Celebrity Face Recognition with InsightFace (AntelopeV2)

This project performs celebrity face recognition in video using InsightFace's AntelopeV2 model. It combines face detection, embedding generation, and face tracking to annotate celebrities' faces in video footage.

Business Value

While a celebrity face recognition model may have limited direct business value outside the entertainment industry, the underlying technologies — face detection, tracking, and recognition — have wide-ranging applications. These include automated security systems, access control, attendance monitoring in schools and workplaces, and customer analytics in retail. This project demonstrates the foundational components of such systems, showcasing their potential for real-world deployment.

---

Features

• Scrapes celebrity face images using Bing
• Extracts embeddings using InsightFace's antelopev2
• Matches faces based on cosine similarity
• Uses Deep SORT to track individuals across frames
• Annotates and exports labeled video
• Modular codebase (scraper, mapper, matcher, face_recognizer, comparator, and bbox)
• Testable via Jupyter notebook

---

Project Structure

.
├── bbox.py                 # Utility for drawing bounding boxes
├── comparator.py           # Calculates cosine similarity
├── face_recognizer.py      # Main script to process and annotate video
├── mapper.py               # Maps embeddings from scraped images
├── matcher.py              # Matches query embedding to known identities
├── scrapper.py             # Downloads celebrity images using Bing
├── embeddings_mapper.json  # Generated after mapping known faces
├── video_input/            # Input videos
├── video_output/           # Output videos with annotations
├── celeb_images/           # Scraped images of celebrities
├── celeb.ipynb             # Jupyter notebook for experimentation

NOTE: I didn't upload any images or videos for a few reasons, but you can still use the scraper to get the images, and use whatever videos you like!

---

Requirements

Install dependencies with:

pip install -r requirements.txt

Main packages:

• insightface
• opencv-python
• numpy
• deep_sort_realtime
• icrawler

---

How It Works

1. Image Collection

   • Uses icrawler to scrape face images of celebrities.
   • See: scrapper.py

2. Embedding Generation

   • Loads each image and generates embeddings using antelopev2.
   • Stores them in embeddings_mapper.json.
   • See: mapper.py

3. Face Recognition in Video

   • Detects faces, tracks them with Deep SORT.
   • Matches each face with known embeddings using cosine similarity.
   • Annotates faces and exports the labeled video.
   • See: face_recognizer.py

---

Example Output

A video of Elon Musk is analyzed and annotated with bounding boxes and names of detected celebrities.

Example output saved to: video_output/elon_musk_output_labeled.mp4

---

Testing via Notebook

A celeb.ipynb file is included to experiment with the components individually and test recognition on images interactively.

---

Customization

• To add new celebrities, more names are present in celebrity_list.csv, you can check it, add or remove names, etc.
• Adjust similarity threshold in matcher.py (default: 0.3).
• Change video source in face_recognizer.py.

---

Notes

• The system only performs recognition on frames sampled every half second (fps / 2).
• Detected faces without confident matches are labeled as "UNKNOWN".
• All scraped images are used for embedding without manual filtering, so quality may affect performance.

---

License

This project is for educational/demo purposes. Ensure compliance with any third-party terms when using scraped celebrity data.

---

Contributions

Pull requests, feedback, and improvements are welcome!