Unbiased Regression Loss for DETRs
This addresses a specific problem in object detection for researchers and practitioners, offering an incremental improvement to existing DETR methods.
The paper tackles the bias in DETR-based detectors where conventional L1 loss favors larger boxes, harming small object detection, by proposing a Sized L1 loss that normalizes box sizes, resulting in consistent improvements on the MS-COCO benchmark.
In this paper, we introduce a novel unbiased regression loss for DETR-based detectors. The conventional $L_{1}$ regression loss tends to bias towards larger boxes, as they disproportionately contribute more towards the overall loss compared to smaller boxes. Consequently, the detection performance for small objects suffers. To alleviate this bias, the proposed new unbiased loss, termed Sized $L_{1}$ loss, normalizes the size of all boxes based on their individual width and height. Our experiments demonstrate consistent improvements in both fully-supervised and semi-supervised settings using the MS-COCO benchmark dataset.