Abstract
ODSmoothGrad is an explainability tool designed specifically for object detectors. It generates saliency maps that highlight the pixels influencing both classification scores and bounding box coordinates. By incorporating the SmoothGrad algorithm, it produces clear, localized visualizations that help researchers verify if a model is attending to the correct features to define an object's boundaries.
Introduction
While deep learning has revolutionized computer vision, these models remain largely "black boxes." Saliency maps—visualizations highlighting pixels that most influence a model's output—have become a staple for explaining image classifiers. However, object detection presents a more complex challenge: the model must not only identify what an object is (classification) but also where it is (localization via bounding boxes).
Most existing explainability research focuses on the "what." When applied to object detectors, previous methods often ignore bounding box parameters ($x_{min}, y_{min}, x_{max}, y_{max}$) or are restricted to specific architectures like YOLO or SSD. ODSmoothGrad addresses these gaps by providing a generalized, noise-reduced saliency method that explains the logic behind both the label and the box.
Methodology
ODSmoothGrad builds upon two foundational techniques: Vanilla Saliency, which takes the gradient of the output with respect to the input image, and SmoothGrad, which reduces visual noise by averaging saliency maps from multiple noisy versions of the input.
The key innovation is applying this process to the five distinct outputs of an object detector: the class logit and the four bounding box regressors. The researchers implemented two versions: a logit-based approach for RetinaNet anchor boxes and a more generalized approach targeting the output tensors of the Detectron2 library. To ensure consistency during the SmoothGrad process, an Intersection-over-Union (IoU) threshold of 0.7 was used to "align" detections across noisy samples before averaging their maps. Validation was performed using the MSCOCO dataset.
Results and Discussion
ODSmoothGrad demonstrates significant improvements in both spatial specificity and visual clarity compared to traditional methods.
Qualitative Analysis
As shown in Figure 1, ODSmoothGrad produces much sharper and more interpretable "hot spots" compared to standard gradient methods like Grad-CAM. While Grad-CAM often produces a single, coarse heatmap for an entire class, ODSmoothGrad generates distinct maps for every individual object detected in the frame.
A key finding is the spatial specificity of the maps: the saliency map for $x_{min}$ (the left edge) highlights the left side of the object, while the map for $x_{max}$ highlights the right. This confirms the model utilizes relevant edge features to regress box boundaries.
Quantitative Validation
The authors validated these maps using the Intersection-over-Foreground (IOF) metric. By dividing ground truth masks into thirds (e.g., left, middle, right), they found that saliency maps for specific parameters (like $x_{min}$) showed significantly higher IOF with the corresponding spatial region (the "left third"). For both RetinaNet and Faster R-CNN, IOF values for relevant spatial parameters were consistently high, typically peaking above 0.6.
Conclusion
ODSmoothGrad provides a robust framework for peering into the decision-making process of object detectors. By generalizing across architectures and focusing on localization parameters, it offers a more complete explanation than classification-only methods. Future work will explore extending this to transformer-based detectors (e.g., DETR) and developing more sophisticated metrics to quantify the "faithfulness" of these maps to the model's internal reasoning.
Summary updated