Uncertainty for Identifying Open-Set Errors in Visual Object Detection

TL;DR

GMM-Det is a real-time uncertainty estimation method that effectively identifies and rejects open-set errors in object detection, enhancing safety in open-world applications.

Contribution

The paper introduces GMM-Det, a novel approach using class-specific Gaussian Mixture Models to detect open-set errors in object detectors with minimal overhead.

Findings

GMM-Det outperforms existing uncertainty methods in open-set error detection.

GMM-Det maintains detection performance while rejecting false positives.

The methodology enables evaluation of open-set detection in existing datasets.

Abstract

Deployed into an open world, object detectors are prone to open-set errors, false positive detections of object classes not present in the training dataset. We propose GMM-Det, a real-time method for extracting epistemic uncertainty from object detectors to identify and reject open-set errors. GMM-Det trains the detector to produce a structured logit space that is modelled with class-specific Gaussian Mixture Models. At test time, open-set errors are identified by their low log-probability under all Gaussian Mixture Models. We test two common detector architectures, Faster R-CNN and RetinaNet, across three varied datasets spanning robotics and computer vision. Our results show that GMM-Det consistently outperforms existing uncertainty techniques for identifying and rejecting open-set detections, especially at the low-error-rate operating point required for safety-critical applications. GMM-Det maintains object detection performance, and introduces only minimal computational overhead. We also introduce a methodology for converting existing object detection datasets into specific open-set datasets to evaluate open-set performance in object detection.