Dynamic Balancing-Based Label Assignment for Object Detection
Yue Liao, Zhaoyan Fang, Huan Li, Dong Jiang, Huadong Li, Niannian Chen · 2023
Label assignment is a key task in object detection. Recent studies have shown that the results of label assignment have an important impact on the detector performance. Existing label assignment algorithms usually use the joint scores of classification and regression to classify positive and negative samples, but the joint relationship is determined by the hyperparameters which are not robust: the hyperparameters, as a static metric, cannot be dynamically adjusted according to the state of the network, and thus the scores also have the problem that they cannot reflect the importance of the samples adequately. To address these problems, this paper proposed a dynamic balancing-based label assignment algorithm (DBLA), which maps the sample points to a two dimensional plane with the regression score as the X-axis and the classification score as the Y-axis, by calculating the statistical characteristics of classification and regression of the candidate samples to determine the dynamic straight lines. The algorithm measures the quality of the samples by the distance from the candidate samples to the straight line, and classifies the positive and negative samples. Based on this, a positive sample weighting strategy which increases the weight of high-quality samples and balances the regression and classification performance of the model is designed. Experimental results on the MS COCO dataset show that DBLA achieves more reasonable in label assignment with 1.4% AP improvement compared to that on the ATSS baseline model by the above two improvements.