Follow Faces

Code for exhibition "Könnt ihr noch?" - Deichkind

koenntihrnoch.mov

_{Note: This code contains the CV/ML part. The code for the interface got implemented by the artist monomvisual - who build the head - inside movement.py.}

---

Description

• Face detection implemented for camera stream with YOLOv8 trained on faces by arnabdhar

• Face tracking

  • implemented with deepsort realtime [Kalman Filter (movement prediction) + Hungarian Algorithm (assignment) + embedding model (using clip_RN50x16 or torchreid)]

    • Target Coordinates Functionality: Three states:

      • ALONE: If no faces are detected, the target point performs a random walk across the screen with occasional pauses.

      • BASIC_BITCH_STARE: If faces are detected, the target point locks onto one face by its ID and follows the center of its bounding box for a random duration.

      • CLOSEUP_STARE: If a face occupies more than x% (default 25%) of the screen area, it is always immediately targeted and stared at, regardless of the current state.

      States and target IDs are re-evaluated when duration expires or when visibility changes.

• No Tracking

  Alternative approach to tracking (fallback if tracking performance and or compute resources are not be sufficient.)

  Two states:

  • ALONE : same as above
  • STARE: If boxes are detected, one is selected as the target, alternating the biggest or a random boy. For 20 seconds, teh chosen target is the box currently closest to the last target point.

• Smooth Hardware Movement:

• Hardware position moves continuously toward updated target coordinates in a separated thread.

Dependencies

Python

I used python 3.13 - you could peg it with uv venv --python 3.13 but properly not necessary

Python packages

You can install them via

pip install -r requirements.txt

Or even better if you use uv, run in the directory of this repo:

uv venv
uv pip install -r requirements.txt

Note:

If you want to use torchreid, you will have to in deepsort realtime source coe and change in deep_sort_realtime/embedder/embedder_pytorch.py: Line 163

tochreid.utils import FeatureExtractor

to

tochreid.reid.utils import FeatureExtractor

Run

• with uv uv run main.py

• standard python main.py

Args

See all arguments:

• with uv uv run main.py --help
• standard python main.py --help

You can

• choose between multiple camera devices
• choose whether the object detection is streaming as video or just in the logs - both ways you get a json payload out with the detections
• set a confidence threshold
• set an IoU threshold
• stare duration (max and min)
• close face threshold (% of the screen)
• choose between tracking with deepsort or alternative approach without tracking

---

This project is licensed under AGPL v3+