PosNeg-Balanced Anchors with Aligned Features for Single-Shot Object Detection

We introduce a novel single-shot object detector to ease the imbalance of foreground-background class by suppressing the easy negatives while increasing the positives.To achieve this,we propose an anchor promotion module(APM)which predicts the probability of each anchor as positive and adjusts their initial locations and shapes to promote both the quality and quantity of positive anchors.In addition,we design an efficient feature alignment module to extract aligned features for fitting the promoted anchors with the help of both location and shape transformation information from the APM.The probabilities from APM are helpful for the detection classifier to identify the easy negatives and to ignore their gradients during training.We assemble the proposed modules to the backbone of VGG-16 and ResNet-101 network with an encoder-decoder architecture.Extensive experiments on MS COCO and PASCAL VOC well demonstrate our model performs competitively with alternative methods(42.8%mAP on MS COCO and 82.7%mAP on PASCAL VOC)and can run at 28.6 FPS.