-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
1059 lines (910 loc) · 39.3 KB
/
Copy pathindex.html
File metadata and controls
1059 lines (910 loc) · 39.3 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>
<head>
<meta charset="utf-8">
<title>AnyThermal: Towards Learning Universal Representations for Thermal Perception</title>
<meta name="description" content="AnyThermal: a universal thermal feature backbone distilled from RGB foundation models, plus the TartanRGBT platform and dataset.">
<meta name="viewport" content="width=device-width, initial-scale=1">
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<link rel="stylesheet" href="./static/css/bulma.min.css">
<link rel="stylesheet" href="./static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="./static/css/bulma-slider.min.css">
<link rel="stylesheet" href="./static/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="./static/css/index.css">
<link rel="icon" href="./static/images/favicon.svg">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="./static/js/fontawesome.all.min.js"></script>
<script src="./static/js/bulma-carousel.min.js"></script>
<script src="./static/js/bulma-slider.min.js"></script>
<script src="./static/js/index.js"></script>
<script type="module" src="https://gradio.s3-us-west-2.amazonaws.com/4.36.1/gradio.js"></script>
<style>
.mia {
font-family: monospace;
font-weight: bold;
}
.publication-authors a {
color: #3273dc; /* normal link color (Bulma default) */
transition: color 0.2s ease;
text-decoration: none;
}
.publication-authors a:hover {
color: #ff7a00; /* hover color (orange like your theme) */
}
</style>
</head>
<body>
<!-- NAVBAR -->
<nav class="navbar" role="navigation" aria-label="main navigation">
<div class="navbar-brand">
<a role="button" class="navbar-burger" aria-label="menu" aria-expanded="false">
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
<span aria-hidden="true"></span>
</a>
</div>
<div class="navbar-menu">
<div class="navbar-start" style="flex-grow:1; justify-content:center;">
<a class="navbar-item" href="#overview">Overview</a>
<a class="navbar-item" href="#method">Method</a>
<a class="navbar-item" href="#platform">Platform</a>
<a class="navbar-item" href="#dataset">Dataset</a>
<a class="navbar-item" href="#results">Results</a>
<a class="navbar-item" href="#scaling">Scaling & Diversity</a>
<a class="navbar-item" href="#resources">Contact</a>
</div>
</div>
</nav>
<!-- HERO / OVERVIEW -->
<section class="hero" id="overview">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title">
<span class="anythermal-gradient">AnyThermal</span>: Towards Learning Universal Representations for Thermal Perception
</h1>
<br>
<div class="has-text-centered" style="margin-bottom:1.5rem;">
<img src="assets/ICRA-VIE26-logo.png"
alt="Additional illustration"
style="
max-width:60%;
border:2px solid #ddd;
border-radius:10px;
padding:2px;
box-shadow:0 2px 6px rgba(0,0,0,0.15);
">
</div>
<div class="is-size-5 publication-authors">
<span class="author-block">
<a href="https://www.parvmaheshwari.com/" target="_blank">Parv Maheshwari</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://jaykarhade.github.io/" target="_blank">Jay Karhade</a>*<sup>1</sup>,
</span>
<span class="author-block">
<a href="http://yogesh-chawla.com/" target="_blank">Yogesh Chawla</a>*<sup>2</sup>,
</span>
<span class="author-block">
<a href="https://riss.ri.cmu.edu/student/isaiah-adu/" target="_blank">Isaiah Adu</a><sup>3</sup>,
</span>
<span class="author-block" style="color:#6a1b9a;">
Florian Heisen<sup>4</sup>,
</span>
<span class="author-block">
<a href="https://www.ri.cmu.edu/ri-people/andrew-porco/" target="_blank">Andrew Porco</a><sup>5</sup>,
</span>
<span class="author-block">
<a href="https://www.ri.cmu.edu/ri-people/andrew-jong/" target="_blank">Andrew Jong</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://www.ri.cmu.edu/ri-people/yifei-liu/" target="_blank">Yifei Liu</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://bse.unl.edu/person/santosh-pitla/" target="_blank">Santosh Pitla</a><sup>2</sup>,
