PROOF OF CONCEPT: (CatBot3000)
experimenting with the YOLOv5 object detection algorithm and the YouTube Bounding Boxes dataset..
PROJECT PLAN: to create an easy-to-use notebook that myself and others can use to learn about and train YOLOv5 object detection models using subsets of the YouTube Bounding Boxes dataset. So far, I have automated the data downloading and preprocessing portion with the process-data.py script. I plan on creating a clean and simple Google Colab notebook that will act as an easy interface for the user.
AUTOMATED SO FAR:
- class selection from the YouTube Bounding Boxes dataset
- downloading the videos (now about ~3x faster using multithreading)
- extracting frames (now much faster by only calling ffmpeg once per video..)
- generating labels in YOLOv5-ready format
- remapping selected classes for YOLOv5 (zero indexed)
- splitting the dataset and organizing files appropriately for YOLOv5 training