Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (11): 209-214.DOI: 10.3778/j.issn.1002-8331.2107-0277

• Graphics and Image Processing • Previous Articles     Next Articles

Research on Occluded Object Detection by Improved RetinaNet

YANG Shan, WANG Jian, HU Li, LIU Bo, ZHAO Hao   

  1. 1.School of Information Engineering, Southwest University of Science and Technology, Mianyang, Sichuan 621010, China
    2.Shenzhen Launch Digital Technology Co., Ltd., Shenzhen, Guangdong 518000, China
    3.School of Information Science and Technology, University of Science and Technology of China, Hefei 230026, China
  • Online:2022-06-01 Published:2022-06-01



  1. 1.西南科技大学 信息工程学院,四川 绵阳 621010
    2.深圳市朗驰欣创科技股份有限公司,广东 深圳 518000
    3.中国科学技术大学 信息科学技术学院,合肥 230026

Abstract: Aiming at the problem of low detection accuracy caused by the dense and overlapping instances, a detection framework with the improved regression loss function and dynamic non-maximum suppression(NMS) is proposed in this paper. Rep-GIoU-Loss derived from the combination of GIoU-Loss and the rejection factor Rep is used for location regression of object, which increases the correlation among regression parameters and reduces the probability of candidate bounding box offset to adjacent truth boxes. As a result, Rep-GIoU-Loss not only improves the regression accuracy of object location effectively, but also enjoys good robustness to the occlusion problem. In addition, a density prediction branch for predicting occlusion degree is added. The computed value from the density prediction branch is utilized as the dynamic threshold of the NMS, which reduces the missed and false detection instances. The detection accuracy of the proposed approach is improved by 1.3 percentage points on Pascal VOC and 2.8 percentage points on the self-made dataset. The experimental results show the effectiveness of the proposed method.

Key words: object detection, occlusion, GIoU, RetinaNet, non-maximum suppression(NMS)

摘要: 针对目标检测任务中目标实例密集、重叠等因素导致的检测精度不高的问题,提出一种改进回归损失函数与动态非极大值抑制的目标检测框架。采用结合排斥因子Rep的GIoU-Loss进行目标位置回归,在增加回归参数间相关性的同时降低候选边框向邻近真值偏移概率。Rep-GIoU-Loss不仅有效提升目标位置回归精度,对目标遮挡情形也具有较好的鲁棒性。此外,增加稠密度预测分支预测目标被遮挡程度,并将遮挡程度预测值作为NMS方法的动态阈值,以减少漏检、虚检目标实例。实验结果表明,改进方法检测精度在PASCAL VOC2007测试数据集上提高了1.3个百分点,自制数据集可提高2.8个百分点,验证了该方法的有效性。

关键词: 目标检测, 遮挡, GIoU, RetinaNet, 非极大值抑制