Thanks for the amazing work of Wei Chen, Xiaogang Jia, Mingfei Wu, and Zhengfa Liang. The goal of this repo is to implement and reproduce the paper Multi-Dimensional Cooperative Network for Stereo Matching which published on ICRA 2022. Original paper could be found via the following links:
I follow the data preparation of PSMNet to download the SceneFlow dataset.
- SceneFlow includes three datasets: flything3d, driving and monkaa.
- You can train MDCNet with some of three datasets, or all of them.
- the following is the describtion of six subfolder.
# the disp folder of Driving dataset
driving_disparity
# the image folder of Driving dataset
driving_frames_cleanpass
# the disp folder of Flything3D dataset
frames_cleanpass
# the image folder of Flything3D dataset
frames_disparity
# the disp folder of Monkaa dataset
monkaa_disparity
# the image folder of Monkaa dataset
monkaa_frames_cleanpass
I pretrain MDCNet on SceneFlow dataset for 24 epochs.
bash scripts/sceneflow_mdcnet.sh
I finetune MDCNet on KITTI 2015 dataset for 300 epochs. Split 80% data for training and 20% for validation.
bash scripts/kitti15_mdcnet.sh
I finetune MDCNet on KITTI 2012 dataset for 300 epochs. Split 80% data for training and 20% for validation.
bash scripts/kitti12_mdcnet.sh
You can inference on kitti raw dataset.
python test_loop.py \
--datapath /home/bsplab/Documents/dataset_kitti/train/2011_09_26_drive_0011_sync \
--output_dir output \
--loadmodel results/kitti15_mdcnet/checkpoint.tar
Train on RTX 2080Ti
Matching Cost Computation | Cost Aggregation | SceneFlow (EPE) | KITTI 2015 D1-all (%) | KITTI 2012 D1-all (%) | Time(s) |
---|---|---|---|---|---|
Correlation | Unet/2D | 1.647 | 3.93% | 5.08% | 0.043 |
Concat | Hourglass/3D | 1.121 | 2.13% | 2.56% | 0.243 |
Correlation+Concat | Unet/2D+DCU+Hourglass/3D | 1.351 | 3.16% | 3.91% | 0.073 |
In this implementation, I use parts of the implementations of the following works:
- PSMNet by Jia-Ren Chang
- GWCNet by Xiaoyang Guo
- CasStereoNet by Xiaodong Gu
- AnyNet by Yan Wang
Thanks for the respective authors for sharing their amazing works.