Poster V2 Webcam Emotion (FER)

Minimal, reproducible live webcam demo for POSTER_V2 facial expression recognition (FER).

Features

• Detects the largest face via OpenCV Haar cascade
• Classifies emotion using POSTER_V2 (AffectNet 8-class checkpoint)
• Draws a bounding box and color-coded label under the face
• Fully local: venv + pinned dependencies; no global installs

Quick Start (macOS/Apple Silicon tested)

• Prereqs: git, Python 3.10–3.13, camera permission for your terminal
• Setup:

  bash scripts/bootstrap.sh

• Run:

  ./.venv/bin/python src/webcam_emotion.py

• Quit: press q

Notes

• We use the AffectNet 8-class checkpoint because the 7-class model in the upstream repo hardcodes Windows paths for pretrains. The 8-class variant loads pretrains from a relative path.
• FER = Facial Expression Recognition. Classes assumed: neutral, happy, sad, surprise, fear, disgust, anger, contempt.
• Prefer another dataset (e.g., RAF-DB 7-class)? Open an issue/PR—we can switch the architecture and paths cleanly.

Project Layout

• src/webcam_emotion.py — minimal webcam script
• scripts/bootstrap.sh — reproducible setup (venv, deps, clone, weights)
• external/POSTER_V2 — upstream repo (ignored by git)
• external/POSTER_V2/models/pretrain — pretrain weights (ignored)
• external/POSTER_V2/checkpoint — dataset checkpoint (ignored)
• .venv — local virtual environment (ignored)

Reproducibility

• Pinned dependencies in requirements.txt (torch, torchvision, timm, thop, numpy, OpenCV, gdown, pillow)
• One-command bootstrap sets up venv, installs pins, clones upstream, downloads weights
• No global installs; everything stays within .venv

Troubleshooting

• Camera permission: System Settings → Privacy & Security → Camera → allow your terminal
• GPU/MPS: the script uses Apple MPS if available; otherwise CPU
• VS Code import hints (optional settings):

  {
    "python.analysis.extraPaths": ["external/POSTER_V2"],
    "python.defaultInterpreterPath": ".venv/bin/python"
  }

Acknowledgements

• Upstream model: POSTER V2 — https://github.com/Talented-Q/POSTER_V2 (MIT)

License

• MIT for this wrapper repo (see LICENSE)
• POSTER_V2 is MIT-licensed in its own repository