davidkodar/rf-detr-seg-medium-onnx

---

license: apache-2.0
library_name: onnx tags: - object-detection - instance-segmentation - rf-detr - onnx pipeline_tag: image-segmentation

RF-DETR-Seg Medium — ONNX export

Locally exported ONNX of the RFDETRSegMedium tier from Roboflow's
rfdetr Python package. Roboflow's training repo distributes PyTorch checkpoints; this is
the ONNX equivalent for downstream inference runtimes that don't embed PyTorch (ort, candle, tract, etc.).

Architecture is identical to the public RF-DETR-Seg Preview tier
mirrored at
bukuroo/RF-DETR-Seg-ONNX;
only the weights differ — Medium trades a small inference-time
penalty for noticeably better classifier accuracy on out-of-distribution scenes.

File

File	Size	SHA-256
rf_detr_seg_medium_v1.onnx	139 MB	ac764d38d19183940d120f3333eb769dbca90100b6955f2cef2b3848565f7cf4

Output structure

Three named outputs (matches Preview tier):

• dets — (1, 200, 4) cxcywh, normalised to [0, 1]
• labels — (1, 200, 91) COCO-91 class logits
• masks — (1, 200, 108, 108) per-detection mask, sigmoid-activated

Input: (1, 3, 432, 432) NCHW f32, ImageNet mean/std normalised, sRGB-decoded.

Conversion

from rfdetr import RFDETRSegMedium                              
RFDETRSegMedium().export(output_dir="./out")

(rfdetr 1.x; CUDA + PyTorch required during export.)

License

Apache 2.0 — same as upstream roboflow/rf-detr. Both code and checkpoints redistribute freely.