</span>
<span class="author-block">
<a href="https://www.ri.cmu.edu/ri-faculty/sebastian-scherer/" target="_blank">Sebastian Scherer</a><sup>1</sup>,
</span>
<span class="author-block">
<a href="https://www.ri.cmu.edu/ri-faculty/wenshan-wang/" target="_blank">Wenshan Wang</a><sup>1</sup>
</span>
</div>
<div class="is-size-6 publication-authors">
<span class="author-block"><sup>1</sup>Robotics Institute, Carnegie Mellon University, Pittsburgh, PA, USA</span><br>
<span class="author-block"><sup>2</sup>Biological Systems Engineering, University of Nebraska–Lincoln, Lincoln, NE, USA</span><br>
<span class="author-block"><sup>3</sup>Mechanical Engineering, Penn State University, University Park, PA, USA</span><br>
<span class="author-block"><sup>4</sup>School of Engineering and Design, Technical University of Munich, Munich, Germany</span><br>
<span class="author-block"><sup>5</sup>Mechanical Engineering, Carnegie Mellon University, Pittsburgh, PA, USA</span>
</div>
<div class="is-size-6 publication-authors" style="margin-bottom: 2rem;">
* Equal Contribution
</div>
<!-- Links -->
<div class="publication-links">
<span class="link-block">
<a href="https://arxiv.org/abs/2602.06203" class="external-link button is-normal is-rounded is-dark" target="_blank">
<span class="icon">
<img src="assets/arxiv_x_logo.png" alt="arXiv" style="height:1.2em;">
</span>
<span>arXiv</span>
</a>
</span>
<span class="link-block">
<span class="link-block">
<a class="external-link button is-normal is-rounded is-dark"
href="https://huggingface.co/datasets/theairlabcmu/TartanRGBT"
target="_blank" rel="noopener">
<span class="icon">
<img src="https://huggingface.co/front/assets/huggingface_logo.svg"
alt="Hugging Face"
style="height:1.2em;">
</span>
<span>Dataset</span>
</a>
</span>
<span class="link-block">
<a href="https://huggingface.co/theairlabcmu/AnyThermal"
class="external-link button is-normal is-rounded is-dark"
target="_blank" rel="noopener">
<span class="icon">
<img src="https://huggingface.co/front/assets/huggingface_logo.svg"
alt="Hugging Face"
style="height:1.2em;">
</span>
<span>Try Model</span>
</a>
</span>
<span class="link-block">
<a href="https://github.com/castacks/AnyThermal/tree/main"
class="external-link button is-normal is-rounded is-dark"
target="_blank" rel="noopener">
<span class="icon"><i class="fab fa-github"></i></span>
<span>AnyThermal</span>
</a>
</span>
<span class="link-block">
<a href="https://github.com/AnyThermal/tartan_rgbt_ws"
class="external-link button is-normal is-rounded is-dark"
target="_blank" rel="noopener">
<span class="icon"><i class="fab fa-github"></i></span>
<span>TartanRGBT Software Stack</span>
</a>
</span>
<span class="link-block">
<a href="https://docs.google.com/document/d/1X5Av5y16SZzcaHOHzIiit1tz_KwJjkek2Er0iPeW7_g/edit?usp=sharing"
class="external-link button is-normal is-rounded is-dark"
target="_blank">
<span class="icon"><i class="fas fa-tools"></i></span>
<span>TartanRGBT Assembly Instructions</span>
</a>
</span>
<span class="link-block">
<a href="assets/AnyThermal_ICRA2026_supplementary.mp4"
class="external-link button is-normal is-rounded is-dark"
target="_blank">
<span class="icon"><i class="fas fa-video"></i></span>
<span>Paper Video</span>
</a>
</span>
<span class="link-block">
<a href="#results" class="external-link button is-normal is-rounded is-dark">
<span class="icon"><i class="fas fa-chart-line"></i></span>
<span>Results</span>
</a>
</span>
</div>
</div>
</div>
<p class="has-text-centered" style="margin-top:1rem; margin-bottom:2.5rem; line-height:1.35;">
<b>TL;DR:</b>
<span style="color:#ff7a00; font-weight:700;">AnyThermal</span> is a task-agnostic backbone achieving state-of-the-art results in thermal place recognition, segmentation, and depth estimation by distilling visual foundation models.
