Cascaded Multi-Channel Feature Fusion for Object Detection

In this paper, we propose and evaluate a novel object detection architecture called Cascaded Multi-Channel Feature Pyramid Network, or CM-FPN. The proposed network, which is based on Feature Pyramid Network by Lin et al., employs multi-stage cascaded top-down feature pyramid networks to extract more semantic multiresolution feature maps for object region proposal and object classification. Depths of the feature maps are adjusted so that the feature maps in the later stages of the cascades where they are more semantic have higher channel depths. Experimental evaluation of the proposed approach has shown that the proposed method produces higher object detection accuracy.