Revisiting Stereo Depth Estimation From a Sequence-to-Sequence Perspective with Transformers. (ICCV 2021 Oral)
This is the official repo for our work Revisiting Stereo Depth Estimation From a Sequence-to-Sequence Perspective with Transformers.
Fine-tuned result on street scene:
Generalization to medical domain when trained only on synthetic data:
If you find our work relevant, please cite
@InProceedings{Li_2021_ICCV,
author = {Li, Zhaoshuo and Liu, Xingtong and Drenkow, Nathan and Ding, Andy and Creighton, Francis X. and Taylor, Russell H. and Unberath, Mathias},
title = {Revisiting Stereo Depth Estimation From a Sequence-to-Sequence Perspective With Transformers},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2021},
pages = {6197-6206}
}
sttr-light
for the new model.STereo TRansformer (STTR) revisits stereo depth estimation from a sequence-to-sequence perspective. The network combines conventional CNN feature extractor and long-range relationship capturing module Transformer. STTR is able to relax prior stereo depth estimation networks in three aspects:
STTR performs comparably well against prior work with refinement in Scene Flow and KITTI 2015. STTR is also able to generalize to MPI Sintel, KITTI 2015, Middlebury 2014 and SCARED when trained only on synthetic data.
Two types of attention mechanism are used: self-attention and cross-attention. Self-attention uses context within the same image, while cross-attention uses context across two images. The attention shrinks from global context to local context as the layer goes deeper. Attention in a large textureless area tends to keep attending dominant features like edges, which helps STTR to resolve ambiguity.
Self-Attention
Cross-Attention
We find that only image-image based attention is not enough. Therefore, we opt in relative positional encoding to provide positional information. This allows STTR to use the relative distance from a featureless pixel to dominant pixel (such as edge) to resolve ambiguity. In the following example, STTR starts to texture the center of the table using relative distance, thus strides parallel to the edges start to show.
Feature Descriptor
We observe that the feature extractor before Transformer actually learns without any explicit supervision to classify pixels into two clusters - textured and textureless. We hypothesize that this implicit learnt classification helps STTR to generalize.
Implicit Learnt Classification
We recommend the following steps to set up your environment
Create your python virtual environment by
conda create --name sttr python=3.6 # create a virtual environment called "sttr" with python version 3.6
(Python version >= 3.6)
Install Pytorch: Please follow link here.
(PyTorch version >= 1.5.1)
Other third-party packages: You can use pip to install the dependencies by
pip install -r requirements.txt
(Optional) Install Nvidia apex: We use apex for mixed precision training to accelerate training. To install, please follow instruction here
You can download the pretrained model from the following links.
Models | Link |
---|---|
STTR (Scene Flow pretrained) | Download link |
STTR (KITTI finetuned) | Download link |
STTR-light (Scene Flow pretrained) | Download link |
jvda
stereo-transformer
|_ dataset (dataloder)
|_ module (network modules, including loss)
|_ utilities (training, evaluation, inference, logger etc.)
Please see sample_data folder for details. We keep the original data folder structure from the official site. If you need to modify the existing structure, make sure to modify the dataloader.