Supporting this is the
<span style="color:#ff7a00; font-weight:700;">TartanRGBT platform</span>, the first open-source framework for hardware-synchronized RGB–thermal data collection.
We leverage this to provide the
<span style="color:#ff7a00; font-weight:700;">TartanRGBT dataset</span>—16,943 diverse RGB–thermal pairs across urban, indoor, off-road, and park environments.
</p>
<!-- Abstract -->
<div class="columns is-centered" id="abstract">
<div class="column is-full">
<h2 class="title is-3 has-text-centered">Abstract</h2>
<!-- Graphical Abstract: full row width -->
<div class="has-text-centered" style="margin-bottom:1.5rem;">
<img src="assets/AnyThermal-intro.png"
alt="AnyThermal graphical abstract"
style="width:100%; border-radius:8px;">
</div>
<div class="content has-text-justified">
<p>
We present <b>AnyThermal</b>, a thermal backbone that captures robust, task-agnostic thermal features suitable for a variety of tasks such as cross-modal place recognition, thermal segmentation, and monocular depth estimation using thermal images.
Existing thermal backbones that follow task-specific training from small-scale data result in utility limited to a specific environment and task.
Unlike prior methods, <b>AnyThermal</b> can be used across a wide range of environments (indoor, aerial, off-road, urban) and tasks without task-specific training.
Our key insight is to distill feature representations from visual foundation models such as DINOv2 into a thermal encoder using data collected across these environments.
To bridge the diversity gap in existing RGB–thermal datasets, we introduce the <b>TartanRGBT platform</b>, the first open-source data collection framework with hardware-synchronized RGB–thermal image acquisition.
Using this payload, we collect the <b>TartanRGBT dataset</b>, a diverse and balanced dataset spanning four environments.
We demonstrate the effectiveness of <b>AnyThermal</b> and <b>TartanRGBT</b>, achieving state-of-the-art results with improvements of up to 36% across diverse environments and downstream tasks.
</p>
</div>
<!-- KEY CONTRIBUTIONS -->
<section class="section section-anchor" id="specs">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column is-four-fifths">
<h2 class="title is-3 has-text-centered">Key Contributions</h2>
<div class="content">
<ul>
<li>
<b style="color:#ff7a00;">AnyThermal Backbone:</b>
A task-agnostic thermal feature extractor developed through knowledge distillation from visual foundation models. When paired with lightweight heads, it achieves state-of-the-art results in thermal segmentation and cross-modal place recognition, while outperforming comparable RGB-based backbones in monocular thermal depth estimation.
</li>
<li>
<b style="color:#ff7a00;">TartanRGBT Platform:</b>
The first open-source, hardware-synchronized framework for simultaneous stereo RGB and thermal image acquisition.
By releasing CAD files and a complete software stack, the platform lowers the barrier to large-scale RGB–thermal data collection.
</li>
<li>
<b style="color:#ff7a00;">TartanRGBT Dataset:</b>
A balanced collection of 16,943 synchronized RGB–thermal pairs spanning urban, indoor, off-road, and aerial environments.
This diverse dataset substantially improves the generalization and downstream performance of thermal perception models.
</li>
</ul>
</div>
</div>
</div>
</div>
</section>
<!-- TEASER: Interactive Demo -->
<section class="hero is-light section-anchor" id="teaser">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h2 class="title is-3">AnyThermal in Action</h2>
<!-- Task Buttons -->
<div class="buttons is-centered" style="margin-bottom:1.5rem;">
<button class="button is-rounded is-dark demo-btn is-active"
onclick="switchDemo('vpr')">
Cross-Modal Place Recognition
</button>
<button class="button is-rounded is-dark demo-btn"
onclick="switchDemo('seg')">
Segmentation
</button>
<button class="button is-rounded is-dark demo-btn"
onclick="switchDemo('depth')">
Depth
</button>
</div>
<!-- Single Display Window -->
<div class="columns is-centered">
<div class="column is-two-thirds has-text-centered">
<img id="demo-image"
src="assets/AnyThermal-VPR.png"
style="width:100%; border-radius:8px; transition:0.3s ease;">
<p id="demo-caption"
class="is-size-6"
style="margin-top:0.5rem;">
Robust place retrieval under severe illumination and appearance changes.
