XiandaGuo/OpenStereo

OpenStereo: A Comprehensive Benchmark for Stereo Matching

What it solves

OpenStereo provides a unified, flexible, and extensible framework for stereo matching, which is the process of estimating depth from two slightly different images of the same scene. It addresses the need for a comprehensive benchmark by integrating numerous datasets and state-of-the-art models into a single environment for easier training, testing, and comparison.

How it works

The project acts as a comprehensive library that supports 17 popular stereo datasets (such as KITTI, SceneFlow, and StereoCarla) and reproduces several state-of-the-art (SOTA) stereo matching methods. It utilizes PyTorch's Distributed Data Parallel (DDP) for efficient training and testing, employs Auto Mixed Precision (AMP) to optimize performance, and integrates TensorRT for high-performance deployment.

Who it’s for

It is designed for academic researchers and developers working on computer vision, specifically those focused on depth estimation, stereo matching, and autonomous driving perception.

Highlights

  • Extensive Dataset Support: Compatible with 17 major stereo datasets, including specialized driving datasets like StereoCarla.
  • Model Zoo: Includes reproductions of various SOTA methods and integrated models like FoundationStereo.
  • Performance Optimizations: Built-in support for DDP, AMP, and TensorRT for faster inference and training.
  • Strong Baselines: Includes the StereoBase model, which achieved a top rank on the KITTI15 leaderboard.

Related

  • Project
    yzfzzz/depth-detect

    A high-performance C++ and TensorRT framework that fuses YOLO object detection with monocular/binocular depth estimation to track objects and estimate their motion states.

  • Project
    modelscope/ms-swift

    A large model and multimodal fine-tuning and deployment framework that supports over 1,000 models and provides a full pipeline from training to deployment.

  • Project
    ShiqiYu/OpenGait

    A flexible and extensible gait analysis framework that provides a comprehensive benchmark and model zoo for identifying individuals through walking patterns using silhouettes, skeletons, and LiDAR data.

  • Project
    Thinklab-SJTU/ThinkMatch

    A research framework for developing and benchmarking deep graph matching algorithms to find node-to-node correspondences between graphs.

  • Project
    OpenBMB/UltraEval-Audio

    A unified open-source framework for evaluating large audio foundation models, supporting both speech understanding and generation across 34 benchmarks.