extra"><td class="tl">Selective-Stereo (IGEV)</td><td>✓</td><td>0.08</td><td>0.04</td><td>0.16</td><td>0.16</td><td>0.91</td><td>0.95</td><td>0.99</td></tr>
<tr class="extra"><td class="tl">IGEV-Stereo</td><td>✓</td><td>0.06</td><td>0.02</td><td>0.13</td><td>0.12</td><td>0.94</td><td>0.97</td><td>0.99</td></tr>
<tr class="extra"><td class="tl">IGEV++</td><td>✓</td><td>0.05</td><td>0.02</td><td>0.12</td><td>0.12</td><td>0.96</td><td>0.97</td><td>0.99</td></tr>
<tr class="extra"><td class="tl">FoundationStereo</td><td>✓</td><td>0.07</td><td>0.04</td><td>0.18</td><td>0.17</td><td>0.93</td><td>0.94</td><td>0.98</td></tr>
<tr class="ours"><td class="tl">LinStereo (Ours, T=8)</td><td>–</td><td><b>0.04</b></td><td><b>0.01</b></td><td><b>0.07</b></td><td><b>0.07</b></td><td><b>0.98</b></td><td><b>0.99</b></td><td><b>1.00</b></td></tr>
</tbody>
</table>
</div>
<p class="caption reveal"><b>Bold</b> = best. Extra data (✓) = trained beyond SceneFlow.
AbsRel/SqRel/RMSE/LogRMSE lower is better; accuracy thresholds A1/A2/A3 higher is better. Scored only on valid-GT pixels
projecting onto the CAD mesh. LinStereo — SceneFlow-only — leads every metric at this near-range,
fine-structure target. Qualitative close-up in
<a href="#supplementary">More qualitative results → Real-world tank</a>.</p>
</div>
</section>
<!-- ================= EFFICIENCY ================= -->
<section class="section section--soft" id="efficiency">
<div class="container">
<h2 class="section__title">Efficiency & runtime</h2>
<p class="lead lead--center reveal">LinStereo trades raw speed for accuracy and zero-shot generalisation: it is a large
model built on a <strong>frozen Depth Anything V3</strong> backbone. The cost is explicit below — yet its
core <strong>PALA</strong> update operator costs about the same per iteration as a local ConvGRU, despite being
<em>global</em> and linear in image size.</p>
<h3 class="subhead reveal">Computational efficiency</h3>
<div class="tablewrap reveal">
<table class="tbl">
<thead><tr><th class="tl">Method</th><th>Params (M)↓</th><th>GFLOPs↓</th><th>Time (ms)↓</th><th>FPS↑</th></tr></thead>
<tbody>
<tr><td class="tl">LightStereo-S</td><td>3.44</td><td><b>45.4</b></td><td><b>9.9</b></td><td><b>101.0</b></td></tr>
<tr><td class="tl">CoEx</td><td>2.73</td><td>70.6</td><td>11.0</td><td>91.0</td></tr>
<tr><td class="tl">CGI-Stereo</td><td>3.50</td><td>81.6</td><td>12.8</td><td>78.1</td></tr>
<tr><td class="tl">Fast-ACVNet</td><td>3.08</td><td>104.6</td><td>13.2</td><td>76.0</td></tr>
<tr><td class="tl">ADStereo</td><td>7.10</td><td>201.6</td><td>15.4</td><td>64.7</td></tr>
<tr><td class="tl">RAFT-Stereo†</td><td>9.87</td><td>296.4</td><td>18.0</td><td>55.7</td></tr>
<tr><td class="tl">RT-IGEV</td><td>4.17</td><td>354.2</td><td>24.7</td><td>40.5</td></tr>
<tr><td class="tl">MobileStereoNet</td><td><b>2.35</b></td><td>175.6</td><td>37.7</td><td>26.5</td></tr>
<tr><td class="tl">AANet</td><td>3.93</td><td>–</td><td>65.8</td><td>15.2</td></tr>
<tr class="ours"><td class="tl">LinStereo (Ours, T=2)</td><td>127.0</td><td>770.4</td><td>80</td><td>12.5</td></tr>
</tbody>
</table>
</div>
<p class="caption reveal"><b>Bold</b> = best. Measured at 480×640 on a single NVIDIA RTX 4500 (24 GB);
LinStereo at <strong>T = 2</strong> (accuracy elsewhere uses T = 8). LinStereo is a large,
accuracy-first model — not a lightweight real-time network.</p>
<h3 class="subhead reveal">Per-iteration update operator</h3>
<div class="tablewrap reveal" style="max-width:480px;margin:0 auto;">
<table class="tbl tbl--compact">
<thead><tr><th class="tl">Update operator</th><th>Latency (ms)↓</th></tr></thead>
<tbody>
<tr class="ours"><td class="tl">PALA (Ours)</td><td>3.50 ± 0.05</td></tr>
<tr><td class="tl">RAFT-Stereo ConvGRU</td><td>3.63 ± 0.06</td></tr>
<tr><td class="tl">IGEV ConvGRU</td><td>3.43 ± 0.03</td></tr>
</tbody>
</table>
</div>
<p class="caption reveal">Per-update latency at 480×640 (± 95% CI). PALA's <em>global</em>,
linear-complexity attention costs essentially the same per iteration as the <em>local</em> ConvGRU operators it replaces.</p>