RF-DETR Small — tabletop-22 (ONNX)
An RF-DETR Small detector (COCO-pretrained, Apache-2.0) fine-tuned on 22 tabletop object classes (ETRI tabletop data), exported to ONNX for VisionServe.
visionserve pull rfdetr-small-etri # this detector alone
visionserve pull rfdetr-gdino-siglip-etri # + GroundingDINO + SigLIP rescoring (open vocabulary)
Files
| file | contents |
|---|---|
model.onnx |
the detector, opset 17, 114 MB |
labels.txt |
23 lines: the 22 class names in head order, then N/A |
I/O contract
| name | shape | notes | |
|---|---|---|---|
| input | input |
[1, 3, 512, 512] f32 |
RGB, squashed to 512×512 (no letterbox, no padding), /255, ImageNet mean/std |
| output | dets |
[1, 300, 4] f32 |
cxcywh, normalised to the 512×512 input |
| output | labels |
[1, 300, 23] f32 |
logits, sigmoid (focal loss, not softmax); NMS-free |
Squash, not letterbox. The fine-tune used rfdetr's square_resize_div_64=True, and rfdetr's own
predict() resizes with F.resize(img, [res, res]). Serving this checkpoint letterboxed is a
train/serve mismatch that measured −7.35 mAP on a sibling checkpoint of the same recipe.
Training
rfdetr.RFDETRSmall(...).train(...) with 100 % of the 31.9 M parameters trainable ("pf-full" of a
partial-unfreeze sweep), 185 training / 62 validation images. Measured through VisionServe on the
62 validation images: 85.33 mAP, 38 ms per request (RTX A6000) — measured letterboxed, i.e. before the squash fix above.
License
Apache-2.0, inherited from RF-DETR (Roboflow). No third-party weights with other licenses.