Validating Bounding Box Annotations

This post details the architecture and dataset generation for Uber's ML-based bounding box validation system. It introduces a four-stage pipeline: input processing, feature extraction (visual, motion, coordinate), ML classification using XGBoost, and post-processing/clustering. The dataset is generated by synthetically introducing jump and ID swap perturbations. Key features include temporal context analysis over an 11-frame window. The system aims to detect ID swaps and position jumps, with future support for freeze errors and scale distortions.