Spatial Sensitive Grad-CAM: Toward Instance-Specific Explanations for Object Detectors by Incorporating Spatial Sensitivity
Toshinori Yamauchi · IEEE Access · 2025
Visual explanations for object detectors are important for ensuring their reliability. Since object detectors detect each instance in an image, generating instance-specific explanations that identify the contributing regions for individual detections is crucial for appropriately interpreting their behavior. However, most existing methods are designed for image classification models, resulting in inappropriate explanations that highlight regions unrelated to the detected instance. In this study, we propose Spatial Sensitive Grad-CAM (SSGrad-CAM), a visual explanation method for object detectors that generates instance-specific explanations. The proposed method extends Grad-CAM, a widely used technique originally developed for classification models, to overcome its limitation in producing instance-specific heat maps. Grad-CAM estimates channel-wise feature importance, but it neglects spatial information relevant to individual instances, leading to inappropriate heat maps that highlight irrelevant regions. To address this, SSGrad-CAM incorporates spatial sensitivity into Grad-CAM using a space map, which is computed by normalizing the magnitude of the gradients. This space map encodes spatial importance unique to each detected instance, allowing the method to generate heat maps that correspond to individual detections. Through extensive experiments, we confirm that the proposed method outperforms existing approaches both quantitatively and qualitatively. These results demonstrate that SSGrad-CAM generates higher-quality heat maps that more accurately identify important regions for each detected instance, enabling a more precise and reliable understanding of object detectors.