Enhancing Feature Fusion Using Attention for Small Object Detection

Jie Li, Yanxiang Gong, Zheng Ma, Mei Xie

Abstract: At present, object detection performance can meet some routine tasks' requirements. However, the detection performance for small-sized objects is far from satisfactory. Therefore, we propose the feature layer attention module and nonlinear positioning loss penalty based on size to improve small object detection performance. Our work proposes the feature layer attention module, which introduces an attention mechanism in the feature layer to enhance the model's attention to small objects. Through the feature fusion scheme proposed in this paper, we solve the problem of insufficient features of small objects to a certain extent and reduce the difficulty of model training. Besides, we introduce a size-based nonlinear penalty in the loss function, which can enhance the penalty for small object positioning errors. The effectiveness of our method has been demonstrated on small object data sets. On VisDrone2019 dataset, the proposed method improves the detection's AP by 2.2%. On TT100k dataset, the proposed method improves the detection's AP by 1.0%.