From 58e9277ea7989b4d9757db7928ec21d9abd0d064 Mon Sep 17 00:00:00 2001 From: Haruko386 Date: Tue, 22 Sep 2026 17:48:43 +0800 Subject: [PATCH] docs: add copyright notice and refresh changelog --- CHANGELOG.md | 116 ++++++++++++++++++++++++++++++--------------------- NOTICE.md | 22 ++++++++++ 2 files changed, 90 insertions(+), 48 deletions(-) create mode 100644 NOTICE.md diff --git a/CHANGELOG.md b/CHANGELOG.md index b1b7eed..99a4805 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -1,65 +1,85 @@ # Changelog -All notable changes to this project will be documented in this file. +This file records the notable changes in ApDepth from the complete Git history. Release dates follow the commit dates referenced by the repository's lightweight tags. ----------- +## Unreleased — ApDepth V2.1 -## [ApDepth-v2-1] - 2026-04-06 +### Added -Framework +- Added direct Stage 2 fine-tuning from the original Stable Diffusion 2.1 checkpoint, without requiring Stage 1 training. +- Added Spatially Debiased Window Transport (SDWT) for geometry-aware local depth supervision, with gradient loss and an optional low-weight frequency loss. +- Added separate configurations for SDWT training, SDWT with FFT, and component-level loss ablations. +- Added optional VAE decoder fine-tuning and checkpoint save/load support. +- Added SafeTensors checkpoint export, recursive folder inference, and resumable dataset inference. +- Added VKITTI far-depth post-training and a gradual frequency-loss transition. -## Goodbye 2025, 2026 is near 🎉 +### Changed -## [ApDepth-v1-2] - 2025-11-25 -> We present a new Loss_Function for our two stage training strats. We use `MSE_Loss` for stage one to -> learn the whole information and stage two uses our propoesd `Latent_Frequence_Loss` to refine edge detail +- Kept the original two-stage and Stage 1 workflows available while documenting three supported training strategies. +- Simplified the repository layout, refreshed model configuration and checkpoint links, and updated CI and Docker packaging. -**Training Framework:** +## [ApDepth-V2-0] — 2026-04-06 -Framework +- Renamed the training and inference package from `marigold` to `apdepth`. +- Reorganized trainer, model configuration, checkpoint directories, and command-line documentation around the ApDepth workflow. +- Updated model download scripts and expanded Docker, training, and inference instructions. -**Inference Framework:** +### Architecture -Framework +

+ ApDepth V2.0 training architecture +

-## [ApDepth-v1-1] - 2025-10-09 -> We use a pretrained model as a "teacher" to "teach" `stable diffusion 2-1` to generate Depth Map. +## [ApDepth-V1-2] — 2026-04-02 -**Training Framework:** +- Extended the two-stage objective from latent frequency refinement to combined latent, pixel, gradient, and surface-normal supervision. +- Added Sobel-based edge masks and revised the evaluation and loss strategies. +- Added DIODE training data support and updated dataset sampling. +- Switched the teacher backbone to Depth Anything V2 Giant and refreshed the training configuration. -Framework +### Architecture -**Inference Framework:** +

