Video Action Localization App is a Streamlit-powered web interface that processes video input to detect and localize actions across frames using spatio-temporal deep learning models. It leverages MMAction2’s SlowFast-ACRN model for action recognition and Faster R-CNN for spatial localization.
A complete, Colab-ready solution for spatio-temporal action detection in videos, featuring:
✅ SlowFast-ACRN for action recognition ✅ Faster R-CNN for human/object detection ✅ Streamlit interface for interactivity ✅ Google Colab backend for GPU-powered inference
✅ Prerequisites
- A Google Colab account
- GPU runtime enabled in Colab (Runtime > Change runtime type > GPU)
- A free Ngrok account
- Your personal Ngrok Auth Token
- 📤 Upload .mp4, .avi, or .mov video files
- 🧠 Perform frame-wise action localization with SlowFast-ACRN
- 📥 Download annotated output video with bounding boxes and action labels
Component Description Backbone SlowFast Network with ACRN Head Detector Faster R-CNN (ResNet-50-FPN) Trained On Kinetics-400 Dataset
This app is designed to run entirely in Google Colab, using GPU resources for efficient inference, while Streamlit provides the user interface.
- Run the notebook sequentially from top to bottom
- You may be prompted to restart runtime after dependency installations
- Once Ngrok tunnel is active, you'll get a public URL to access your Streamlit app
├── Action_Localization.ipynb # 🚀 Main Colab notebook for setup & execution ├── app.py # 💻 Streamlit app logic ├── checkpoints/ # 🧠 Model weight files │ ├── slowfast-acrn.pth # - Action recognition weights │ └── faster_rcnn.pth # - Object detection weights ├── mmaction2/ # 📦 MMAction2 framework (cloned repo) │ ├── configs/ # - Model configuration files │ ├── demo/ # - Sample media for testing └── requirements.txt # 📌 Python dependencies
👤 Author Ibrahim Mustapha Mechatronics Engineering Student | Computer Vision & AI Enthusiast 🔗 GitHub 🔗 linkedin 📧 Email