Classification bias and regression bias adjustment for long-tailed traffic sign detection and recognition

Abstract

Traffic sign detection and recognition (TSDR), as a pivotal technology in Intelligent Transportation System, has attracted growing focus and widespread application in recent times. However, in practical applications, the complexity and variability of road conditions lead to a pronounced long-tailed distribution in traffic signs, i.e., a few classes account for a large proportion of instances while most classes contain only a few instances. This long-tailed distribution leads to significant bias during training. In this paper, we identify that such bias issues exist not only in the classification branch but also in the regression branch. Therefore, we first propose the Classification Bias Adjustment (CA) module to address classification bias. This module combines margin adjustment and gradient adjustment strategies based on the mean classification scores to alleviate classification bias. Meanwhile, we propose the Regression Bias Adjustment (RA) module to address regression bias. This module re-weights the regression loss for each class in accordance with the mean Intersection over Union (IoU) to alleviate regression bias. Through comprehensive experiments on the TT100K and GTSDB datasets, it has been validated that our proposed approach has greater effectiveness than the existing state-of-the-art methods.