CE-FPN: enhancing channel information for object detection

Feature pyramid network (FPN) has been an efficient framework to extract multi-scale features in object detection. However, current FPN-based methods mostly suffer from the intrinsic flaw of channel reduction, which brings about loss semantical information. And miscellaneous feature maps may cause serious aliasing effects. In this paper, we present a novel enhancement (CE-FPN) alleviate these problems. Specifically, inspired by sub-pixel convolution, propose skip fusion (SSF) perform both and upsampling. Instead original 1 × convolution linear upsampling, it mitigates information due reduction. Then context (SCE) for extracting stronger representations, is superior other utilization rich convolution. Furthermore, introduce attention guided module (CAG) optimize final integrated on each level. It alleviates effect only with few computational burdens. We evaluate our approaches Pascal VOC MS COCO benchmark. Extensive experiments show that CE-FPN achieves competitive performance more lightweight compared state-of-the-art detectors.