计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (11): 151-159.DOI: 10.3778/j.issn.1002-8331.2202-0260

• 模式识别与人工智能 • 上一篇    下一篇

无锚框目标检测模型特征任务不对齐研究

郝帅征,刘宏哲   

  1. 1.北京联合大学 北京市信息服务工程重点实验室,北京 100101
    2.北京联合大学 机器人学院,北京 100101
  • 出版日期:2023-06-01 发布日期:2023-06-01

Research on Feature Misalignment Between Tasks in Anchor-Free Models

HAO Shuaizheng, LIU Hongzhe   

  1. 1.Beijing Key Laboratory of Information Service Engineering, Beijing Union University, Beijing 100101, China
    2.College of Robots, Beijing Union University, Beijing 100101, China
  • Online:2023-06-01 Published:2023-06-01

摘要: 通常的目标检测模型由分类任务和回归任务构成。由于不同的任务驱动因素,模型中头部对应的这两个任务分支网络对来自同一输入图片、同一个实例的特征具有不同的敏感性。这就造成了检测模型对于相同位置的特征、分类效果和回归效果相差巨大的问题,也就是任务特征不对齐的问题。但是通用的目标检测后处理办法,仅以分类分数作为非极大抑制过程的标准,带来了大量回归质量较差、但置信度很高的检测结果。对现代化的无锚框网络展开不对齐问题的研究分析,将问题进一步拆解为尺度层级上的不对齐和空间位置上的不对齐。提出了参数量代价最小的解决方案:使用可变形卷积模块对检测模型头部网络的感受野进行微调,使用考虑样本点对齐效果的标签分配机制进行对齐样本点的挖掘,创新性地解决了上述两个子问题。进一步的详细实验和对比分析证明了该工作的有效性和实用性,以及对不同特征提取骨干网络的鲁棒性。

关键词: 目标检测, 深度学习, 无锚框检测器, 标签分配机制

Abstract: General object detection models consist of classification and regression branches. Due to different task drivers, they have a different sensibility to the features from the exact instances. That causes a vast performance gap, the so-called task-feature misalignment problem. Based on the assumption that the candidate result with high classification confidence has a high regression quality, the standard prediction method employs only the classification score as the criterion in NMS procedures. That leads to many prediction results with high classification scores but poor regression qualities. This paper mainly researches the misalignment problem in modern anchor-free detection models, specifically decomposing the problem with scale and spatial misalignment. It proposes to resolve the problem at minimal cost-a minor modification of the head network, which tweaks the receptive field of two tasks individually, and a new label assignment method mining the most aligned feature samples. The experiments show that, compared to the baseline FCOS, a one-stage anchor-free object detection model, the model consistently gets around 3 AP improvements with different backbones, demonstrating the method’s simplicity and efficiency.

Key words: object detection, deep learning, anchor-free models, label assignment scheme