Scene Flow
SCENE_FLOW
|_ RGB_finalpass
|_ TRAIN
|_ A
|_0000
|_ disparity
|_ TRAIN
|_ A
|_0000
|_ occlusion
|_ TRAIN
|_ left
MPI Sintel
MPI_Sintel
|_ training
|_ disparities
|_ final_left
|_ final_right
|_ occlusions (occlusions of left border of objects)
|_ outofframe (occlusion of left border of images)
KITTI 2015
KITTI_2015
|_ training
|_ disp_occ_0 (disparity including occluded region)
|_ image_2 (left image)
|_ image_3 (right image)
MIDDLEBURY_2014
MIDDLEBURY_2014
|_ trainingQ
|_ Motorcycle (scene name)
|_ disp0GT.pfm (left disparity)
|_ disp1GT.pfm (right disparity)
|_ im0.png (left image)
|_ im1.png (right image)
|_ mask0nocc.png (left occlusion)
|_ mask1nocc.png (right occlusion)
SCARED
SCARED
|_ training
|_ disp_left
|_ img_left
|_ img_right
|_ occ_left
If you don't have a GPU, you can use Google Colab:
If you have a GPU and want to run locally:
stereo-transformer
folder.sh scripts/pretrain.sh
--dataset_directory
argument in the .sh
file to where Scene Flow data is stored, i.e. replace PATH_TO_SCENEFLOW
sh scripts/kitti_finetune.sh
--dataset_directory
argument in the .sh
file to where KITTI data is stored, i.e. replace PATH_TO_KITTI
stereo-transformer
folder.sh scripts/kitti_toy_eval.sh
stereo-transformer
folder.The result of STTR may vary by a small fraction depending on the trial, but it should be approximately the same as the tables below.
Expected result of STTR (sceneflow_pretrained_model.pth.tar
) and STTR-light (sttr_light_pretrained_model.pth.tar
).
Sceneflow
3px Error | EPE | Occ IOU | |
---|---|---|---|
STTR | 1.26 | 0.45 | 0.92 |
STTR-light | 1.54 | 0.50 | 0.97 |
Generalization without fine-tuning.
MPI Sintel | KITTI 2015 | Middleburry-Q | SCARED | |||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|
3px Error | EPE | Occ IOU | 3px Error | EPE | Occ IOU | 3px Error | EPE | Occ IOU | 3px Error | EPE | Occ IOU | |
STTR | 5.75 | 3.01 | 0.86 | 6.74 | 1.50 | 0.98 | 6.19 | 2.33 | 0.95 | 3.69 | 1.57 | 0.96 |
STTR-light | 5.82 | 2.95 | 0.69 | 7.20 | 1.56 | 0.95 | 5.36 | 2.05 | 0.76 | 3.30 | 1.19 | 0.89 |
Expected 3px error result of kitti_finetuned_model.pth.tar
Dataset | 3px Error | EPE |
---|---|---|
KITTI 2015 training | 0.79 | 0.41 |
KITTI 2015 testing | 2.01 | N/A |
I don't see occlusion from Scene Flow dataset. What should I do?
Scene Flow dataset can be downloaded at https://lmb.informatik.uni-freiburg.de/resources/datasets/SceneFlowDatasets.en.html. However, you may notice that the Full datasets has disparity and images, but not occlusion. What you need to do is to download the occlusion from the DispNet/FlowNet2.0 dataset subsets and use the provided training list on the right named train to only use the subset of Full datasets with the occlusion data. Please see utilities/subsample_sceneflow.py
for more details of subsampling the Full datasets.
How much memory does it require to train/inference?
We provide a flexible design to accommodate different hardware settings.
downsample
parameter to reduce memory consumption at the cost of potential performance degradation.dataset/scene_flow.py
.What are occluded regions?
"Occlusion" means pixels in the left image do not have a corresponding match in right images. Because objects in right image are shifted to the left a little bit compared to the right image, thus pixels in the following two regions generally do not have a match:
Why there are black patches in predicted disparity with values 0?
The disparities of occluded region are set to 0.
Why do you read disparity map including occluded area for KITTI during training?
We use random crop as a form of augmentation, thus we need to recompute occluded regions again. The code for computing occluded area can be found in dataset/preprocess.py.
How to reproduce feature map visualization in Figure 4 of the paper?
The feature map is taken after the first LayerNorm in Transformer. We use PCA trained on the first and third layer to reduce the dimensionality to 3.
This project is under the Apache 2.0 license. Please see LICENSE for more information.
We try out best to make our work easy to transfer. If you see any issues, feel free to fork the repo and start a pull request.
Special thanks to authors of SuperGlue, PSMNet and DETR for open-sourcing the code. We also thank GwcNet, GANet, Bi3D, AANet for open-sourcing the code. We thank Xiran for MICCAI pre-processing.