LRP-DS: Lightweight RepPoints with Decoupled Sampling Point Set

Most object detection methods use rectangular bounding boxes to represent the object, while representative points network (RepPoints) employs a point set describe object. The RepPoints can provide more fine-grained localization and facilitates classification. However, it ignores difference between classification tasks. Therefore, lightweight with decoupling of sampling (LRP-DS) is proposed in this paper. Firstly, MobileNet-V2 Feature Pyramid Networks (FPN) employed as backbone realize network, rather than Resnet. Secondly, considering tasks, are decoupled, by introducing free method. Finally, due introduction method, problem mismatch accuracy confidence highlighted, so score independently. final structure paper achieves 73.3% mean average precision (mAP) on VOC07 test dataset, which 1.9% higher original same MobileNetV2 FPN. Our LRP-DS has speed 20FPS for input image (1000, 600), RTX2060 GPU, nearly twice fast ResNet50 Experimental results show effectiveness our