Update README.md

README.md

@@ -1,272 +1,98 @@
-# Model Card for DINOv2-S/B/L/g
+---
+license: cc-by-nc-4.0
+task_categories:
+  - depth-estimation
+tags:
+  - stereo-matching
+  - disparity
+  - stereo4d
+  - foundationstereo
+pretty_name: FFS Stereo4D
+size_categories:
+  - 100K<n<1M
+---
+
+# FFS Stereo4D
+
+Disparity maps for stereo matching, generated from the [Stereo4D](https://github.com/niconielsen32/Stereo4D) dataset using [FoundationStereo](https://github.com/NVlabs/FoundationStereo).
+
+## Dataset Structure
 
-These are Vision Transformer models trained following the method described in the papers:
-"DINOv2: Learning Robust Visual Features without Supervision"
-and
-"Vision Transformers Need Registers".
-
-We provide 8 models: 
-- 1 ViT-g trained from scratch with 3 ViT-S/B/L models distilled from the ViT-g, without registers.
-- 1 ViT-g trained from scratch with 3 ViT-S/B/L models distilled from the ViT-g, with registers.
-
-## Model Details
-The model takes an image as input and returns a class token and patch tokens, and optionally 4 register tokens.
-
-The embedding dimension is: 
-- 384 for ViT-S.
-- 768 for ViT-B.
-- 1024 for ViT-L.
-- 1536 for ViT-g.
-
-The models follow a Transformer architecture, with a patch size of 14. In the case of registers, we add 4 register tokens, learned during training, to the input sequence after the patch embedding.
-
-For a 224x224 image, this results in 1 class token + 256 patch tokens, and optionally 4 register tokens.
-
-The models can accept larger images provided the image shapes are multiples of the patch size (14). 
-If this condition is not verified, the model will crop to the closest smaller multiple of the patch size.
-
-### Model Description
-
-- **Developed by:** Meta AI
-- **Model type:** Vision Transformer
-- **License:** Apache License 2.0
-
-- **Repository:** https://github.com/facebookresearch/dinov2
-- **Paper:** https://arxiv.org/abs/2304.07193
-- **Demo:** https://dinov2.metademolab.com/
-
-## Uses
-
-The models are vision backbones providing multi-purpose features for downstream tasks.
-
-### Direct Use
-
-The models can be used without fine-tuning, with downstream classifiers as simple as linear layers, to obtain competitive results:
-- on depth estimation, semantic segmentation, using linear layers.
-- on image classification, using k-NN classifiers on the class token.
-- on image classification, with logistic regression classifiers applied on the class token.
-- on image classification, with a linear layer applied on the class token and the average of the patch tokens.
-- on image retrieval using nearest neighbors.
-
-### Downstream Use
-
-It is technically possible to perform fine-tuning on the models, for small gains (we measured +2% on ImageNet-1k classification). 
-We recommend keeping this as a very last step and only when necessary, as the features already provide good performance out-of-the-box.
-
-## Bias, Risks, and Limitations
-
-Despite improvements thanks to the training method not using annotations, we still observe significant biases in our models toward rich households from Western countries.
-
-### Recommendations
-
-We expect fine-tuning will increase the biases in the features produced by the model as they will be tuned to the fine-tuning labels.
-
-## How to Get Started with the Model
-
-Use the code below to get started with the model.
-
-```python
-import torch
-
-# DINOv2
-dinov2_vits14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vits14')
-dinov2_vitb14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitb14')
-dinov2_vitl14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitl14')
-dinov2_vitg14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitg14')
-
-# DINOv2 with registers
-dinov2_vits14_reg = torch.hub.load('facebookresearch/dinov2', 'dinov2_vits14_reg')
-dinov2_vitb14_reg = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitb14_reg')
-dinov2_vitl14_reg = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitl14_reg')
-dinov2_vitg14_reg = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitg14_reg')
+```
+data/train/
+  metadata.csv
+  disparity/
+    0000000/
+      {vid_id}_frame_{frame_idx:06d}.png
+    0000001/
+      ...
 ```
 
-## Training Details
+- **Disparity images**: 16-bit 784×784 PNG files storing per-pixel disparity values. You can read by following: https://github.com/NVlabs/FoundationStereo/blob/master/scripts/vis_dataset.py
+- **metadata.csv**: Links each disparity image back to its source YouTube video.
 
-### Training Data
+### Metadata Columns
 
-- **Training data:** LVD-142M (see paper)
-- **Training regime:** fp16 using PyTorch-FSDP mixed-precision.
