Hierarchical Deep Stereo Matching on High Resolution Images

[project webpage]

Qualitative results on Middlebury:

Performance on Middlebury benchmark (y-axis: error, the lower the better):

Able to handle large view variation of high-res images (as a submodule in Open4D, CVPR 2020):

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

• Middlebury model (more robust)
• KITTI model (specified to KITTI)
• Google drive (backup link)

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

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

1. Download and extract training data in folder /d/. Training data include Middlebury train set, HR-VS, KITTI-12/15, ETH3D, and SceneFlow.
2. Run

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --maxdisp 384 --batchsize 28 --database /d/ --logname log1 --savemodel /somewhere/  --epochs 10

3. 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.