This library provides a comprehensive framework for dense matching tasks, including implementation, training, and evaluation of various deep learning models. It supports numerous standard datasets for geometric, optical flow, and semantic matching, along with scripts for performance analysis and score generation. The library is designed for researchers and practitioners in computer vision working on dense correspondence problems.

How It Works

The library implements several state-of-the-art dense matching architectures, including GLU-Net, PWC-Net, PDC-Net, and WarpC, often incorporating novel loss functions like Probabilistic Warp Consistency (PWC) and Globally Optimized Correspondence Volumes (GOCor). These methods leverage PyTorch and often utilize multi-stage or multi-scale approaches for improved accuracy, especially in challenging scenarios with large appearance changes or viewpoint variations.

Quick Start & Requirements

• Install: Create a conda environment (conda create -n dense_matching_env python=3.7, conda activate dense_matching_env), then pip install numpy opencv-python torch torchvision matplotlib imageio jpeg4py scipy pandas tqdm gdown pycocotools timm.
• Dependencies: CUDA is required. CuPy is a required dependency for CUDA operations (pip install cupy or pip install cupy-cudaXX for specific CUDA versions). Git submodules must be initialized (git submodule update --init --recursive).
• Configuration: Create admin/local.py to specify dataset paths.
• Pre-trained Models: Download using assets/download_pre_trained_models.sh.
• Documentation: Official Docs

Highlighted Details

• Implements and provides pre-trained models for multiple CVPR/NeurIPS/ICCV award-winning methods (GLU-Net, PDC-Net, WarpC, GOCor).
• Supports a wide array of benchmark datasets: MegaDepth, RobotCar, ETH3D, HPatches, KITTI, Sintel, TSS, PF-Pascal, PF-Willow, Spair-71k, Caltech-101, YFCC100M, ScanNet.
• Includes detailed evaluation scripts for correspondence and pose estimation tasks.
• Offers a flexible training framework with various dataset integrations and custom training options.

Maintenance & Community

The project is primarily maintained by Prune Truong. Updates are noted in the changelog, with recent activity including refactoring and new demo additions.

Licensing & Compatibility

The repository does not explicitly state a license in the README. However, it mentions borrowing code from other projects, which may have their own licenses. Users should verify licensing for commercial use.

Limitations & Caveats

• The installation instructions specify Python 3.7 and PyTorch 1.0, which are older versions. Compatibility with newer versions may require adjustments.
• CuPy installation requires matching the CUDA version, which can be a point of friction.
• Data preparation for various datasets involves downloading large files and specific directory structures.