+| Column | Description |
+|---|---|
+| `file_name` | Relative path to the disparity image |
+| `vid_id` | Clip identifier (matches the `.npz` calibration file) |
+| `frame_idx` | Frame index in the rectified stereo output |
+| `youtube_video_id` | YouTube video ID of the source 360 video |
+| `timestamp_us` | Timestamp in microseconds in the original video |
+| `timestamp_sec` | Timestamp in seconds |
+| `video_frame_index` | Estimated frame number in the original video |
+| `fps` | FPS of the source video |
 
-### Training Procedure 
+## Retrieving Source RGB Frames
 
-- **Training objective:**
-  - DINO self-distillation loss with multi-crop
-  - iBOT masked-image modeling loss
-  - KoLeo regularization on [CLS] tokens
-- **Architectures:**
-  - ViT-S (21M params): Patch size 14, embedding dimension 384, 6 heads, MLP FFN
-  - ViT-B (86M params): Patch size 14, embedding dimension 768, 12 heads, MLP FFN
-  - ViT-L (0.3B params): Patch size 14, embedding dimension 1024, 16 heads, MLP FFN
-  - ViT-g (1.1B params): Patch size 14, embedding dimension 1536, 24 heads, SwiGLU FFN
-- **Distillation:**
-  - Distillation follows the standard DINOv2 pretraining procedure, except the teacher is a pretrained ViT-g, frozen.
+This dataset contains **disparity maps only**. Due to the copyrights of these videos, users need to download on your own behalf. The corresponding left/right RGB stereo pairs can be recovered by:
 
-## Evaluation
+1. Following [stereo4d toolkit](https://github.com/Stereo4d/stereo4d-code) to download the YouTube video using `youtube_video_id`.
+2. Seek to `timestamp_sec` (or `video_frame_index`) to locate the source frame.
+3. Apply equirectangular rectification using the Stereo4D calibration `.npz` files to obtain the left and right perspective images.
 
-We refer users to the associated papers for the evaluation protocols.
+## Generation Pipeline
 
-<table>
-  <tr>
-    <th colspan="2"></th>
-    <th colspan="3">ImageNet-1k</th>
-    <th>NYU-Depth v2</th>
-    <th>SUN-RGBD</th>
-    <th>ADE20k</th>
-    <th>iNaturalist 2018</th>
-    <th>Oxford-H</th>
-  </tr>
-  <tr>
-    <th rowspan="2">model</th>
-    <th rowspan="2">with <br /> registers</th>
-    <th>classif. (acc)</th>
-    <th>classif. (acc)</th>
-    <th>classif. V2 (acc)</th>
-    <th>depth (RMSE)</th>
-    <th>depth (RMSE)</th>
-    <th>segm. (mAP)</th>
-    <th>classif. (acc)</th>
-    <th>retrieval (mAP)</th>
-  </tr>
-  <tr>
-    <!-- <th>^</th> -->
-    <th>k-NN</th>
-    <th>linear</th>
-    <th>linear</th>
-    <th>linear<br />4 layers</th>
-    <th>NYU-D transfer</th>
-    <th>multiscale</th>
-    <th>linear</th>
-    <th>nearest neighbor</th>
-  </tr>
-  <tr>
-    <td>ViT-S/14</td>
-    <td align="center">:x:</td>
-    <td align="right">79.0%</td>
-    <td align="right">81.1%</td>
-    <td align="right">70.8%</td> 
-    <td align="right">0.417</td> 
-    <td align="right">0.431</td> 
-    <td align="right">47.2</td> 
-    <td align="right">69.5%</td> 
-    <td align="right">43.2</td> 
-  </tr>
-  <tr>
-    <td>ViT-S/14</td>
-    <td align="center">:white_check_mark:</td>
-    <td align="right">79.1%</td>
-    <td align="right">80.9%</td>
-    <td align="right">71.0%</td> 
-    <td align="right">N/A</td> 
-    <td align="right">N/A</td> 
-    <td align="right">N/A</td> 
-    <td align="right">67.6%</td> 
-    <td align="right">39.5</td> 
-  </tr>
-  <tr>
-    <td>ViT-B/14</td>
-    <td align="center">:x:</td>
-    <td align="right">82.1%</td>
-    <td align="right">84.5%</td>
-    <td align="right">74.9%</td>
-    <td align="right">0.362</td> 
-    <td align="right">0.400</td> 
-    <td align="right">51.3</td> 
-    <td align="right">76.3%</td> 
-    <td align="right">49.5</td> 
-  </tr>
-    <td>ViT-B/14</td>
-    <td align="center">:white_check_mark:</td>
-    <td align="right">82.0%</td>
-    <td align="right">84.6%</td>
-    <td align="right">75.6%</td>
-    <td align="right">N/A</td> 
-    <td align="right">N/A</td> 
-    <td align="right">N/A</td> 
-    <td align="right">73.8%</td> 
-    <td align="right">51.0</td> 
-  </tr>
-  <tr>
-    <td>ViT-L/14</td>
-    <td align="center">:x:</td>
-    <td align="right">83.5%</td>
-    <td align="right">86.3%</td>
-    <td align="right">77.6%</td>
-    <td align="right">0.333</td> 
-    <td align="right">0.396</td> 
-    <td align="right">53.1</td> 
-    <td align="right">79.8%</td> 
-    <td align="right">54.0</td> 
-  </tr>
-  <tr>
-    <td>ViT-L/14</td>
-    <td align="center">:white_check_mark:</td>
-    <td align="right">83.8%</td>
-    <td align="right">86.7%</td>
-    <td align="right">78.5%</td>
-    <td align="right">N/A</td> 
-    <td align="right">N/A</td> 
-    <td align="right">N/A</td> 
-    <td align="right">80.9%</td> 
-    <td align="right">55.7</td> 
-  </tr>
-  <tr>
-    <td>ViT-g/14</td>
-    <td align="center">:x:</td>
-    <td align="right">83.5%</td>
-    <td align="right">86.5%</td>
-    <td align="right">78.4%</td>
-    <td align="right">0.298</td> 
-    <td align="right">0.362</td> 
-    <td align="right">53.0</td> 
-    <td align="right">81.6%</td> 
-    <td align="right">52.3</td> 
-  </tr>
-  <tr>
-  <tr>
-    <td>ViT-g/14</td>
-    <td align="center">:white_check_mark:</td>
-    <td align="right">83.7%</td>
-    <td align="right">87.1%</td>
-    <td align="right">78.8%</td>
-    <td align="right">N/A</td> 
-    <td align="right">N/A</td> 
-    <td align="right">N/A</td> 
-    <td align="right">81.5%</td> 
-    <td align="right">58.2</td> 
-  </tr>
-</table>
+1. **Source**: YouTube 360 videos from the Stereo4D dataset.
