This project aims at providing the necessary building blocks for easily creating dense matching models, e.g. Stereo Matching, Scene Flow using PyTorch 1.7.1 or higher.
Our architecture is based on two wellknown detection framework: mmdetection and maskrcnn-benchmark. By integating their major features, our architecture is suitable for dense matching, and achieves robust performance！
-
Modular Design
We decompose the matching framework into different components and one can easily construct a customized matching framework by combining different modules.
-
Support of multiple frameworks out of box
The toolbox directly supports popular and contemporary matching frameworks, e.g. AcfNet, GANet, GwcNet, PSMNet, GC-Net etc.
-
visualization
The toolbox provides various visualization, e.g., cost volume / distribution visualization, colorful disparity map and so on.
-
State of the art
The toolbox stems from the codebase developed by AcfNet team who ranked 1st at KITTI 2012 and 3rd at KITTI 2015 in 2019, and we keep pushing it forward.
- Multi-GPU training and inference
- Mixed precision training: trains faster with less GPU memory on NVIDIA tensor cores.
- Batched inference: can perform inference using multiple images per batch per GPU
Please refer to INSTALL.md for installation and dataset preparation.
Please see GETTING_STARTED.md for the basic usage of DenseMatchingBenchmark.
- inference
- visulization tool(for cost volume)
All our reimplemented methods will provide checkpoint(including all training details, e.g., TensorBoard Record) in corresponding config file ResultOf{ModelName}
(x): means the result in original paper
| Model | FLOPS | Parameters | FPS | Time(ms) | SceneFlow (EPE) | KITTI 2012 | KITTI 2015 (D1-all) |
|---|---|---|---|---|---|---|---|
| PSMNet | 938.186G | 5.225M | 1.67 | 599.2 | 1.112 (1.090) | 2.33 (2.32) | |
| AcfNet | 1239.0G | 5.559M | 1.38 | 723.1 | 0.831 (0.867) | ||
| StereoNet-8x-single | 78.512G | 399.066K | 19.17 | 52.2 | 1.533 (1.525) | ||
| DeepPruner | |||||||
| AnyNet | 1.476G | 46.987K | 3.190 (~3.2) |
Update Time: 24/04/2020
- Integrate Optical Flow task into our architecture
- Inference a model is avaliable
- Unify the format of module's name
Update Time: 10/11/2020
- Fix a bug in the loss of confidence estimation network, and retrain the AcfNet-Adaptive Network, checkpoints are uploaded here
Update Time: 28/01/2021
- Add demo.py to run a demo.
- Support visualization for cost distribution, run view_cost.py to take a look. We implement an UI for friendly interaction.
- For detailed HowToUse, please refer to GETTING_STARTED.md
Update Time: 31/05/2021
- Support Pytorch1.7.1, not test for all projects if it works, but you can try, if not work or get worse results, please checkout to: 177c56ca1952f54d28e6073afa2c16981113a2af
We appreciate all contributions to improve DenseMatchingBenchmark. Please refer to CONTRIBUTING.md for the contributing guideline.
DenseMatchingBenchmark is an open source project that is contributed by researchers and engineers from various colledges and companies. We appreciate all the contributors who implement their methods or add new features, as well as users who give valuable feedbacks. We wish that the toolbox and benchmark could serve the growing research community by providing a flexible toolkit to reimplement existing methods and develop their own new matching algorithm.
If you use this toolbox or benchmark in your research, please cite this project.
@article{zhang2020adaptive,
title={Adaptive Unimodal Cost Volume Filtering for Deep Stereo Matching},
author={Zhang, Youmin and Chen, Yimin and Bai, Xiao and Yu, Suihanjin and Yu, Kun and Li, Zhiwei and Yang, Kuiyuan},
journal={AAAI},
year={2020}
}
This repo is currently maintained by Youmin Zhang(@youmi-zym) and Yimin Chen (@Minwell-dht)
DenseMatchingBenchmark is released under the MIT license. See LICENSE for additional details.