martijnfolmer / CVSimilarityViaEmbedding

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Finding similarity between embedded representations of people

These classes can be used to track people or objects in a video without having to save the camera data, maintaining privacy. It does this by running object detection (using a YOLO network), then using the embedding layer of the famous CLIP model to change the image into an input tensor of size (1x512). These input tensors can then be compared to eachother in following frames. This allows us to ID a person or object as being the same person over time, without having to store their likeness.

Note that this is very much a proof of concept and many improvements can be made, use at your own risk.

Tracking with two people

A video of the results can be found here

About

License:Apache License 2.0


Languages

Language:Python 100.0%