Dynamic Balance Net: Correlation-enhanced Two-stage Object Detection Network with IoU-Loss

The most widely used object detectors to date are based on the two-stage approach popularized by R-CNN, where two locators cooperate on estimating objects' position information. However, when a network reaches the best performance point, the first locator is under-fit and the second locator is over-fit because two locators are trained by same loss function while the first locator has harder work. This blocks two locators being fitted at the same time and limits two-stage networks' performance. In this paper, we propose a correlation-enhanced IoU-loss to tackle this problem. More precisely, our method adjusts loss functions according to IoU between output and ground-truth, which enhances first locator's loss and achieves better inner cooperation. Extensive experiments on the PASCAL-VOC and MS COCO datasets show the effectiveness of the correlation-enhanced IoU-loss, as well as its compatibility with and adaptivity to two-stage object detectors. All networks included in the experiment obtain a no less than 3 per cent enhancement.