For years, "real-time object detection" was synonymous with the YOLO family — convolutional, fast, and tuned hard on COCO. DETR-style transformer detectors were more principled but too slow to compete on latency. RF-DETR's wager is that this tradeoff is no longer fundamental: bolt a DETR detection head onto a self-supervised DINOv2 backbone, and the transformer becomes fast enough for real-time inference while inheriting representations that generalize past COCO.
What Sets It Apart
- One architecture, three tasks: detection, instance segmentation, and keypoint estimation share the same backbone and recipe, so you scale up or down without switching frameworks.
- Six sizes from Nano (~30M params) to 2XL (~127M), letting you pick a point on the speed-accuracy curve instead of retraining a different model family.
- It is benchmarked not only on COCO but on RF100-VL — 100 real-world datasets spanning domains like aerial, medical, and industrial imagery — where it transfers noticeably better than YOLO-class models. That domain-transfer edge, more than a fraction of a COCO point, is the practical argument for it.
Great Fit If — Look Elsewhere If
Great fit if you deploy detection on your own non-COCO data and want strong out-of-the-box transfer plus a real-time latency budget, or if you need detection, segmentation, and keypoints from a single stack. Look elsewhere if you are locked into an existing YOLO toolchain and edge target where every millisecond and exported-format quirk matters, or if you need permissive licensing across every variant — the Plus extensions ship under a separate, non-Apache license.
