Pyramid Stereo Matching Network

This repository contains the code (in PyTorch) for "Pyramid Stereo Matching Network" paper (CVPR 2018) by Jia-Ren Chang and Yong-Sheng Chen.

changelog

2020/12/20: Update PSMNet: now support torch 1.6.0 / torchvision 0.5.0 and python 3.7, Removed inconsistent indentation.

2020/12/20: Our proposed Real-Time Stereo can be found here Real-time Stereo.

Citation

@inproceedings{chang2018pyramid,
  title={Pyramid Stereo Matching Network},
  author={Chang, Jia-Ren and Chen, Yong-Sheng},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={5410--5418},
  year={2018}
}

Contents

1. Introduction
2. Usage
3. Results
4. Contacts

Introduction

Recent work has shown that depth estimation from a stereo pair of images can be formulated as a supervised learning task to be resolved with convolutional neural networks (CNNs). However, current architectures rely on patch-based Siamese networks, lacking the means to exploit context information for finding correspondence in illposed regions. To tackle this problem, we propose PSMNet, a pyramid stereo matching network consisting of two main modules: spatial pyramid pooling and 3D CNN. The spatial pyramid pooling module takes advantage of the capacity of global context information by aggregating context in different scales and locations to form a cost volume. The 3D CNN learns to regularize cost volume using stacked multiple hourglass networks in conjunction with intermediate supervision.

Usage

Dependencies

• Python 3.7
• PyTorch(1.6.0+)
• torchvision 0.5.0
• KITTI Stereo
• Scene Flow

Usage of Scene Flow dataset
Download RGB cleanpass images and its disparity for three subset: FlyingThings3D, Driving, and Monkaa.
Put them in the same folder.
And rename the folder as: "driving_frames_cleanpass", "driving_disparity", "monkaa_frames_cleanpass", "monkaa_disparity", "frames_cleanpass", "frames_disparity".

Notice

1. Warning of upsample function in PyTorch 0.4.1+: add "align_corners=True" to upsample functions.
2. Output disparity may be better with multipling by 1.17. Reported from issues #135 and #113.

Train

As an example, use the following command to train a PSMNet on Scene Flow

python main.py --maxdisp 192 \
               --model stackhourglass \
               --datapath (your scene flow data folder)\
               --epochs 10 \
               --loadmodel (optional)\
               --savemodel (path for saving model)

As another example, use the following command to finetune a PSMNet on KITTI 2015

python finetune.py --maxdisp 192 \
                   --model stackhourglass \
                   --datatype 2015 \
                   --datapath (KITTI 2015 training data folder) \
                   --epochs 300 \
                   --loadmodel (pretrained PSMNet) \
                   --savemodel (path for saving model)

You can also see those examples in run.sh.

Evaluation

Use the following command to evaluate the trained PSMNet on KITTI 2015 test data

python submission.py --maxdisp 192 \
                     --model stackhourglass \
                     --KITTI 2015 \
                     --datapath (KITTI 2015 test data folder) \
                     --loadmodel (finetuned PSMNet) \

Pretrained Model

※NOTE: The pretrained model were saved in .tar; however, you don't need to untar it. Use torch.load() to load it.

Update: 2018/9/6 We released the pre-trained KITTI 2012 model.

Update: 2021/9/22 a pretrained model using torch 1.8.1 (the previous model weight are trained torch 0.4.1)

Test on your own stereo pair

python Test_img.py --loadmodel (finetuned PSMNet) --leftimg ./left.png --rightimg ./right.png

Results

Evaluation of PSMNet with different settings

※Note that the reported 3-px validation errors were calculated using KITTI's official matlab code, not our code.

Results on KITTI 2015 leaderboard

Leaderboard Link

Qualitative results

Left image

Predicted disparity

Error

Visualization of Receptive Field

We visualize the receptive fields of different settings of PSMNet, full setting and baseline.

Full setting: dilated conv, SPP, stacked hourglass

Baseline: no dilated conv, no SPP, no stacked hourglass

The receptive fields were calculated for the pixel at image center, indicated by the red cross.

Contacts

[email protected]

Any discussions or concerns are welcomed!