</p>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- METHOD -->
<section class="section section-anchor" id="method">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">AnyThermal Backbone</h2>
<div class="columns is-vcentered">
<div class="column is-half">
<h3 class="title is-4">RGB–Thermal Distillation</h3>
<div class="content has-text-justified">
<p>
AnyThermal uses two ViT-B/14 DINOv2 encoders: a frozen RGB teacher and a trainable thermal student, both initialized with pre-trained RGB weights.
</p>
<p>
Thermal images are converted to three channels and passed through the student; a contrastive loss aligns the CLS-token embeddings of corresponding RGB–thermal pairs, encouraging shared global semantics while relaxing pixel-perfect alignment.
</p>
<p>
Distillation is performed across multiple RGB–thermal datasets, enabling the backbone to learn environment-agnostic thermal representations that generalize across tasks and domains.
In particular, <b style="color:#ff7a00;">AnyThermal</b> is trained using five datasets spanning diverse environments, including urban driving datasets such as ViVID++ (outdoor sequences), STheREo, Freiburg, and <b style="color:#ff7a00;">TartanRGBT</b> (ours); the Boson Nighttime Dataset for aerial scenarios; and <b style="color:#ff7a00;">TartanRGBT</b> (ours) for both indoor and off-road environments.
</p>
</div>
</div>
<div class="column is-half has-text-centered">
<img src="assets/student_teacher.png" alt="AnyThermal distillation overview" style="max-width:100%;">
<p class="is-size-7">
Task-agnostic knowledge distillation from a frozen RGB DINOv2 teacher to a trainable thermal student (<b style="color:#ff7a00;">AnyThermal</b>), enabling label-free learning of generalizable thermal features across environments.
</p>
</div>
</div>
<div class="columns">
<div class="column">
<h3 class="title is-4">Task Heads (Frozen Backbone)</h3>
<div class="content">
<ul>
<li><b>Cross-modal Place Recognition:</b> SALAD head over AnyThermal features with triplet loss.</li>
<li><b>Thermal Segmentation:</b> a lightweight two-layer MLP head over ViT patch embeddings.</li>
<li><b>Mono-thermal Depth:</b> MiDaS-style encoder–decoder using multi-scale AnyThermal features.</li>
</ul>
</div>
</div>
</div>
</div>
</section>
<!-- PLATFORM -->
<section class="hero is-white section-anchor" id="platform">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">TartanRGBT Platform</h2>
<div class="columns is-centered">
<div class="column is-half has-text-centered">
<img src="assets/tartanrgbt_setup.gif" alt="TartanRGBT platform CAD and wiring" style="width:100%; border-radius:8px;">
<p class="is-size-7">
CAD model and wiring overview of the handheld TartanRGBT platform.
</p>
</div>
<div class="column is-half has-text-centered">
<img src="assets/cameras_setup.gif" alt="TartanRGBT camera setup" style="width:100%; border-radius:8px;">
<p class="is-size-7">
Camera and sensor placement in the TartanRGBT platform.
</p>
</div>
</div>
<div class="columns">
<div class="column">
<div class="content has-text-justified">
<p>
The TartanRGBT platform is a handheld rig that captures synchronized stereo RGB, stereo thermal, and IMU at 30 Hz using a ZED X camera, two FLIR Boson 640+ thermal cameras, and an NVIDIA Jetson AGX Orin 64 GB computer.
</p>
<p>
All cameras are hardware-timed: the ZED X pair is factory-synced, and a trigger from the capture card synchronizes the thermal cameras via external sync pins in slave mode.
</p>
<p>
A custom 3D-printed enclosure with ergonomic handles, cooling fans, and exposed ports makes the system field-ready, while Docker-based auto-launch and a single recording button simplify operation.
