[Paper Review] ODAM: Gradient-based instance-specific visual explanations for object detection

Chenyang Zhao, Antoni B. Chan|arXiv (Cornell University)|Apr 13, 2023

Explainable Artificial Intelligence (XAI)4 citations

TL;DR

ODAM is a gradient-based, instance-specific visual explanation method for object detectors that generates high-quality, noise-robust heat maps for individual object predictions across one- and two-stage detectors. It improves object discrimination through a novel training scheme (Odam-Train) and enables better non-maximum suppression (Odam-NMS), outperforming state-of-the-art methods in accuracy and robustness on crowded benchmarks.

ABSTRACT

We propose the gradient-weighted Object Detector Activation Maps (ODAM), a visualized explanation technique for interpreting the predictions of object detectors. Utilizing the gradients of detector targets flowing into the intermediate feature maps, ODAM produces heat maps that show the influence of regions on the detector's decision for each predicted attribute. Compared to previous works classification activation maps (CAM), ODAM generates instance-specific explanations rather than class-specific ones. We show that ODAM is applicable to both one-stage detectors and two-stage detectors with different types of detector backbones and heads, and produces higher-quality visual explanations than the state-of-the-art both effectively and efficiently. We next propose a training scheme, Odam-Train, to improve the explanation ability on object discrimination of the detector through encouraging consistency between explanations for detections on the same object, and distinct explanations for detections on different objects. Based on the heat maps produced by ODAM with Odam-Train, we propose Odam-NMS, which considers the information of the model's explanation for each prediction to distinguish the duplicate detected objects. We present a detailed analysis of the visualized explanations of detectors and carry out extensive experiments to validate the effectiveness of the proposed ODAM.

Motivation & Objective

• To address the lack of instance-specific visual explanations in object detection, where existing methods like Grad-CAM produce class-specific heat maps that highlight all objects of the same category.
• To overcome the computational inefficiency and noisy outputs of perturbation-based methods like D-RISE, which require numerous inference passes and depend on mask resolution.
• To improve model interpretability by enabling explanation of individual prediction attributes, such as class and bounding box coordinates, rather than just classification.
• To enhance object discrimination in explanations by training detectors to produce consistent explanations for the same object and distinct ones for different objects.
• To leverage instance-level explanations for improving non-maximum suppression (NMS), particularly in crowded scenes with overlapping detections.

Proposed method

• ODAM computes gradient-weighted activation maps by backpropagating gradients from each prediction attribute (e.g., class score, box coordinates) to intermediate feature maps, producing instance-specific heat maps.
• The method applies gradient-based attribution to feature maps, similar to Grad-CAM, but adapted for object detection to localize influence per predicted instance rather than per class.
• Odam-Train introduces a consistency loss to encourage similar explanations for the same object across multiple detections and a separation loss to ensure distinct explanations for different objects.
• The training scheme improves localization and discrimination in explanation maps without modifying the detector architecture.
• Odam-NMS uses the instance-level heat maps from ODAM to guide NMS, prioritizing detections with more localized and distinct explanations to suppress duplicates.
• The approach is general and compatible with one-stage detectors (e.g., FCOS), two-stage detectors (e.g., Faster R-CNN), and various backbones and heads.

Experimental results

Research questions

• RQ1 Can gradient-based attribution produce instance-specific visual explanations for object detectors, rather than class-specific ones?
• RQ2 How can model training be modified to improve the faithfulness and discrimination of visual explanations for individual object instances?
• RQ3 Can instance-level explanations from ODAM enhance the performance of non-maximum suppression in crowded scenes with overlapping objects?
• RQ4 How does ODAM compare to state-of-the-art methods like D-RISE and Grad-CAM in terms of explanation quality, efficiency, and robustness to object size?
• RQ5 What is the trade-off between explanation faithfulness and object discrimination, and how can it be managed?

Key findings

• ODAM produces higher-quality, less noisy heat maps than D-RISE and Grad-CAM, with better localization and robustness to object size, as shown in qualitative comparisons.
• Odam-Train significantly improves explanation localization and object discrimination, reducing confusion between similar objects and enhancing consistency for the same object.
• On the CrowdHuman dataset, Odam-NMS achieves the best Jaccard Index (81.1) and lowest Missing Rate (44.5) among all NMS variants for FCOS, outperforming NMS, Soft-NMS, and FeatureNMS.
• For Faster R-CNN, Odam-NMS achieves the highest Average Precision (88.1) and the lowest Missing Rate (42.8), demonstrating superior performance in crowded scenes.
• Despite slightly longer inference time due to heat map generation, Odam-NMS maintains competitive speed and achieves the best balance of accuracy and robustness in NMS.
• The results confirm that instance-level explanations from ODAM can effectively guide NMS, validating the interpretability and utility of the proposed method.