PointRend is a novel method for efficient and high-quality image segmentation. By drawing parallels between classical computer graphics rendering techniques and challenges in pixel labeling tasks, the paper presents PointRend, a neural network module that performs point-based segmentation predictions at adaptively selected locations using an iterative subdivision algorithm. PointRend can be integrated into existing state-of-the-art models for both instance and semantic segmentation. It achieves excellent results with a simple design, producing crisp object boundaries where previous methods oversmooth and showing significant gains on COCO and Cityscapes datasets for both instance and semantic segmentation. PointRend's efficiency allows for higher resolution outputs that are otherwise impractical in terms of memory or computation compared to existing approaches. The paper introduces PointRend as a method for image segmentation by viewing it as a rendering problem. It uses a subdivision strategy to adaptively select a non-uniform set of points for label computation. PointRend can be applied to both instance and semantic segmentation tasks. It uses a point-wise feature representation and a small point head subnetwork to predict output labels from the point-wise features. The method is evaluated on instance and semantic segmentation tasks using COCO and Cityscapes benchmarks, showing qualitative and quantitative improvements. PointRend improves strong Mask R-CNN and DeepLabV3 models by a significant margin. PointRend is a general module that can be implemented in various ways. It accepts CNN feature maps and outputs high-resolution predictions. It makes predictions only on carefully selected points, extracting a point-wise feature representation by interpolating from the feature maps. The point head is a small neural network trained to predict labels from the point-wise features. The method is applied to instance segmentation by refining coarse masks through iterative subdivision, and to semantic segmentation by predicting labels on a regular grid. The paper discusses the method's application to instance and semantic segmentation, showing that PointRend outperforms existing methods in both tasks. It is efficient, allowing for high-resolution outputs with significantly fewer computational resources. The method is evaluated on multiple datasets and shows improvements in both quantitative and qualitative aspects. PointRend is also shown to be effective in combination with other models, such as DeeplabV3 and SemanticFPN, for semantic segmentation tasks. The method is efficient and effective, with a simple design that achieves high performance in both instance and semantic segmentation tasks.PointRend is a novel method for efficient and high-quality image segmentation. By drawing parallels between classical computer graphics rendering techniques and challenges in pixel labeling tasks, the paper presents PointRend, a neural network module that performs point-based segmentation predictions at adaptively selected locations using an iterative subdivision algorithm. PointRend can be integrated into existing state-of-the-art models for both instance and semantic segmentation. It achieves excellent results with a simple design, producing crisp object boundaries where previous methods oversmooth and showing significant gains on COCO and Cityscapes datasets for both instance and semantic segmentation. PointRend's efficiency allows for higher resolution outputs that are otherwise impractical in terms of memory or computation compared to existing approaches. The paper introduces PointRend as a method for image segmentation by viewing it as a rendering problem. It uses a subdivision strategy to adaptively select a non-uniform set of points for label computation. PointRend can be applied to both instance and semantic segmentation tasks. It uses a point-wise feature representation and a small point head subnetwork to predict output labels from the point-wise features. The method is evaluated on instance and semantic segmentation tasks using COCO and Cityscapes benchmarks, showing qualitative and quantitative improvements. PointRend improves strong Mask R-CNN and DeepLabV3 models by a significant margin. PointRend is a general module that can be implemented in various ways. It accepts CNN feature maps and outputs high-resolution predictions. It makes predictions only on carefully selected points, extracting a point-wise feature representation by interpolating from the feature maps. The point head is a small neural network trained to predict labels from the point-wise features. The method is applied to instance segmentation by refining coarse masks through iterative subdivision, and to semantic segmentation by predicting labels on a regular grid. The paper discusses the method's application to instance and semantic segmentation, showing that PointRend outperforms existing methods in both tasks. It is efficient, allowing for high-resolution outputs with significantly fewer computational resources. The method is evaluated on multiple datasets and shows improvements in both quantitative and qualitative aspects. PointRend is also shown to be effective in combination with other models, such as DeeplabV3 and SemanticFPN, for semantic segmentation tasks. The method is efficient and effective, with a simple design that achieves high performance in both instance and semantic segmentation tasks.