New RF-DETR Segmentation Checkpoints from Nano to 2XLarge

RF-DETR-Seg is a real-time transformer architecture for instance segmentation, developed as an extension of the RF-DETR detection model, with a segmentation head added to maintain real-time inference capabilities. The release includes pre-trained checkpoints across various model sizes, from Nano to 2XLarge, available under the Apache 2.0 license, and introduces new object detection checkpoints for larger models based on DINOv3 under the Platform Model License 1.0. RF-DETR-Seg demonstrates superior accuracy over other real-time models like YOLO26, with varying model sizes catering to different latency and resolution needs. The architecture supports fine-tuning on custom datasets, with training facilitated through the RF-DETR Python package or directly on the Roboflow Platform, enhancing adaptability to diverse hardware and dataset configurations. The model continues to achieve state-of-the-art results in real-time instance segmentation, maintaining practical latency while expanding support for additional computer vision tasks.