BLNet: Boundary Points Localization Network for Object Detection
Jiaoyang An, Bo Ma · 2020
Currently, almost all of the two-stage object detectors treat bounding box localization as an offset regression problem in the second stage. However, the spatial information in each Region of Interest (RoI) feature map is not considered in this pipeline. In this paper, we propose a novel boundary points localization network (BLNet) to predict the location of four boundary points (topmost, bottommost, leftmost, rightmost) of objects on RoI feature maps with a fully convolutional network. In addition, in order to compensate for the low resolution of the heatmaps, we use a differentiable operation called soft-argmax to convert the heatmaps into the numerical coordinates directly. Experiments on PASCAL VOC 2007 and 2012 datasets demonstrate that our BLNet significantly outperforms the traditional regression-based methods. Using ResNet-101 as the backbone, our method achieves 80.9% mAP on VOC 2007 and 78.7% mAP on VOC 2012 dataset.