The paper introduces RepPoints, a novel representation for object detection that provides finer localization and semantic feature extraction compared to traditional bounding boxes. RepPoints are learned to adaptively position themselves over an object, bounding its spatial extent and indicating semantically significant local areas. Unlike bounding boxes, RepPoints do not require anchors and can be used in anchor-free object detectors. The authors demonstrate that an anchor-free detector using RepPoints achieves competitive performance with state-of-the-art anchor-based methods, achieving 46.5 AP and 67.4 AP$_{50}$ on the COCO test-dev benchmark using a ResNet-101 model. The paper also discusses the differences between RepPoints and deformable RoI pooling, showing that while both can improve feature extraction, RepPoints provide more accurate geometric localization.The paper introduces RepPoints, a novel representation for object detection that provides finer localization and semantic feature extraction compared to traditional bounding boxes. RepPoints are learned to adaptively position themselves over an object, bounding its spatial extent and indicating semantically significant local areas. Unlike bounding boxes, RepPoints do not require anchors and can be used in anchor-free object detectors. The authors demonstrate that an anchor-free detector using RepPoints achieves competitive performance with state-of-the-art anchor-based methods, achieving 46.5 AP and 67.4 AP$_{50}$ on the COCO test-dev benchmark using a ResNet-101 model. The paper also discusses the differences between RepPoints and deformable RoI pooling, showing that while both can improve feature extraction, RepPoints provide more accurate geometric localization.