arnavdhole/kennedy-mouse-pipeline

This repository provides stage-one geometric preprocessing for mouse-pose datasets. The user workflow is to load Parquet data, apply dataset-specific keypoint choices and filtering, establish a scale and canonical reference, then scale, center, and rotate each frame into normalized poses for downstream models. Diagnostics include alignment plots and evaluation metrics saved as JSON. Dataset configs and notebooks support lab-specific variations. The README describes a later autoencoder stage, but its implementation is outside this repository; notebook internals were not sampled.