</p>
</div>
</div>
</div>
<div class="columns">
<div class="column">
<h3 class="title is-4">Calibration & Registration</h3>
<div class="content">
<ul>
<li>Thermal intrinsics are calibrated using a custom heated checkerboard, followed by fisheye rectification of thermal images.</li>
<li>Extrinsics between RGB and thermal cameras are derived from the CAD model, enabling 3D-aware alignment.</li>
<li>Registered RGB–thermal pairs are produced by estimating dense depth with FoundationStereo, transforming points into the thermal frame, and projecting with thermal intrinsics.</li>
</ul>
</div>
</div>
</div>
<div class="columns is-centered" style="margin-top:2rem;">
<div class="column has-text-centered">
<table style="margin:auto; border-collapse:separate; border-spacing:2px; width:100%;">
<tr>
<td><img src="static/images/00001970_alpha000.png" style="width:90%;"></td>
<td><img src="static/images/00001970_alpha025.png" style="width:90%;"></td>
<td><img src="static/images/00001970_alpha050.png" style="width:90%;"></td>
<td><img src="static/images/00001970_alpha075.png" style="width:90%;"></td>
<td><img src="static/images/00001970_alpha100.png" style="width:90%;"></td>
</tr>
<tr>
<td><img src="static/images/00002980_alpha000.png" style="width:90%;"></td>
<td><img src="static/images/00002980_alpha025.png" style="width:90%;"></td>
<td><img src="static/images/00002980_alpha050.png" style="width:90%;"></td>
<td><img src="static/images/00002980_alpha075.png" style="width:90%;"></td>
<td><img src="static/images/00002980_alpha100.png" style="width:90%;"></td>
</tr>
<tr>
<td><img src="static/images/00002390_alpha000.png" style="width:90%;"></td>
<td><img src="static/images/00002390_alpha025.png" style="width:90%;"></td>
<td><img src="static/images/00002390_alpha050.png" style="width:90%;"></td>
<td><img src="static/images/00002390_alpha075.png" style="width:90%;"></td>
<td><img src="static/images/00002390_alpha100.png" style="width:90%;"></td>
</tr>
<tr>
<td class="is-size-6" style="text-align:center;">α = 0.00</td>
<td class="is-size-6" style="text-align:center;">α = 0.25</td>
<td class="is-size-6" style="text-align:center;">α = 0.50</td>
<td class="is-size-6" style="text-align:center;">α = 0.75</td>
<td class="is-size-6" style="text-align:center;">α = 1.00</td>
</tr>
</table>
<p class="is-size-6" style="margin-top:0.5rem;">
Alpha-blended RGB–thermal overlays across different scenes (α controls the transparency of the thermal image over the RGB image), showing pixel-wise alignment.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- DATASET -->
<section class="section section-anchor" id="dataset" style="padding-top:1rem;">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">TartanRGBT Dataset</h2>
<div class="columns is-multiline is-centered is-variable is-4">
<div class="column is-half has-text-centered">
<img src="assets/indoor.gif"
alt="Indoor results"
style="width:100%; border-radius:8px;">
<p class="is-size-6 has-text-centered" style="margin-top:0.5rem;">Indoor</p>
</div>
<div class="column is-half has-text-centered">
<img src="assets/parks.gif"
alt="Park results"
style="width:100%; border-radius:8px;">
<p class="is-size-6 has-text-centered" style="margin-top:0.5rem;">Parks</p>
</div>
<div class="column is-half has-text-centered">
<img src="assets/offroad.gif"
alt="Off-road results"
style="width:100%; border-radius:8px;">
<p class="is-size-6 has-text-centered" style="margin-top:0.5rem;">Off-road</p>
</div>
<div class="column is-half has-text-centered">
<img src="assets/urban.gif"
alt="Urban results"
style="width:100%; border-radius:8px;">
<p class="is-size-6 has-text-centered" style="margin-top:0.5rem;">Urban</p>
</div>
</div>
<div class="columns is-centered">
<div class="column is-full">
<div class="content has-text-justified">
<p>
The TartanRGBT dataset consists of 16,943 synchronized, registered RGB–thermal pairs sampled at 1 Hz for non-redundant distillation, covering indoor, urban driving, parks, and off-road environments.
