Feature Fused Detector for Small Objects

Detecting small objects remains an unsolved challenge for object detection task due to their poor resolution and limited contexts. In this study, we put forward a novel method named Feature Fused Detector (FFD), which fuses multi-scale features from shallow to deep layers for object detection. FFD enhances the semantic information of small objects by combining two feature fusion modules named Multi-scale Feature Module (MFM) and TransHead. MFM uses the residual connection and a spatial pyramid pooling layer to connect features from different scales. TransHead is the detection head that contains convolution layers and Transformer Blocks, which forms three prediction branches by fusing top-down features. Comparison experiments are conducted on the Priority Research Application dataset, and the results demonstrate that our FFD outperforms YOLOv3 by 2.8% and achieves 0.8% improvement over YOLOv5l in AP metric, showing that our proposed method can be used in small objects scenarios.