The amorphous bottleneck: a recall-optimized ensemble for anomaly detection

Abstract

Object detectors trained on rigid, well-defined targets tend to break down when the thing they are looking for has soft edges that blur into the surrounding texture, a compound failure we call the Amorphous Bottleneck. We study it on a road defect benchmark and find that most labeled cracks sit below the psychophysical Weber threshold of contrast, which leaves annotators disagreeing with each other and makes raw precision a misleading signal of model quality. Our response is data centric and recall first. Starting from a single Mask R-CNN baseline, we train nineteen augmentation specialists covering spatial, pixel level, and occlusion variations, then audit each one by tracking, per instance, which true defects it catches and which it misses. Predictions are merged with Weighted Boxes Fusion, and the final lineup is chosen by greedy forward selection against an asymmetric safety weighted score. Inspecting the confident false positives by hand showed that most were real defects the original annotation had skipped over, so the apparent loss of precision is largely a labeling artifact rather than the model hallucinating. The recipe should carry over to other low contrast, non-rigid anomaly tasks where supervision is noisy. On the held-out split, the six-model ensemble selected by this procedure raises operational recall to 0.9089 at a 27:1 rescue to regression ratio and holds mAP@50:95 steady.