Open-source project
open-mmlab/mmaction2 avatar
open-mmlab/mmaction2

MMAction2: OpenMMLab's PyTorch Video Understanding Toolbox, Its Branch Switch and Its 2023 Release Line

OpenMMLab's Next Generation Video Understanding Toolbox and Benchmark

5,197 stars1,354 forksPythonApache-2.0

At a glance

What is it?
MMAction2 is a PyTorch toolbox covering action recognition, localization, spatio-temporal detection, skeleton-based action detection and video retrieval. Its default branch moved from master to main, and the newest GitHub release is v1.2.0 from October 2023.
Who is it for?
MMAction2 is a sound choice for researchers who already work in the OpenMMLab stack and want configurable baselines for action recognition, detection and retrieval on video. Treat it as a stable but aging codebase: the latest release is v1.2.0 from 2023-10-12, the last push was 2026-03-18, and the quick install pins an old conda Python, so check version compatibility before building on it.
Can I use it commercially?
Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Activity is slowing. The repository last received commits 8 months ago.
What is it written in?
Mainly Python, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 13, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • open-mmlab/mmaction25,197 ★ · +38
Star history
Repository2026-09-132026-11-13
open-mmlab/mmaction25.2k5.2k

38 Stars in the last 61 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

The default branch moved from master to main, so older tutorials point at 0.x code

The first thing the README says is about branches. The default branch has been switched to `main`, which was previously `1.x`, from `master`, which holds the current 0.x line. Users are encouraged to migrate to the latest version with more supported models, stronger pre-training checkpoints and simpler coding, and a Migration Guide on Read the Docs covers the details.

In practice this means two codebases share one project name. A blog post, Colab notebook or forum answer written against the old master branch may use configuration and API conventions that do not match the 1.x code you clone today. A model config copied from the old tree will not simply run on main.

The release history confirms the split: v1.0.0 was published on 2023-04-07, v1.1.0 on 2023-07-04 and v1.2.0 on 2023-10-12. When you search for help, check which major version the answer assumes, and read the Migration Guide before porting anything from 0.x.

Training depends on a stack of OpenMMLab packages that must be version-matched

MMAction2 is built on PyTorch and sits on top of other OpenMMLab libraries. Its README lists PyTorch, MMCV and MMEngine as dependencies, plus MMDetection and MMPose as optional ones. The repository's own requirements.txt only pulls in three sub-files: requirements/build.txt, requirements/optional.txt and requirements/tests.txt.

That split has a consequence. The core OpenMMLab packages are installed through the `mim` tool, not through the requirements file alone, so a plain `pip install -r requirements.txt` does not give you a working setup. The optional MMDetection and MMPose packages are only needed if you use features built on detection and pose, and the README does not spell out which demos need them.

Because MMCV and MMEngine carry compiled components and tie to specific PyTorch builds, the practical risk is a mismatch between PyTorch, CUDA and MMCV. The install note in the README warns about this itself, telling you to check that the PyTorch and cudatoolkit versions the conda command installs match your environment.

The quick install creates a conda environment on Python 3.8

The README collapses the installation into one block and points to the full instructions on Read the Docs for details:

shell
conda create --name openmmlab python=3.8 -y
conda activate openmmlab
conda install pytorch torchvision -c pytorch  # This command will automatically install the latest version PyTorch and cudatoolkit, please check whether they match your environment.
pip install -U openmim
mim install mmengine
mim install mmcv
mim install mmdet  # optional
mim install mmpose  # optional
git clone https://github.com/open-mmlab/mmaction2.git
cd mmaction2
pip install -v -e .

Two details deserve attention. The environment is created with Python 3.8, an older interpreter than many current machines default to, so expect to adjust it if your other tools need a newer version. And the final step is an editable install from a clone of the repository, `pip install -v -e .`, which means you work from source on the default branch, not from a tagged package. The PyPI badge points to a published mmaction2 package, but the quick instructions do not use it.

Five video tasks are covered, and everything else is out of scope

The Major Features list says the toolbox implements algorithms for five video understanding tasks: action recognition, action localization, spatio-temporal action detection, skeleton-based action detection and video retrieval. The README illustrates them with Kinetics-400 recognition, NTU-RGB+D-120 skeleton-based recognition, and spatio-temporal detection results on AVA-2.1 and Kinetics-400.

The design is modular. A video understanding framework is decomposed into components, and a customised framework is built by combining modules. That is what lets one codebase serve all five tasks through configuration files under the configs directory.

