Unbiased Regression Loss for DETRs

TL;DR

This paper proposes a new unbiased regression loss for DETR detectors that normalizes box sizes to improve detection accuracy, especially for small objects, demonstrated through experiments on MS-COCO.

Contribution

The paper introduces the Sized L1 loss, a novel normalization technique that reduces bias towards larger boxes in DETR-based detectors.

Findings

Improved detection performance on MS-COCO benchmark.

Consistent gains in both fully-supervised and semi-supervised settings.

Reduction of size bias in object detection.

Abstract

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.