Input: A regular color photo (PNG, JPEG, etc.)
<aside> ðŸ›
Find Objects in a Photo, Fast
A modern object detector from Roboflow. Point it at any photo, and it draws boxes around the things it finds - people, cars, dogs, chairs, food, and 76 other everyday categories - with a confidence score for each one. The "Medium" version is the all-around pick: meaningfully more accurate than tiny detectors, much faster than the largest ones.
</aside>
RF-DETR-Medium looks at a photo and tells you what's in it and where each thing is. For every object it spots, it returns three pieces of information: a label ("person", "car", "dog"), a box drawn around the object, and a confidence score from 0 to 1 saying how sure it is. It's been trained to recognize 80 everyday categories from the well-known COCO dataset - people, animals, vehicles, household items, food, sports equipment, and more. The model is part of Roboflow's RF-DETR family, a new generation of object detectors built to be both accurate and fast enough for real-time use. The "Medium" size is the balanced choice for most projects: more accurate than the small variants, faster than the large ones.
<aside>
Input: A regular color photo (PNG, JPEG, etc.)
</aside>
<aside>
<aside>
Label – What it is (e.g., "person", "car", "dog", "chair")
Box – Where it is in the image, given as [left, top, right, bottom] pixel coordinates
Confidence – How sure the model is, from 0.0 to 1.0
</aside>
| Model size | 33.7 million parameters — a mid-size model |
| Speed | Real-time on modern GPUs; the "Medium" pick on the speed-vs-accuracy spectrum |
| Where it sits in the family | RF-DETR comes in multiple sizes — Medium is the balanced default. Smaller variants are faster but less accurate; larger ones are more accurate but slower |
| How it compares to YOLO | At the same speed bracket, RF-DETR-Medium is more accurate than similarly fast YOLO detectors (e.g., +1.6 mAP over YOLO26m at slightly faster speed). The trade-off: RF-DETR uses more parameters, so it needs more memory than YOLO of the same "size letter" |
| What it recognizes | 80 everyday object categories (the standard COCO list) |
| How it detects | One pass over the image — no need for older multi-step detection tricks |
| Good for | Production object detection where accuracy matters and you want better results than YOLO at similar speed |
| Architecture | RF-DETR — DINOv2-with-registers ViT backbone + multi-scale projector + Deformable DETR decoder |
| Backbone Detail | DINOv2-with-registers style ViT with RF-DETR's windowed / full attention alternation for efficient self-attention |
| Multi-scale Fusion | RF-DETR multi-scale projector (C2f-style blocks in the LW-DETR lineage) aggregates multi-level backbone features before the decoder |
| Decoder | Deformable DETR-style decoder with multi-scale deformable cross-attention |
| Object Queries | DETR-style queries with bipartite matching and auxiliary decoder losses for training stability |
| Training Losses | Classification + bounding-box L1 + GIoU, with auxiliary supervision on intermediate decoder layers |
| Training Tricks | Group DETR (parallel decoder copies during training for faster convergence) |
| Family Discovery | Weight-sharing Neural Architecture Search (NAS) over practical accuracy/latency knobs — Medium is one subnet on the frontier |
| Parameters | 33.7M (F32 weights) |
| Training Data | COCO 2017 (80 categories, as reflected in config.id2label) |
| Supported Frameworks | Hugging Face Transformers (RfDetrForObjectDetection), Safetensors |
| Paper | "RF-DETR: Neural Architecture Search for Real-Time Detection Transformers" — arXiv:2511.09554 (Robinson et al., Nov 2025) |