Bayesian Optimisation of Existing Object Detection Methods for New Contexts

Abstract

Pre-trained object detectors exhibit strong variations in performance when applied in different contexts, e.g., daytime vs. nighttime performance, up-close vs. long range detection. These variations limit the usefulness of pre-trained detectors for safety critical applications like autonomous driving, which require consistent decision making in every context. Retraining for all contexts is often impossible or prohibitively expensive due to the need for large amounts of labels in each context. Instead, we propose a probabilistic calibration layer which takes these context dependencies into account to translate the detection score produced by a pre-trained detector into a conditional probability of presence. As a proof of concept, we demonstrate that reinterpreting the confidence scores of three commonly used detectors based on the estimated distance to the supposed object yields an improvement in average precision of pedestrian detection of up to 3% on the NuScenes dataset.