ODSmoothGrad: Generating Saliency Maps for Object Detectors
This provides a tool for improving explainability in object detection, but it is incremental as it adapts existing methods to a new task.
The paper tackles the problem of generating saliency maps for object detectors, extending explainability techniques from image classification to include both classification scores and bounding box parameters, and demonstrates this on one-stage and two-stage detectors with comparisons to classifier-based methods.
Techniques for generating saliency maps continue to be used for explainability of deep learning models, with efforts primarily applied to the image classification task. Such techniques, however, can also be applied to object detectors, not only with the classification scores, but also for the bounding box parameters, which are regressed values for which the relevant pixels contributing to these parameters can be identified. In this paper, we present ODSmoothGrad, a tool for generating saliency maps for the classification and the bounding box parameters in object detectors. Given the noisiness of saliency maps, we also apply the SmoothGrad algorithm to visually enhance the pixels of interest. We demonstrate these capabilities on one-stage and two-stage object detectors, with comparisons using classifier-based techniques.