RobustStereoMatching
To download checkpoint files for our method, go to https://drive.google.com/drive/folders/1OWNqiLUY4n-RfhOFJzxLYbQv8SUCwvGx?usp=sharing (this link does not have any personal information) and download all files to checkpoint/MCTNet/
To test our trained model:
python test.py --left_image=examples/left.png --right_image=examples/right.png --backbone=1
Not setting --backbone=1 will use the one without backbone.
To train and evaluate our model, please download from:
- SceneFlow - https://lmb.informatik.uni-freiburg.de/resources/datasets/SceneFlowDatasets.en.html
- KITTI2015 - http://www.cvlibs.net/datasets/kitti/eval_scene_flow.php?benchmark=stereo
Our default directory setting is:
data/
FlyingThings3D/
frames_finalpass/
disparity/
KITTI2015/
training/
testing/
projects/
RobustStereoMatching/
train.py
...
To train our model:
./train.sh
To evaluate model (dataset: 1 - SceneFlow, 3 - KITTI2015):
python evaluation.py --whichModel=2 --dataset=1 --backbone=1
To test other models such as GANet (https://github.com/feihuzhang/GANet) or PSMNet (https://github.com/JiaRenChang/PSMNet), please go to their repo and put it in the folder models. GANet will need to be compiled.
To test stereo-constrained attacks:
python pgd_attack.py --whichModel=2 --dataset=1 --backbone=1 --e=0.03 --total_iter=20
To test synthetic patch attacks:
python synthetic_patch_attack.py
Use --test_patch_shift to test generated patch at different disparity levels.