On the Accuracy of Open Video Quality Metrics For Local Decision in Av1 Video Codec

Visual explanations are important to increase models' transparency. Grad-CAM is an effective method because of its high class discrimination, no requirement of architectural changes, and so on. However, in detection tasks, because Grad-CAM only focuses on the importance of features but does not have spatial sensitivity, it generates heatmaps in which not related regions to detected objects are also highlighted. in this study, we propose Spatial Sensitive Grad-CAM (SSGrad-CAM), which can generate appropriate heatmaps for object detectors. SSGrad-CAM modifies the heatmap generated from Grad-CAM with space maps computed by normalizing the magnitude of gradients. in this manner, SSGrad-CAM can incorporate spatial sensitivity and focus on the importance of both features and space. Through experiments, we confirm SSGrad-CAM can generate appropriate heatmaps for detection results, and also confirm it can generate when models detect objects by paying high attention to their peripheral regions, as well.