RepPoints is a novel representation for object detection that uses a set of sample points to represent objects, enabling both localization and recognition. Unlike traditional bounding boxes, which provide only a coarse localization, RepPoints adaptively position themselves to capture the spatial extent of objects and semantically significant local areas. This representation is learned using weak localization supervision and implicit recognition feedback, allowing for end-to-end training without the need for anchors. The proposed RepPoints-based detector, RPDet, achieves state-of-the-art performance on the COCO benchmark, with 46.5 AP and 67.4 AP50 using a ResNet-101 model. RPDet is anchor-free, replacing traditional bounding box representations with RepPoints for object proposals and final localization targets. This approach improves performance compared to bounding box-based methods and is more efficient than one-stage detectors like RetinaNet. RepPoints are flexible and can be used with deformable convolution, enhancing feature extraction and localization. The method is compared to existing approaches, including deformable RoI pooling, and is shown to provide a more accurate geometric representation of objects. Experiments demonstrate that RepPoints outperform bounding boxes in localization and recognition, and the method is effective in various object detection tasks. The proposed approach offers a promising direction for improving object detection by learning richer and more natural representations of objects.RepPoints is a novel representation for object detection that uses a set of sample points to represent objects, enabling both localization and recognition. Unlike traditional bounding boxes, which provide only a coarse localization, RepPoints adaptively position themselves to capture the spatial extent of objects and semantically significant local areas. This representation is learned using weak localization supervision and implicit recognition feedback, allowing for end-to-end training without the need for anchors. The proposed RepPoints-based detector, RPDet, achieves state-of-the-art performance on the COCO benchmark, with 46.5 AP and 67.4 AP50 using a ResNet-101 model. RPDet is anchor-free, replacing traditional bounding box representations with RepPoints for object proposals and final localization targets. This approach improves performance compared to bounding box-based methods and is more efficient than one-stage detectors like RetinaNet. RepPoints are flexible and can be used with deformable convolution, enhancing feature extraction and localization. The method is compared to existing approaches, including deformable RoI pooling, and is shown to provide a more accurate geometric representation of objects. Experiments demonstrate that RepPoints outperform bounding boxes in localization and recognition, and the method is effective in various object detection tasks. The proposed approach offers a promising direction for improving object detection by learning richer and more natural representations of objects.