vt-vl-lab/video-data-aug — explained in plain English
Analysis updated 2026-07-14 · repo last pushed 2021-10-26
Build a system that classifies sports highlights using limited training data.
Detect specific activities in surveillance footage without massive labeled datasets.
Analyze user-generated video content by training models with fewer manual labels.
Experiment with data augmentation techniques on standard academic video datasets like UCF-101 and HMDB-51.
| vt-vl-lab/video-data-aug | krishnaik06/eda_sweetviz | kaopanboonyuen/saie2026 | |
|---|---|---|---|
| Stars | 33 | 25 | 22 |
| Language | Jupyter Notebook | Jupyter Notebook | Jupyter Notebook |
| Last pushed | 2021-10-26 | 2020-06-06 | — |
| Maintenance | Dormant | Dormant | — |
| Setup difficulty | hard | easy | moderate |
| Complexity | 5/5 | 1/5 | 3/5 |
| Audience | researcher | data | researcher |
Figures from each repo's GitHub metadata at analysis time.
Requires significant hardware including 8 high-end GPUs and familiarity with command-line deep learning workflows.
This repository contains the code for a research project called "Learning Representational Invariances for Data-Efficient Action Recognition." In plain terms, it is a tool that helps computers learn to recognize human actions in videos, like swinging a baseball bat or doing a cartwheel, using far fewer labeled examples than normally required. The core benefit is making AI video understanding more practical when you don't have a massive dataset of manually categorized clips. The project works by teaching the AI to focus on what matters (the person and their movement) while ignoring things that shouldn't change the answer (like the background, lighting, or camera angle). It does this through a combination of supervised learning, where the AI learns from labeled examples, and semi-supervised learning, where it learns from both labeled and unlabeled video. The "data augmentation" aspect means it intentionally creates variations of the training videos, cropping, altering, or transforming them, so the AI becomes more robust. The code is built on top of an existing video analysis framework called MMAction2. This is primarily a tool for machine learning researchers and engineers working on video recognition. If you're building a system that needs to classify sports highlights, detect specific activities in surveillance footage, or analyze user-generated video content, this approach could help you get good results without needing to collect and label hundreds of thousands of training videos. The project includes ready-to-use configurations for standard academic video datasets like UCF-101 and HMDB-51. The code requires significant hardware, the researchers used 8 high-end GPUs for their experiments. It's built in PyTorch, a popular deep learning toolkit, and the documentation assumes familiarity with command-line training workflows. The README doesn't go into detail about the underlying research methodology, but it links to a project page for those who want to understand the theoretical contributions before diving into the code.
A research tool that helps computers learn to recognize human actions in videos using far fewer labeled examples. It creates variations of training videos so AI focuses on the action, not the background.
Mainly Jupyter Notebook. The stack also includes PyTorch, MMAction2, Jupyter Notebook.
Dormant — no commits in 2+ years (last push 2021-10-26).
The license is not mentioned in the repository documentation.
Setup difficulty is rated hard, with roughly 1day+ to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
Verify against the repo before relying on details.