Awesome
Hierarchical Deep Stereo Matching on High Resolution Images
[project webpage]
<img src="./architecture.png" width="800">Qualitative results on Middlebury:
<img src="http://www.contrib.andrew.cmu.edu/~gengshay/wordpress/wp-content/uploads/2019/06/cvpr19-middlebury1-small.gif" width="400">Performance on Middlebury benchmark (y-axis: error, the lower the better):
<img src="./middlebury-benchmark.png" width="400">Able to handle large view variation of high-res images (as a submodule in Open4D, CVPR 2020):
<img src="http://www.contrib.andrew.cmu.edu/~gengshay/wordpress/wp-content/uploads/2020/02/cvpr19-dance.gif" width="800">Requirements
- tested with python 2.7.15 and 3.6.8
- tested with pytorch 0.4.0, 0.4.1 and 1.0.0
- a few packages need to be installed, for eamxple, texttable
Weights
Note: The .tar file can be directly loaded in pytorch. No need to uncompress it.
Inference
Test on CrusadeP and dancing stereo pairs:
CUDA_VISIBLE_DEVICES=3 python submission.py --datapath ./data-mbtest/ --outdir ./mboutput --loadmodel ./weights/final-768px.tar --testres 1 --clean 1.0 --max_disp -1
Evaluate on Middlebury additional images:
CUDA_VISIBLE_DEVICES=3 python submission.py --datapath ./path_to_additional_images --outdir ./output --loadmodel ./weights/final-768px.tar --testres 0.5
python eval_mb.py --indir ./output --gtdir ./groundtruth_path
Evaluate on HRRS:
CUDA_VISIBLE_DEVICES=3 python submission.py --datapath ./data-HRRS/ --outdir ./output --loadmodel ./weights/final-768px.tar --testres 0.5
python eval_disp.py --indir ./output --gtdir ./data-HRRS/
And use cvkit to visualize in 3D.
Example outputs
<img src="data-mbtest/CrusadeP/im0.png" width="400"> left image <img src="mboutput/CrusadeP/capture_000.png" width="400"> 3D projection <img src="mboutput/CrusadeP-disp.png" width="400"> disparity map <img src="mboutput/CrusadeP-ent.png" width="400"> uncertainty map (brighter->higher uncertainty)Parameters
- testres: 1 is full resolution, and 0.5 is half resolution, and so on
- max_disp: maximum disparity range to search
- clean: threshold of cleaning. clean=0 means removing all the pixels.
Data
train/val
- Middlebury (train set and additional images)
- High-res-virtual-stereo (HR-VS)
- KITTI-2012&2015
- SceneFlow
- Eth3D
test
High-res-real-stereo (HR-RS) It has been taken off due to licensing issue. Please use the Argoverse dataset.
Train
- Download and extract training data in folder /d/. Training data include Middlebury train set, HR-VS, KITTI-12/15, ETH3D, and SceneFlow.
- Run
CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --maxdisp 384 --batchsize 28 --database /d/ --logname log1 --savemodel /somewhere/ --epochs 10
- Evalute on Middlebury additional images and KITTI validation set. After 40k iterations, average error on Middlebury additional images excluding Shopvac (perfect+imperfect, 24 stereo pairs in total) with half-res should be around 5.7.
Citation
@InProceedings{yang2019hsm,
author = {Yang, Gengshan and Manela, Joshua and Happold, Michael and Ramanan, Deva},
title = {Hierarchical Deep Stereo Matching on High-Resolution Images},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2019}
}
Acknowledgement
Part of the code is borrowed from MiddEval-SDK, PSMNet, FlowNetPytorch and pytorch-semseg. Thanks SorcererX for fixing version compatibility issues.