Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hybrid Classification-Regression Adaptive Loss for Dense Object Detection

Yanquan Huang, Liu Wei Zhen, Yun Hao, Mengyuan Zhang, Qingyao Wu, Zikun Deng, Xueming Liu, Hong Deng

TL;DR

The paper addresses cross-task inconsistency and hard-sample emphasis in dense object detection by introducing Hybrid Classification-Regression Adaptive Loss (HCRAL). HCRAL integrates a Residual of Classification and IoU (RCI) module for cross-task supervision, a Conditioning Factor (CF) to focus on difficult samples within each task, and Expanded Adaptive Training Sample Selection (EATSS) to enlarge informative positives beyond standard sampling. It demonstrates substantial performance gains on COCO test-dev across multiple one-stage detectors, achieving state-of-the-art results with FCOS+ATSS and notable improvements on RetinaNet and ATSS, including setups with backbones like Res2Net and deformable convolutions. The work shows that jointly modeling cross-task consistency and adaptive sample selection yields robust improvements and generalizes across architectures.

Abstract

For object detection detectors, enhancing model performance hinges on the ability to simultaneously consider inconsistencies across tasks and focus on difficult-to-train samples. Achieving this necessitates incorporating information from both the classification and regression tasks. However, prior work tends to either emphasize difficult-to-train samples within their respective tasks or simply compute classification scores with IoU, often leading to suboptimal model performance. In this paper, we propose a Hybrid Classification-Regression Adaptive Loss, termed as HCRAL. Specifically, we introduce the Residual of Classification and IoU (RCI) module for cross-task supervision, addressing task inconsistencies, and the Conditioning Factor (CF) to focus on difficult-to-train samples within each task. Furthermore, we introduce a new strategy named Expanded Adaptive Training Sample Selection (EATSS) to provide additional samples that exhibit classification and regression inconsistencies. To validate the effectiveness of the proposed method, we conduct extensive experiments on COCO test-dev. Experimental evaluations demonstrate the superiority of our approachs. Additionally, we designed experiments by separately combining the classification and regression loss with regular loss functions in popular one-stage models, demonstrating improved performance.

Hybrid Classification-Regression Adaptive Loss for Dense Object Detection

TL;DR

The paper addresses cross-task inconsistency and hard-sample emphasis in dense object detection by introducing Hybrid Classification-Regression Adaptive Loss (HCRAL). HCRAL integrates a Residual of Classification and IoU (RCI) module for cross-task supervision, a Conditioning Factor (CF) to focus on difficult samples within each task, and Expanded Adaptive Training Sample Selection (EATSS) to enlarge informative positives beyond standard sampling. It demonstrates substantial performance gains on COCO test-dev across multiple one-stage detectors, achieving state-of-the-art results with FCOS+ATSS and notable improvements on RetinaNet and ATSS, including setups with backbones like Res2Net and deformable convolutions. The work shows that jointly modeling cross-task consistency and adaptive sample selection yields robust improvements and generalizes across architectures.

Abstract

For object detection detectors, enhancing model performance hinges on the ability to simultaneously consider inconsistencies across tasks and focus on difficult-to-train samples. Achieving this necessitates incorporating information from both the classification and regression tasks. However, prior work tends to either emphasize difficult-to-train samples within their respective tasks or simply compute classification scores with IoU, often leading to suboptimal model performance. In this paper, we propose a Hybrid Classification-Regression Adaptive Loss, termed as HCRAL. Specifically, we introduce the Residual of Classification and IoU (RCI) module for cross-task supervision, addressing task inconsistencies, and the Conditioning Factor (CF) to focus on difficult-to-train samples within each task. Furthermore, we introduce a new strategy named Expanded Adaptive Training Sample Selection (EATSS) to provide additional samples that exhibit classification and regression inconsistencies. To validate the effectiveness of the proposed method, we conduct extensive experiments on COCO test-dev. Experimental evaluations demonstrate the superiority of our approachs. Additionally, we designed experiments by separately combining the classification and regression loss with regular loss functions in popular one-stage models, demonstrating improved performance.
Paper Structure (14 sections, 10 equations, 3 figures, 9 tables, 1 algorithm)

This paper contains 14 sections, 10 equations, 3 figures, 9 tables, 1 algorithm.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. METHOD
  4. LOSS FUNCTION DESIGN
  5. Classification Loss Function
  6. Regression Loss Function
  7. Expand Adaptive Training Sample Selection
  8. EXPERIMENTS
  9. Ablation Studies
  10. Evaluation of Individual Method Contributions
  11. Comparison with State-of-the-Art
  12. Generalization and Superiority of HCRA
  13. Conclusion
  14. Appendix

Figures (3)

  • Figure 1: The diagram provides an overview of our HCRA loss composition and the positive-negative sample selection strategy. Both the classification and regression loss functions incorporate adaptive modules, such as the RCI (residual of cls and IoU) module and the CF (conditioning factor) module. In the context of ambiguous anchors A, B, C, and D, the values associated with the RCI module and CF module are visually represented in the bar chart. Notably, anchors C and D represent samples newly introduced by EATSS (Expanded Anchor Target Sampling Strategy) in comparison to ATSS. Please see more details about EATSS in Algorithm \ref{['EATSS-code']}.
  • Figure 2: A diagram of structure of HCRAL. HCRAL includes CF and RCI. RCI, derived from the regression and classification predictions, aims to emphasize consistency. Meanwhile, CF is designed to focus on difficult-to-train samples. To futher exploring the performance of HCRAL, EATSS strategy is adopted.
  • Figure 3: Illustrations of various curves of different parameter. (a) The distribution of IoU and classification scores. (b) The curves of $\omega$ with different $\mu$ when $p^\ast=0$. (c) The curves of $t$ with different $\gamma$.