Quantification and mitigation of border-level localization deviation for object detectors

Abstract

Environmental perception is a critical module of automated driving systems for detecting obstacles and providing a decision-making basis for planning and control modules. In recent years, many localization deviation correction methods have emerged. However, these methods rarely study the border-level deviations, and current evaluation metrics cannot quantify the border-level deviation. To solve this problem, we model the maximum border localization deviation as a Gaussian distribution and propose a series of quantitative metrics to represent the border localization deviation as the absolute sum of the mean and variance of the distribution. Based on it, we proposed a predictive distribution fusion module to embed the predictive information into detection head networks, making the heads rethink and learn to reduce deviation. Experimental results demonstrate that our method can be flexibly integrated with various state-of-the-art detectors, further improving detection accuracy by approximately 1.0 mAP and enhancing the overall localization quality score by more than 6%. At an inference speed of 26.7 FPS, it achieves a detection accuracy of 43.3 mAP in urban road environments. The code and trained models are available at https://github.com/unbelieboomboom/RefineHead.