Feature Maps Channel Augmentation for Object Detection

Taking advantage of feature maps channels is one of the key challenges in object detection. In this study, we propose an optimization solution to further utilize the inter-channel relationship of feature maps. According to the exploration experiments, we have proposed a new Feature Maps Channel Augmentation module (FMCA) that can be inserted into mainstream structures to better establish the strong and weak dependency relationship between channels. We used comparison and ablation experiments and found that the FMCA module achieved a certain improvement under the circumstances that the detection speed was less affected. On the COCO dataset, our FMCA with ResNet-50 backbone has achieved relatively better results.