Multi-branch Bounding Box Regression for Object Detection

Localization and classification are two important components in the task of visual object detection. In recent years, object detectors have increasingly focused on creating various localization branches. Bounding box regression is vital for two-stage detectors. Therefore, we propose a multi-branch bounding box regression method called Multi-Branch R-CNN for robust object localization. Multi-Branch R-CNN is composed of the fully connected head and the fully convolutional head. The fully convolutional head focuses on the utilization of spatial semantics. It is complementary to the fully connected head that prefers local features. The features extracted from the two localization branches are fused, then flow to the next stage for classification and regression. The two branches cooperate to predict more precise localization, which significantly improves the performance of the detector. Extensive experiments were conducted on public PASCAL VOC and MS COCO benchmarks. On the COCO dataset, our Multi-Branch R-CNN with ResNet-101 backbone achieved state-of-the-art single model results by obtaining an mAP of 43.2. Extensive comparative experiments prove the effectiveness of the proposed method.