Spatial Sensitive Grad-CAM++: Improved Visual Explanation for Object Detectors via Weighted Combination of Gradient Map

Toshinori Yamauchi; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 8164-8168

Visual explanations for object detectors are important tasks and several methods have been proposed. One of the methods is Spatial Sensitive Grad-CAM which can generate instance-specific heat maps. However it may generate heat maps that highlight only part of the important regions because it computes the importance of the features from the average of the gradient map. To tackle this problem we propose Spatial Sensitive Grad-CAM++ which computes the importance of features from the weighted combination of the gradient map originally introduced in Grad-CAM++ and extended by us to the object detectors. With this improvement it can generate heat maps that more accurately capture the important regions for the detected results. Through experiments we confirm that it outperforms other methods which indicates that it generates higher quality heat maps for the object detectors.