CoEx

PyTorch implementation of our paper:

Correlate-and-Excite: Real-Time Stereo Matching via Guided Cost Volume Excitation
Authors: Antyanta Bangunharcana, Jae Won Cho, Seokju Lee, In So Kweon, Kyung-Soo Kim, Soohyun Kim
IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021

[Project page]

We propose a Guided Cost volume Excitation (GCE) and top-k soft-argmax disparity regression for real-time and accurate stereo matching.

Contents

• Installation
• Datasets
  • Data for demo
  • If you want to re-train the models
  • Data directories
• Demo on KITTI raw data
  • Model zoo
• Re-training the model

Installation

We recommend using conda for installation:

conda env create -f environment.yml
conda activate coex

Datasets

Data for demo

For a demo of our code on the KITTI dataset, download the "[synced+rectified data]" from raw KITTI data. Unzip and place the extracted folders following the directory tree below.

If you want to re-train the models

Sceneflow dataset
Download the finalpass data of the Sceneflow dataset as well as the Disparity data.

KITTI 2015
Download kitti15 dataset, and unzip data_scene_flow.zip, rename it as kitti15, and move it into SceneFlow directory as shown in the tree below.

KITTI 2012
Download kitti12 dataset. Unzip data_stereo_flow.zip, rename it as kitti12, and move it into SceneFlow directory as shown in the tree below.

The paths where the code can find the dataset can be modified inside the config files, but make sure the directory names of driving, flyingthings3d_final, monkaa.

Data directories

In our setup, the dataset is organized as follows

../../data
└── datasets
    ├── KITTI_raw
    |   ├── 2011_09_26
    |   │   ├── 2011_09_26_drive_0001_sync
    |   │   ├── 2011_09_26_drive_0002_sync
    |   |       :
    |   |
    |   ├── 2011_09_28
    |   │   ├── 2011_09_28_drive_0001_sync
    |   │   └── 2011_09_28_drive_0002_sync
    |   |       :
    |   |   :    
    |
    └── SceneFlow
        ├── driving
        │   ├── disparity
        │   └── frames_finalpass
        ├── flyingthings3d_final
        │   ├── disparity
        │   └── frames_finalpass
        ├── monkaa
        │   ├── disparity
        │   └── frames_finalpass
        ├── kitti12
        │   ├── testing
        │   └── training
        └── kitti15
            ├── testing
            └── training

Demo on KITTI raw data

The pretrained KITTI model is already included in './logs'. Run

python demo.py

to perform stereo matching on raw kitti sequence.

For more demo results, checkout our Project page

Re-training the model

To re-train the model, configure './configs/stereo/cfg_yaml', and run

python stereo.py

Citation

If you find our work useful in your research, please consider citing our paper

@inproceedings{bangunharcana2021coex,
  title={Correlate-and-Excite: Real-Time Stereo Matching via Guided Cost Volume Excitation},
  author={Bangunharcana, Antyanta and Cho, Jae Won and Lee, Seokju and Kweon, In So and Kim, Kyung-Soo and Soohyun Kim},
  booktitle={IROS},
  year={2021}
}

Acknowledgements

Part of the code is adopted from previous works: PSMNet, AANet