Lightened Context Extraction Network for Object Detection

Multi-scale feature maps have long been used for object detection,especially for small objects with weak appearance.Feature Pyramid Network(FPN)accomplishes multi-scale feature fusion in an end-to-end framework,which attracts significant attention. In order to take full advantages of feature maps at each level,we follow part of AC-FPN(Attention-guided Context Feature Pyramid Network) to stack feature maps after every dilation convolution.The number of parameters keeps increasing during feature fusion,we propose to lighten our model by Channel Group Average without bringing in new parameters.Experiments on PASCAL VOC2007,we observe mAP is increased by 9.45% comparing to SSD+MobileNet+FPN and 2.04% comparing to Faster R-CNN.Furthermore,the mAP can be increased by another 1.08% when the DIoU loss being replaced with our new LSIoU(Location and Size-Based IoU)loss function.