SOTA Instance Segmentation with RF-DETR

Released in March 2025, the RF-DETR model architecture has expanded to include RF-DETR Segmentation, establishing a new benchmark for real-time image segmentation by outperforming YOLO11 models in both speed and accuracy on the Microsoft COCO Segmentation benchmark. The RF-DETR Segmentation model incorporates a segmentation head inspired by MaskDINO and employs a non-hierarchical ViT backbone, allowing it to generate high-resolution features through bilinear upsampling for mask creation. This design enables the model to achieve over 30 FPS on a T4 GPU with an end-to-end latency of 5.6ms, corresponding to more than 170 FPS. The model's architecture leverages a shared feature space between the segmentation head and object decoder, enhancing learning efficiency and accuracy. RF-DETR Segmentation can be trained using the Roboflow platform, the open-source rfdetr Python package, or Autodistill, and offers deployment options through Roboflow Workflows. The model's performance, measured in terms of mAP and latency, positions it as a significant advancement in segmentation technology, with plans for further model family expansions and a forthcoming paper detailing the architecture.