Debugging Human Pose Estimation with Explainable AI

This paper examines the challenges of false positives and unstable detections in real-time object detection algorithms for human pose estimation. It analyzes potential causes like cluttered backgrounds, partial occlusions, low resolution, fast motion, poor image quality, and occlusions across frames. Example images demonstrate these issues. To explain the model's predictions, Grad-CAM visualizes important regions and LIME shows how perturbations influence outputs. These techniques identify when background areas incorrectly affect detections. The code implements the analysis and explanation process with detailed comments for understanding. The techniques help debug errors and increase model transparency. The paper investigates false positives, where non-humans are wrongly classified as people, and blinking, with inconsistent detections in sequential frames, in real-time object detection algorithms. It proposes a systematic analysis using explainable AI techniques like Grad-CAM and LIME. Grad-CAM visualizes key regions impacting predictions and LIME reveals perturbations altering outputs, supported by sample images. The goal is increasing model interpretability and enabling effective error correction. Overall, this research aims to demystify pose estimation models’ behavior to assist data scientists in enhancing reliability.