github.com/open-mmlab/mmtracking ↗
OpenMMLab Video Perception Toolbox. It supports Video Object Detection (VID), Multiple Object Tracking (MOT), Single Object Tracking (SOT), Video Instance Segmentation (VIS) with a unified framework.
Open this visualization on its own page →
Contributors
4
Lines of Code
882
From
2020-07-28
To
2021-01-05
About open-mmlab/mmtracking
MMTracking is an open-source video perception toolbox built by OpenMMLab using PyTorch. It provides a unified framework for four major video analysis tasks: video object detection, multiple object tracking, single object tracking, and video instance segmentation. This is notable as the first open-source platform to consolidate these different tracking and detection paradigms into a single modular system.
The toolbox is built with modular design principles, allowing users to construct custom methods by combining different components. It integrates closely with other OpenMMLab projects, particularly MMDetection, enabling users to leverage any detector by simply modifying configuration files. The framework is optimized for performance, with all operations running on GPUs and training and inference speeds comparable to or faster than alternative implementations. MMTracking includes multiple state-of-the-art methods across its supported tasks, such as ByteTrack and OC-SORT for multi-object tracking, STARK and MixFormer for single object tracking, and MaskTrack R-CNN for video instance segmentation.
The project supports numerous standard benchmarks and datasets across all four task categories, including MOT Challenge and DanceTrack for tracking, LaSOT and GOT10k for single object tracking, and YouTube-VIS for video instance segmentation. Comprehensive documentation, tutorials, and a model zoo with pre-trained weights make it accessible for researchers and practitioners. The project welcomes community contributions and is released under the Apache 2.0 license.