</p>
<p>
Each sequence includes stereo RGB, stereo thermal, IMU, and thermal FFC status.
We will also release the data as ROS bag files, allowing users to extract synchronized sensor streams at their desired sampling frequency.
</p>
</div>
<h3 class="title is-4">Diversity Compared to Existing RGB–T Datasets</h3>
<table class="table is-striped is-fullwidth is-hoverable">
<thead>
<tr>
<th>Dataset</th>
<th>Platform</th>
<th>RGB–T Pairs @1 Hz</th>
<th>Sync</th>
<th>Registered</th>
<th>Indoor</th>
<th>Off-road</th>
<th>Aerial</th>
<th>Urban Drive</th>
<th>Urban Park</th>
</tr>
</thead>
<tbody>
<tr>
<td>MS2</td>
<td>Vehicle</td>
<td>16,215</td>
<td>Yes</td>
<td>No</td>
<td>No</td>
<td>No</td>
<td>No</td>
<td>Yes</td>
<td>No</td>
</tr>
<tr>
<td>ViVID++</td>
<td>Handheld/Vehicle</td>
<td>14,824</td>
<td>Mixed</td>
<td>No</td>
<td>Limited</td>
<td>No</td>
<td>Yes</td>
<td>Yes</td>
<td>No</td>
</tr>
<tr>
<td>CART</td>
<td>Handheld/Drone</td>
<td>9,678</td>
<td>Mixed</td>
<td>Yes</td>
<td>No</td>
<td>Yes</td>
<td>Yes</td>
<td>Yes</td>
<td>No</td>
</tr>
<tr>
<td>OdomBeyondVision</td>
<td>Drone/UGV/Handheld</td>
<td>7,129</td>
<td>Yes</td>
<td>No</td>
<td>Yes</td>
<td>No</td>
<td>No</td>
<td>No</td>
<td>No</td>
</tr>
<tr>
<td><b>TartanRGBT (Ours)</b></td>
<td>Handheld</td>
<td>16,943</td>
<td><b>Yes</b></td>
<td><b>Yes</b></td>
<td><b>Yes</b></td>
<td><b>Yes</b></td>
<td>No</td>
<td><b>Yes</b></td>
<td><b>Yes</b></td>
</tr>
</tbody>
</table>
</div>
</div>
</div>
</div>
</section>
<!-- RESULTS -->
<section class="hero is-light section-anchor" id="results">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Results Across Tasks</h2>
<div class="columns">
<div class="column">
<h3 class="title is-4">Cross-Modal Place Recognition</h3>
<div class="content">
<p>
AnyThermal with a SALAD head (AnyThermal-VPR) surpasses RGB-only and RGB–thermal baselines on MS2 (urban), CART (aerial), and OdomBeyondVision (indoor) in Recall@1.
