基于余弦相似度的分类定位一致性损失

Consistency Loss Between Classification and Localization Based on Cosine Similarity

主流的目标检测模型将检测分为分类和定位两个子任务,分类和定位各自具有独立的子网络,且在训练过程中采用互相独立的损失函数。这种模型结构和训练方式忽略了分类和定位之间的相互联系,使得模型预测的类别得分无法体现预测框的定位质量,进一步导致高定位质量的预测在非极大值抑制(NMS)阶段被低定位质量的预测抑制,损害了模型的检测精度。针对该问题,提出了一种一致性损失的概念,该损失通过在训练过程中约束模型预测的类别得分和定位质量的排名相似度,提升了二者的一致程度。基于FCOS-ResNet50模型与PASCAL VOC数据集,所提的损失函数能够提升约1.3个百分点的mAP_(0.5)、4.3个百分点的mAP_(75)和5.4个百分点的mAP_(90)。

Object detection is decoupled into two subtasks,namely,classification and localization,in mainstream detectors.Each task possesses a separate detection subnetwork and is trained with an independent loss function.In this way,the correlation between classification and localization is disregarded,and thus the classification score predicted by the model is not capable of representing the localization quality of the prediction box.Consequently,predictions of high localization quality may be suppressed by their poorly localized counterparts in the procedure of Non-Maximum Suppression(NMS),inducing precision degradation.To tackle this problem,a consistency loss is proposed to constrain the rank similarity between the classification score predicted by the model and the localization quality in training process to reinforce about their consistency.Based on FCOS-ResNet50 model and PASCAL VOC dataset,the proposed loss function brings about 1.3 percentage points of mAP_(0.5),4.3 percentage points of mAP_(75),and 5.4 percentage points of mAP_(90) improvements.