Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unbiased Regression Loss for DETRs

Edric, Ueta Daisuke, Kurokawa Yukimasa, Karlekar Jayashree, Sugiri Pranata

TL;DR

A novel unbiased regression loss is introduced for DETR-based detectors that normalizes the size of all boxes based on their individual width and height, and demonstrates consistent improvements in both fully-supervised and semi-supervised settings using the MS-COCO benchmark dataset.

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.

Unbiased Regression Loss for DETRs

TL;DR

A novel unbiased regression loss is introduced for DETR-based detectors that normalizes the size of all boxes based on their individual width and height, and demonstrates consistent improvements in both fully-supervised and semi-supervised settings using the MS-COCO benchmark dataset.

Abstract

In this paper, we introduce a novel unbiased regression loss for DETR-based detectors. The conventional 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 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.

Paper Structure

This paper contains 19 sections, 3 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Scale Invariant Regression Loss
  4. Transformer-based Object Detection
  5. Semi-Supervised Object Detection
  6. Proposed Method
  7. Preliminary
  8. Overview
  9. Overall Losses
  10. Sized $L_{1}$ Loss
  11. Experiments
  12. Datasets and Evaluation Metrics
  13. Implementation Details
  14. Comparison
  15. Semi-Supervised Object Detection
  16. ...and 4 more sections

Figures (3)

  • Figure 1: Discrepancy in COCO $mAP$ for small, medium, and large boxes with Semi-DETR
  • Figure 2: Model overview of Semi-DETR, with the regression component of the supervised and unsupervised losses replaced with Sized $L_{1}$ loss
  • Figure 3: Sized $L_{1}$ loss effectively normalizes all boxes to equal width and height before loss calculation