README.md

metadata

title: GenVidBench - Video Action Recognition
emoji: 🎬
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.47.2
app_file: app.py
pinned: false
license: apache-2.0
short_description: State-of-the-art video action recognition using MMAction2

GenVidBench - Video Action Recognition

A powerful video analysis tool that uses state-of-the-art deep learning models to recognize actions and activities in videos. Built on top of MMAction2 framework with a user-friendly Gradio interface.

🚀 Features

• Action Recognition: Identify actions and activities in videos using TSN (Temporal Segment Networks)
• Top-5 Predictions: Get the most likely actions with confidence scores
• Multiple Formats: Support for MP4, AVI, MOV, and other video formats
• Real-time Processing: Fast inference optimized for web deployment
• User-friendly Interface: Clean and intuitive Gradio web interface

🎯 Model Details

This demo uses:

• Model: TSN (Temporal Segment Networks) with ResNet-50 backbone
• Dataset: Trained on Kinetics-400 dataset (400 action classes)
• Framework: MMAction2 (OpenMMLab)
• Input: RGB video frames
• Output: Top-5 action predictions with confidence scores

🛠️ Technical Stack

• Backend: Python, PyTorch, MMAction2
• Frontend: Gradio
• Video Processing: OpenCV, Decord
• Deployment: Hugging Face Spaces

📖 How to Use

1. Upload Video: Click the upload area or drag and drop your video file
2. Wait for Processing: The model will analyze your video (usually takes a few seconds)
3. View Results: See the top 5 predicted actions with confidence scores

💡 Tips for Best Results

• Video Length: Shorter videos (under 30 seconds) process faster
• Video Quality: Clear, well-lit videos work best
• Action Clarity: Videos with clear, distinct actions yield better results
• Supported Formats: MP4, AVI, MOV, and other common video formats

🔬 Supported Actions

The model can recognize 400 different action classes from the Kinetics-400 dataset, including:

• Sports activities (basketball, soccer, tennis, etc.)
• Daily activities (cooking, cleaning, reading, etc.)
• Physical exercises (push-ups, jumping jacks, etc.)
• Musical activities (playing instruments, singing, etc.)
• And many more!

🏗️ Architecture

Video Input → Frame Sampling → Feature Extraction → Classification → Top-5 Predictions

📊 Performance

• Accuracy: State-of-the-art performance on Kinetics-400
• Speed: Optimized for real-time inference
• Memory: Efficient GPU/CPU utilization

🤝 Contributing

This project is part of the GenVidBench framework. Contributions are welcome!

📄 License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

🙏 Acknowledgments

• MMAction2 - The underlying framework
• OpenMMLab - For the excellent computer vision tools
• Hugging Face - For the deployment platform

---

Note: This is a demonstration of video action recognition capabilities. For production use, consider additional validation and error handling.