+2. **Rectification**: Equirectangular frames are rectified and cropped to 1024×1024 perspective stereo pairs.
+3. **Disparity estimation**: FoundationStereo computes dense disparity at 784×784 resolution (resized by `scale=0.765625` of the 1024×1024 input).
 
-## Environmental Impact
+### Camera Parameters
 
-- **Hardware Type:** Nvidia A100
-- **Hours used:** 22,000 for ViT-g, 4,500 for ViT-S distillation, 5,300 for ViT-B distillation, 8,000 for ViT-L distillation
-- **Cloud Provider:** Private infra
-- **Compute Region:** USA
-- **Carbon Emitted:** 7t CO2eq
+The following assumed parameters are used for depth and normal map computation:
 
-#### Hardware
+| Parameter | Value | Notes |
+|---|---|---|
+| Baseline | 0.063 m | From Stereo4D calibration, the assumed interpupillary distance for the VR180 cameras. |
+| HFOV | 60° | Matches `output_hfov` in rectification |
+| fx, fy | ~678.8 px | `width / (2 * tan(HFOV/2))`, for 784×784 |
+| cx, cy | 392 px | Image center |
 
-Nvidia A100 GPUs
+Depth is derived as: `depth = fx * baseline / disparity`.
 
-#### Software
 
-PyTorch 2.0,
-xFormers 0.0.18
+## Citation
 
-**BibTeX**
+If you use this dataset, please consider cite:
 
-```
-@misc{oquab2023dinov2,
-  title={DINOv2: Learning Robust Visual Features without Supervision},
-  author={Oquab, Maxime and Darcet, Timothée and Moutakanni, Theo and Vo, Huy and Szafraniec, Marc and Khalidov, Vasil and Fernandez, Pierre and Haziza, Daniel and Massa, Francisco and El-Nouby, Alaaeldin and Howes, Russell and Huang, Po-Yao and Xu, Hu and Sharma, Vasu and Li, Shang-Wen and Galuba, Wojciech and Rabbat, Mike and Assran, Mido and Ballas, Nicolas and Synnaeve, Gabriel and Misra, Ishan and Jegou, Herve and Mairal, Julien and Labatut, Patrick and Joulin, Armand and Bojanowski, Piotr},
-  journal={arXiv:2304.07193},
-  year={2023}
+```bibtex
+@article{wen2026fastfoundationstereo,
+  title={Fast-FoundationStereo: Real-Time Zero-Shot Stereo Matching},
+  author={Bowen Wen and Shaurya Dewan and Stan Birchfield},
+  journal={CVPR},
+  year={2026}
+}
+@article{wen2025foundationstereo,
+  title={FoundationStereo: Zero-Shot Stereo Matching},
+  author={Wen, Bowen and Trepte, Matthew and Aribido, Joseph and Kautz, Jan and Birchfield, Stan and Wan, Yao},
+  journal={CVPR},
+  year={2025}
 }
-@misc{darcet2023vitneedreg,
-  title={Vision Transformers Need Registers},
-  author={Darcet, Timothée and Oquab, Maxime and Mairal, Julien and Bojanowski, Piotr},
-  journal={arXiv:2309.16588},
-  year={2023}
+@inproceedings{jin2025stereo4d,
+  title={{Stereo4D: Learning How Things Move in 3D from Internet Stereo Videos}},
+  author={Jin, Linyi and Tucker, Richard and Li, Zhengqi and Fouhey, David and Snavely, Noah and Holynski, Aleksander},
+  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
+  year={2025},
 }
 ```