</p>
<div class="table-container">
<table class="table is-bordered is-fullwidth is-narrow">
<thead>
<tr>
<th>Model</th>
<th>Backbone</th>
<th>Head</th>
<th>MS2 R@1</th>
<th>CART R@1</th>
<th>OBV R@1</th>
</tr>
</thead>
<tbody>
<tr>
<td>DINOv2</td>
<td>DINOv2</td>
<td>CLS</td>
<td>27.21</td>
<td>25.98</td>
<td>29.49</td>
</tr>
<tr>
<td>SALAD</td>
<td>DINOv2</td>
<td>SALAD</td>
<td>76.97</td>
<td>49.38</td>
<td>38.94</td>
</tr>
<tr>
<td>ImageBind</td>
<td>ViT-H</td>
<td>CLS</td>
<td>0.79</td>
<td>1.13</td>
<td>10.25</td>
</tr>
<tr>
<td>SGM</td>
<td>ResNet-18</td>
<td>NetVLAD</td>
<td>20.02</td>
<td>45.59</td>
<td>21.05</td>
</tr>
<tr>
<td>AnyThermal</td>
<td>AnyThermal</td>
<td>CLS</td>
<td>75.39</td>
<td>45.45</td>
<td>45.40</td>
</tr>
<tr>
<td><b>AnyThermal-VPR</b></td>
<td>AnyThermal</td>
<td>SALAD</td>
<td><b>81.11</b></td>
<td><b>56.00</b></td>
<td><b>53.17</b></td>
</tr>
</tbody>
</table>
</div>
</div>
</div>
</div>
<!-- VPR qualitative results -->
<h3 class="title is-4">Qualitative Cross-Modal Place Recognition</h3>
</h4>
<div class="columns is-centered" style="margin-top:1rem;">
<div class="column is-two-thirds has-text-centered">
<img src="assets/VPR_combined.gif"
alt="Cross-modal place recognition comparison"
style="width:100%; border-radius:8px;">
</div>
</div>
<p class="is-size-7 has-text-centered" style="margin-top:0.5rem;">
<b>Middle:</b> The RGB-only baseline retrieves an incorrect database match under significant illumination and appearance changes.<br>
<b>Right:</b> <b style="color:#ff7a00;">AnyThermal</b> correctly retrieves the true location by leveraging robust thermal representations that are invariant to lighting variations.
</p>
<div class="column">
<h3 class="title is-4">Thermal Segmentation</h3>
<div class="content">
<p>
On MFNet, AnyThermal with a two-layer MLP segmentation head achieves 53.47% mIoU and runs at 6.79 FPS on Orin, outperforming RTFFNet-152 and MCNet while being up to 3.6× faster than the closest competitor.
</p>
<div class="table-container">
<table class="table is-fullwidth is-narrow">
<thead>
<tr>
<th>Model</th>
<th>Params (M)</th>
<th>mIoU (%)</th>
<th>FPS (Orin)</th>
</tr>
</thead>
<tbody>
<tr>
<td>RTFNet-152</td>
<td>196.37</td>
<td>47.00</td>
<td>8.37</td>
</tr>
<tr>
<td>MCNet</td>
<td>54.65</td>
<td>51.95</td>
<td>1.88</td>
</tr>
<tr>
<td>RGB DINO-SEG</td>
<td>87.02</td>
<td>45.46</td>
<td>6.79</td>
</tr>
<tr>
<td><b>AnyThermal-SEG</b></td>
<td>87.02</td>
<td><b>53.47</b></td>
<td>6.79</td>
</tr>
</tbody>
</table>
</div>
<!-- Segmentation qualitative results (vertical stack) -->
<div class="columns is-centered is-multiline" style="margin-top:1.5rem;">
<div class="column is-full has-text-centered">
<img src="assets/image.png"
alt="Segmentation qualitative result 1"
style="width:100%; max-width:900px; border-radius:8px;">
<p class="is-size-7">Baseline segmentation vs. AnyThermal prediction.</p>
</div>
<div class="column is-full has-text-centered">
<img src="assets/thermal_quant_2.png"
alt="Segmentation qualitative result 2"
style="width:100%; max-width:900px; border-radius:8px;">
<p class="is-size-7">Improved boundary and class consistency with AnyThermal.</p>
</div>
</div>
<div class="column">
<h3 class="title is-4">Mono-Thermal Depth Estimation</h3>
<div class="content">
<p>
Plugging AnyThermal into MiDaS on MS2 yields lower AbsRel and RMSE than EfficientNet-Lite3 and RGB DINOv2 backbones.
</p>
<div class="table-container">
<table class="table is-fullwidth is-narrow">
<thead>
<tr>
<th>Backbone</th>
<th>AbsRel ↓</th>
<th>SqRel ↓</th>
<th>RMSE ↓</th>
<th>RMSElog ↓</th>
</tr>
</thead>
<tbody>
<tr>
<td>EfficientNet-Lite3</td>
<td>0.1015</td>
<td>0.3955</td>
<td>2.9587</td>
<td>0.1417</td>
</tr>
<tr>
<td>DINOv2 ViT-B/14</td>
<td>0.0905</td>
<td>0.3177</td>
<td>2.7493</td>
<td>0.1208</td>
</tr>
<tr>
<td><b>AnyThermal</b></td>
<td><b>0.0883</b></td>
<td><b>0.3142</b></td>
<td><b>2.7432</b></td>
<td><b>0.1182</b></td>
</tr>
</tbody>
</table>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- SCALING DATA & DIVERSITY -->
<section class="section section-anchor" id="scaling">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Scaling Data & Diversity</h2>
<div class="columns is-centered">
<div class="column is-four-fifths">
<div class="content has-text-justified">
<p>
Ablation studies over pre-training datasets show that simply adding more urban data leads to performance saturation.
