CONTEXT AUGMENTATION AND FEATURE REFINEMENT NETWORK FOR TINY OBJECT DETECTION
Tiny objects are hard to detect due to their low resolution and small size. The poor detection performance of tiny objects is mainly caused by the limitation of network and the imbalance of training dataset. A new feature pyramid network is proposed to combine context augmentation and feature refinement. The features from multi-scale dilated convolution are fused and injected into feature pyramid network from top to bottom to supplement context information. The channel and spatial feature refinement mechanism is introduced to suppress the conflicting formation in multi-scale feature fusion and prevent tiny objects from being submerged in the conflict information. In addition, a data enhancement method called copy-reduce-paste is proposed, which can increase the contribution of tiny objects to loss during training, ensuring a more balanced training. Experimental results show that the mean average precision of target targets on the VOC dataset of the proposed network reaches 16.9% (IOU=0.5:0.95), which is 3.9% higher than YOLOV4, 7.7% higher than CenterNet, and 5.3% higher than RefineDet.
PDF Abstract
