Introduction
Object detection (OD) models serve as critical components in autonomous systems, yet they remain notoriously brittle when encountering "abnormal" scenarios. In real-world deployments, models often face partially occluded objects, out-of-distribution (OOD) samples, or adversarial attacks designed to induce silent failures.
A significant limitation of current evaluation frameworks is their reliance on ground-truth labels (e.g., mAP), which are unavailable during real-time inference. Furthermore, standard uncertainty quantification methods typically only assess the confidence of objects that have already been detected. If a model fails to propose a bounding box entirely, these methods offer no signal. There is a critical need for a practical, unsupervised metric that can flag when a model's internal perception diverges from its final predictions.
Methodology
The researchers propose Distinctive Localization (DiL), a model-agnostic metric designed to quantify internal model inconsistency. The core intuition is that an abnormality exists if a model’s internal layers perceive an object (high saliency) but its final output layer fails to localize it.
The DiL Pipeline
As illustrated in Figure 1, the DiL computation involves three primary steps:
- Saliency Mapping: Using Explainable AI (XAI) techniques like Grad-CAM, the system generates a saliency map representing the "objectness" perceived by the model across the entire input scene.
- Localization Components: Two values are extracted from the saliency map:
- Complete Localization (CL): The integral of saliency values across the entire image.
- Background Localization (BL): The integral of saliency values in regions not covered by any predicted bounding boxes.
- The DiL Ratio: The final metric is defined as $DiL = BL / CL$.
A low DiL score indicates that the model's predictions align with its internal focus. Conversely, a high DiL score suggests the model "sees" significant features in the background that it failed to convert into formal detections, signaling high uncertainty.
Experimental Setup
The efficacy of DiL was evaluated across nine distinct use cases using a diverse set of benchmarks:
- Standard Datasets: COCO 2017, OccludedPASCAL 3D+, and ImageNet subsets.
- Synthetic Occlusion: The E-PO dataset, featuring realistic occlusions generated via DALL-E 2.
- Adversarial Robustness: Adv-COCO and Adv-Superstore, containing digital and physical adversarial patches.
- Out-of-Distribution: OOD-Superstore, featuring retail items absent from the training distribution.
Key Insights: Perception vs. Prediction
The fundamental strength of DiL lies in capturing the friction between a model's mid-stage "perception" and its post-processing "decision." In many failure cases, the internal activations of the backbone or Region Proposal Network (RPN) correctly identify object features, but these are suppressed during Non-Maximum Suppression (NMS) or filtered by high confidence thresholds.
As shown in Figure 2, a clean scene where all objects are detected yields a negligible DiL (0.014). However, in a scene with partial occlusion where the detector fails, the DiL score spikes (0.77), providing a clear signal of abnormality even without ground-truth labels.
Beyond detection, the authors introduced a Dynamic Decision Threshold (DDT). By using the DiL score to modulate the model's confidence requirements in real-time, the system can automatically "unhide" objects that were initially suppressed due to uncertainty.
Results and Discussion
DiL demonstrated superior performance in identifying abnormalities compared to traditional methods like Monte Carlo Dropout. In standard COCO benchmarks, clean scenes maintained a mean DiL of ~0.15, while OOD scenes rose to ~0.91.
Applying the DDT mechanism led to substantial improvements in model robustness. In specific OOD retail scenarios, the authors reported recall improvements of up to 260% without significant degradation in precision. This suggests that DiL can effectively recover "lost" detections in challenging environments.
Limitations and Future Work
Despite its success, DiL faces challenges with extremely small objects where saliency maps lack the necessary granularity to distinguish signal from noise. Additionally, the metric's reliability is inherently tied to the quality of the underlying XAI technique. Future research will focus on extending DiL to Vision Transformer (ViT) architectures and dense prediction tasks such as semantic segmentation, as well as refining the balance between precision and recall in clean scenes.
Summary updated