|
|
--- |
|
|
title: Video Human Fall Detection With TimeSformer |
|
|
emoji: 🌍 |
|
|
colorFrom: blue |
|
|
colorTo: pink |
|
|
sdk: gradio |
|
|
sdk_version: 5.25.0 |
|
|
app_file: app.py |
|
|
pinned: false |
|
|
license: apache-2.0 |
|
|
short_description: Fall detector with TimeSformer |
|
|
--- |
|
|
|
|
|
# Video Human Detection Demo using TimeSformer |
|
|
|
|
|
This is a Hugging Face Spaces demo project that uses TimeSformer – a video transformer model – for video-based human detection (action recognition). In this demo, we use the pre-trained model `microsoft/timesformer-base-finetuned-k400` from Hugging Face, which has been fine‑tuned on the Kinetics‑400 dataset. The model is capable of classifying a video into one of 400 human action categories. |
|
|
|
|
|
## Overview |
|
|
|
|
|
- **Model:** We use a TimeSformer model (`microsoft/timesformer-base-finetuned-k400`) to classify video clips. |
|
|
- **Feature Extractor:** The demo employs the Hugging Face `AutoFeatureExtractor` for video to process and prepare video frames. |
|
|
- **Inference:** The model outputs a set of predicted action labels with scores. These predictions help detect human actions in the video. |
|
|
- **Interface:** Built with Gradio, the demo lets the user upload a video file. The application extracts frames from the video, processes them with the model, and displays the top action predictions. |
|
|
|
|
|
## Setup and Deployment |
|
|
|
|
|
1. **Requirements:** See `requirements.txt` for the list of required packages. |
|
|
2. **Run Locally:** You can run the demo locally using: |
|
|
```bash |
|
|
python app.py |
|
|
``` |
|
|
3. **Deploy on Hugging Face Spaces:** |
|
|
Simply push these files to a new repository under HF Spaces. The app is designed to run with ZeroGPU if available and it is fully compatible with CPU-only environments. |
|
|
|
|
|
## Notes |
|
|
|
|
|
- **Video Preprocessing:** The demo extracts frames using OpenCV and passes them to the feature extractor. The number of frames and the resolution are set to default values that can be adjusted. |
|
|
- **Model Performance:** TimeSformer is computationally heavy – for real-time use, consider using a smaller or distilled variant, or reduce the number of frames processed. |
|
|
- **ZeroGPU Support:** The app uses the `@spaces.GPU` decorator (from the HF Spaces ZeroGPU environment) if available; otherwise, it will run on CPU. |
|
|
|
|
|
Enjoy testing human detection in videos with this demo! |
|
|
|
