Semantically grounded explainability for pedestrian detection using anatomical proportions and grid-based visualization

Explainable computer vision aims to understand how and why complex vision systems make decisions. However, existing explainability methods often highlight relevant image regions without providing semantically meaningful explanations for human users. In this work, we analyze the behavior of a pedestrian detector under semantically meaningful image perturbations. To overcome the interpretability limitations of existing explainability approaches, pedestrian occlusions are defined according to anatomical proportions derived from Leonardo da Vinci’s golden ratio, enabling a human-understandable decomposition of body regions. In addition, a grid-based visualization of the input and latent representations is employed to provide a structured global view of the detector behavior. Experimental results show that although the detector attention is strongly concentrated on the leg region, the largest performance degradation occurs when the upper body is occluded, with the Average Precision decreasing from 86.51% for fullbody pedestrians to 35.25% under upper-body occlusion. These results suggest that pedestrian detectors rely on complementary semantic body cues beyond the most visually salient regions. The proposed framework provides complementary local and global explainability insights into pedestrian detection models. The code, visualizations, and experimental results are made publicly available at: https://github.com/eurecom-fscv/ExFMA Semantically Grounded Explainability for Pedestrian Detection