Scale-balanced loss for object detection

• A matching imbalance in current object detection pipelines is pointed out. It can lead to poor performance of detecting objects with different scales. An innovative loss function called scale-balanced proposed alleviate the imbalance. Experiments demonstrate effectiveness loss, especially small improved significantly. Object an important field computer vision. Nevertheless, a research area that has so far not received much attention study into anchor strategy and anchor-based detection, particular detection. clear larger sizes tend match more anchors than smaller ones. This may result objects. be alleviated by paying fewer anchors. We propose flexible for which compatible popular methods. The method, does add any extra computational cost original pipelines. By re-weighting strategy, method significantly improves accuracy multi-scale Comprehensive experiments indicate achieved excellent generalization when separately applied some attained up 15% improvements on recall rates medium both PASCAL VOC MS COCO dataset. also beneficial AP improvement 1.5%.