+ ApDepth V1.2 training architecture +

-Framework +

+ ApDepth V1.2 inference architecture +

-## [ApDepth-v1-0] - 2025-09-23 -### Added ->We change Marigold from `Stochastic multi-step generation` to `Deterministic one-step perception`. And achieve the same result. - -Framework - -### Result -| Method | # Training Samples | | NYUv2 | | KITTI | | ETH3D | | ScanNet | | DIODE | | Avg. Rank | -| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | -| | **Real** | **Synthetic** | **AbsRel↓** | **δ1↑** | **AbsRel↓** | **δ1↑** | **AbsRel↓** | **δ1↑** | **AbsRel↓** | **δ1↑** | **AbsRel↓** | **δ1↑** | | -| Marigold (w/o) | —* | 74K | 6.0 | 95.9 | 10.5 | 90.4 | 7.1 | 95.1 | 6.9 | 94.5 | 31.0 | 77.2 | - | -| **Marigold** | | | 5.5 | 96.4 | **9.9** | **91.6** | **6.5** | **96.0** | 6.4 | 95.1 | 30.8 | **77.3** | - | -| Ours | —* | 74K | **5.3** | **96.5** | 10.7 | 89.3 | **6.5** | 95.8 | **6.0** | **96.3** | **30.0** | 77.0 | - | - -## [Marigold-depth-v1-0] - 2023-12-04 -The Result is showed in table -| Method | # Training Samples | | NYUv2 | | KITTI | | ETH3D | | ScanNet | | DIODE | | Avg. Rank | -| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | -| | **Real** | **Synthetic** | **AbsRel↓** | **δ1↑** | **AbsRel↓** | **δ1↑** | **AbsRel↓** | **δ1↑** | **AbsRel↓** | **δ1↑** | **AbsRel↓** | **δ1↑** | | -| DiverseDepth [56] | 320K | — | 11.7 | 87.5 | 19.0 | 70.4 | 22.8 | 69.4 | 10.9 | 88.2 | 37.6 | 63.1 | 7.6 | -| MiDaS [35] | 2M | — | 11.1 | 88.5 | 23.6 | 63.0 | 18.4 | 75.2 | 12.1 | 84.6 | 33.2 | 71.5 | 7.3 | -| LeReS [57] | 300K | 54K | 9.0 | 91.6 | 14.9 | 78.4 | 17.1 | 77.7 | 9.1 | 91.7 | 27.1 | 76.6 | 5.2 | -| Omnidata [13] | 11.9M | 310K | 7.4 | 94.5 | 14.9 | 83.5 | 16.6 | 77.8 | 7.5 | 93.6 | 33.9 | 74.2 | 4.8 | -| HDN [60] | 300K | — | 6.9 | 94.8 | 11.5 | 86.7 | 12.1 | 83.3 | 8.0 | 93.9 | 24.6 | **78.0** | 3.2 | -| DPT [36] | 1.2M | 188K | 9.8 | 90.3 | 10.0 | 90.1 | 7.8 | 94.6 | 8.2 | 93.4 | **18.2** | 75.8 | 3.9 | -| Marigold (w/o) | —* | 74K | 6.0 | 95.9 | 10.5 | 90.4 | 7.1 | 95.1 | 6.9 | 94.5 | 31.0 | 77.2 | 2.5 | -| **Marigold** | | | **5.5** | **96.4** | **9.9** | **91.6** | **6.5** | **96.0** | **6.4** | **95.1** | 30.8 | 77.3 | **1.4** | -
- -![cover](doc/teaser_collage_transparant.png) +## [ApDepth-V1-1] — 2025-11-01 + +- Introduced teacher-guided Stable Diffusion 2.1 fine-tuning with Depth Anything V2 depth priors. +- Updated the training pipeline, datasets, evaluation path, and V1.1 configurations for the teacher-guided workflow. + +### Architecture + +

+ ApDepth V1.1 training architecture +

+ +

+ ApDepth V1.1 inference architecture +

+ +## [ApDepthv1] — 2025-10-09 + +- Added the first frequency-domain loss experiments and an eight-channel UNet input path. +- Added training utilities, example assets, and the initial training and inference workflow documentation. + +## [ApDepth] — 2025-09-23 + +- Released the first ApDepth implementation. +- Converted the Marigold-style stochastic multi-step process into deterministic single-step depth prediction based on Stable Diffusion 2.1. +- Added the initial training, inference, evaluation, and dataset pipeline. + +## [Backup] — 2025-09-13 + +- Archived the early repository bootstrap and imported baseline implementation before the first ApDepth release. + +[ApDepth-V2-0]: https://github.com/Haruko386/ApDepth/releases/tag/ApDepth-V2-0 +[ApDepth-V1-2]: https://github.com/Haruko386/ApDepth/releases/tag/ApDepth-V1-2 +[ApDepth-V1-1]: https://github.com/Haruko386/ApDepth/releases/tag/ApDepth-V1-1 +[ApDepthv1]: https://github.com/Haruko386/ApDepth/releases/tag/ApDepthv1 +[ApDepth]: https://github.com/Haruko386/ApDepth/releases/tag/ApDepth +[Backup]: https://github.com/Haruko386/ApDepth/releases/tag/Backup diff --git a/NOTICE.md b/NOTICE.md new file mode 100644 index 0000000..2f6b398 --- /dev/null +++ b/NOTICE.md @@ -0,0 +1,22 @@ +# Copyright Notice + +

+ + cvrs-ys801 + +      + + Haruko386 + +

+ +

+ cvrs-ys801 +  &  + Haruko386 +

+ +ApDepth and this accompanying work are jointly created and copyrighted by +[cvrs-ys801](https://github.com/cvrs-ys801) and +[Haruko386](https://github.com/Haruko386). Use and redistribution are governed +by the terms in [LICENSE.txt](LICENSE.txt).