Efficient Scale Divide and Conquer Network for Object Detection.

Multi-scale methods are often used to improve the accuracy of detection models. However, this approach is usually computationally expensive. In this paper, we introduce an efficient Scale Divide and Conquer Network (ScaleDCNet) based on the keypoint object detection framework, which accomplishes independent detection at each scale with minimal cost. To achieve this goal, we propose a new one-stage detection network, which consists of a scale feature selection module and a scale-aware keypoint matching approach. The scale feature selection module obtains features of various sizes through an attention mechanism to divide and conquer the scale problem. The scale-aware keypoint matching method is used to make independent predictions for the positions of objects of different sizes. With the above approaches, our ScaleNet improves the detection ability of objects of different sizes, reduces the probability of missing and false detection, and performs better in scale deformation and occlusion. We evaluate on the MS-COCO dataset and our model achieves high accuracy with low computational resources compared to baseline and other algorithms.