Hole Detection Model
A deep learning pipeline for automated detection and analysis of holes in time-lapse videos.
Python Deep Learning Image Analysis
Overview
Trained a U-Net for detecting holes in time-lapse videos, to help the lab understand and quantify cooperative digging behaviour in larvae. More details in paper when released. It’s surprisingly difficult when trying to nmail down both the exact definition of a hole in this context, and how to identify holes of interest.
Key Features
- Model — U-Net model with lots of options around loss functions (focal, Tversky etc.), augmentations, LR schedulers to try and eek out performance
- Prediction — Inference over large video datasets via SLURM array jobs
- Postprocessing & analysis — separated from prediction to allow iteration without re-running inference
- Overlay generation — overlays predicted masks on original video for QC and visualisation