CPH DETR: Comprehensive Regression Loss for End-to-End Object Detection

In the DEtection TRansformer (DETR) model, one-to-one matching is employed for categorizing positive and negative samples. However, this approach presents an issue: among multiple predictions highly overlapping with the target, only one is labeled as a positive sample, while the remainder are categorized as negative samples. Consequently, these negative samples cause misalignment of low category confidence and accurate position regression, and interfere with the classification branch judging positive samples. To address this issue, we propose a novel regression loss termed Comprehensive Regression Loss (CPH Regression Loss). It optimizes not only the positive samples but also the suboptimal samples, which are negative samples highly overlapping with the positive ones. Compared with traditional regression loss based on L1 loss and Generalized Intersection over Union loss, our loss incorporates the proximity between suboptimal samples and positive samples as new loss term, so as to devise the one-dimensional spatial constraint ComPreHensive L1 Loss (CPH L1 Loss) and the two-dimensional spatial constraint ComPreHensive Generalized Intersection over Union Loss (CPH GIoU Loss). The approach aligns regression accuracy and confidence scores by moving suboptimal samples away from positive samples, thus facilitating optimization. We refer to the model trained with CPH Regression Loss as ComPreHensive DETR (CPH DETR). Compared to the base model, our optimal model achieves a notable improvement of +3.7% AP.