The boundary matters. If your work is video captioning, generation or anything outside the five named tasks, the README does not offer it as a target. The v1.2.0 notes show how the scope grows within those tasks: the VindLU multi-modality algorithm and training of ActionClip, lightweight MobileOne TSN and TSM models, the MSVD video retrieval dataset, SlowOnly K700 features for training localization models, and video and audio demos.

The demo folder holds one script per task, including webcam and ONNX variants

The repository ships a demo directory that is more than a single example. It contains demo.py, demo_inferencer.py, demo_skeleton.py, demo_audio.py, demo_spatiotemporal_det.py, demo_spatiotemporal_det_onnx.py, demo_video_structuralize.py and long_video_demo.py, plus webcam_demo.py and webcam_demo_spatiotemporal_det.py. Notebooks demo.ipynb and mmaction2_tutorial.ipynb sit alongside sample videos for skeleton and spatio-temporal detection.

The file names show the range: there are scripts for skeleton-based input, audio, long videos, a live webcam, and an ONNX route for spatio-temporal detection. The demo_configs and fuse folders hold supporting material.

The README does not give command-line arguments for these scripts, so the demo README inside the folder is where invocation is documented. That makes the demo folder a good starting point when you want to see a model run on your own clip before training anything. Run one on a short video first to confirm that your PyTorch and MMCV install is sound.

The newest GitHub release is from October 2023, and the last push was in March 2026

Maintenance needs a careful reading. The recent release list ends at v1.2.0 titled MMAction2 V1.2.0 Release, dated 2023-10-12, preceded by v1.1.0 and v1.0.0. The last push to the repository was on 2026-03-18, which is more than six months before this article, so the project cannot be called actively maintained on that evidence.

The repository is not archived, and it still has CI and a Codecov badge, issue trackers and documentation on Read the Docs. But no release has followed v1.2.0 in the list, so the code on main may include unreleased changes and the packaged version you install from PyPI may lag behind it. If you pin a version for a paper or a product, choose a tagged release and record it.

Issues opened today may also wait a long time for an answer. For long-running work, plan to read and patch the source yourself, and treat the Apache-2.0 license as the guarantee you do have: you may modify and redistribute the code under its terms.

The model zoo reaches back to 2014 papers, with checkpoints documented elsewhere

Results and models are listed in the model zoo on Read the Docs, and the README summarises them in a collapsible table grouped by task. Under Action Recognition, the visible entries include C3D from CVPR 2014, TSN from ECCV 2016, I3D from CVPR 2017, C2D and R(2+1)D from CVPR 2018, TRN from ECCV 2018, and TSM from ICCV 2019. Each entry links to a README in the matching configs folder, such as configs/recognition/tsn.

This arrangement tells you where to look for details. Accuracy numbers, checkpoint links and training settings live in those per-model READMEs and in the model zoo statistics page, not in the top-level README. The repository also keeps model-index.yml and dataset-index.yml as machine-readable indexes.

When you pick a model, start from its config README, confirm the dataset it was trained on, and check whether v1.2.0 or the current main branch is what produced the published results. The README's table is a catalogue of what is implemented, not a ranking of what performs best.

Editorial conclusion

MMAction2 is a sound choice for researchers who already work in the OpenMMLab stack and want configurable baselines for action recognition, detection and retrieval on video. Treat it as a stable but aging codebase: the latest release is v1.2.0 from 2023-10-12, the last push was 2026-03-18, and the quick install pins an old conda Python, so check version compatibility before building on it.

Frequently asked questions

What is MMAction2 used for?

MMAction2 is an open-source PyTorch toolbox for video understanding, part of OpenMMLab. It covers action recognition, action localization, spatio-temporal action detection, skeleton-based action detection and video retrieval.

How do I install MMAction2?

The README's quick instructions create a conda environment with Python 3.8, install PyTorch, then use mim to install mmengine and mmcv. You then clone the repository and run pip install -v -e . from the clone.

Which branch of MMAction2 should I use?

Use main, which is now the default branch and holds the 1.x line. The older master branch holds the 0.x line, and the README links a Migration Guide for moving between them.

Official sources

  1. License: Apache-2.0
  2. open-mmlab/mmaction2 on GitHub
  3. Project website
  4. README
  5. Releases
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/open-mmlab-mmaction2.svg)](https://hysenlabs.com/projects/open-mmlab-mmaction2)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/open-mmlab-mmaction2.svg)](https://hysenlabs.com/projects/open-mmlab-mmaction2)

Was this useful?