[RFC] Proposal to Include mmstereo (Stereo Matching Library) into OpenMMLab

[shenshanf]

Hello! I am a researcher in stereo vision and am currently developing mmstereo, a stereo matching algorithm library based on MMEngine. This project aims to provide a unified framework for training and evaluating stereo matching algorithms. The project is currently under development and not yet open-sourced, but we plan to join the OpenMMLab ecosystem and make it open source after completing the core functionalities.

The project has implemented the following core features:

1. Complete Data Processing Pipeline

   • Stereo matching dataset loaders
   • Common data processing pipeline implementations
   • Basic data structures (DispMap, ErrorMap, etc.)

2. Model Design and Implementation

   • Basic architecture for stereo matching models
   • Common modules (cost volume construction, cost aggregation, etc.)
   • Loss function implementations

3. Evaluation and Visualization

   • Evaluation system for stereo matching models
   • Visualization tools for disparity maps and error maps

Near-term Development Plans:

1. Comprehensive testing of implemented modules
2. Support for more stereo matching algorithms
3. Documentation and examples improvement
4. Code structure standardization for open source

I noticed that OpenMMLab currently doesn't have a dedicated library for stereo matching algorithms. Given the importance of stereo matching in autonomous driving, robotics, and other fields, I believe mmstereo could be a valuable addition to the OpenMMLab ecosystem.

Questions:

1. Is OpenMMLab interested in including a stereo matching algorithm library?
2. What are the specific requirements for the project before open-sourcing?
3. What would be the next steps in the process?

Thank you for your attention!