In contrast, incorporating <b style="color:#ff7a00;">TartanRGBT</b> significantly improves cross-modal place recognition, thermal segmentation, and depth estimation across domains.
</p>
<p>
The observed trends indicate that <b style="color:#ff7a00;">AnyThermal</b> has not yet plateaued with the current scale of RGB–thermal data, suggesting further gains from broader and more diverse data collection.
</p>
<p>
Training on a single aerial dataset introduces domain bias, reducing performance in urban scenes while improving results on aerial benchmarks.
This highlights the importance of multi-domain RGB–thermal data for learning transferable thermal representations.
</p>
</div>
<div class="columns is-multiline is-centered" style="margin-top:1.5rem;">
<div class="column is-one-third has-text-centered">
<img src="assets/thermal_segmentation.jpg"
alt="Scaling effect on thermal segmentation"
style="width:100%; border-radius:8px;">
<p class="is-size-7">
Thermal segmentation performance vs. pre-training datasets.
</p>
</div>
<div class="column is-one-third has-text-centered">
<img src="assets/vpr.jpg"
alt="Scaling effect on cross-modal place recognition"
style="width:100%; border-radius:8px;">
<p class="is-size-7">
Cross-modal place recognition performance vs. dataset diversity.
</p>
</div>
<div class="column is-one-third has-text-centered">
<img src="assets/depth.jpg"
alt="Scaling effect on monocular thermal depth estimation"
style="width:100%; border-radius:8px;">
<p class="is-size-7">
Monocular thermal depth estimation error vs. dataset diversity.
</p>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- HOW TO CITE -->
<section class="section section-anchor" id="citation">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">BibTex</h2>
<div class="columns is-centered">
<div class="column is-four-fifths">
<div class="content">
<p>
If you find our repository useful, please cite our paper in your work:
</p>
<pre><code>@misc{maheshwari2026anythermallearninguniversalrepresentations,
title={AnyThermal: Towards Learning Universal Representations for Thermal Perception},
author={Parv Maheshwari and Jay Karhade and Yogesh Chawla and Isaiah Adu and Florian Heisen and Andrew Porco and Andrew Jong and Yifei Liu and Santosh Pitla and Sebastian Scherer and Wenshan Wang},
year={2026},
eprint={2602.06203},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2602.06203},
}</code></pre>
</div>
</div>
</div>
</div>
</section>
<!-- RESOURCES -->
<section class="section section-anchor" id="resources">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered">Paper & Contact</h2>
<div class="columns is-centered">
<div class="column is-four-fifths has-text-centered">
<div class="buttons is-centered">
<a href="https://arxiv.org/abs/2602.06203"
class="button is-rounded is-dark"
target="_blank" rel="noopener">
<span class="icon">
<img src="assets/arxiv_x_logo.png"
alt="arXiv"
style="height:1.2em;">
</span>
<span> arXiv</span>
</a>
<a href="mailto:parvm@andrew.cmu.edu" class="button is-rounded is-light">
<span class="icon"><i class="fas fa-envelope"></i></span>
<span>Contact</span>
</a>
</div>
</div>
</div>
</div>
</section>
<!-- FOOTER -->
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content has-text-centered">
<p class="is-size-6">
This website is adapted from the
<a href="https://github.com/nerfies/nerfies.github.io">Nerfies template</a>,
which is licensed under a
<a href="https://creativecommons.org/licenses/by-sa/4.0/">
Creative Commons Attribution-ShareAlike 4.0 International License.
</a>
</p>