Link to Youtube video: https://www.youtube.com/watch?v=4xLwqETWKOY
This repo contains all the code used to run the creepy face from that video. It is based on a sample script for person detection with depth estimation, in the Luxonis DepthAI-Python Repository.
The key features are:
- Mostly unaltered person detection from Luxonis repo, using Mobilenet for person detection, with stereo depth for determining the closest person and their distance.
- Added a separate thread to do all the management of servos, eye brightness, and responding to the pressure plate activity for the costume-ID sequence
- It tries to keep the (approx) eye location of the closest person within the center of the image by pulsing the servos to move the face/camera towards the detection.
- For costume ID it saves the most recent image from the camera, base64-encodes it, and sends it to a REST server on the local network.
- The costume ID server is running CLIP, which is an extremely powerful zero-shot classifier from OpenAI. It reads the list of costumes from costumes.txt, and computes embeddings for each one plugged into this statement: "A kid wearing a X costume".
- When a costume request is received, it encodes the image, and finds the closest text embedding. That is the returned costume.
- It also generates some cool output to show you the top 5 results.