Add LibreRFDETRm-ui (UI-DETR-1), match upstream RF-DETR resize, make Results.plot() render every task

[EHxuban11]

What: add the UI-DETR-1 weights as an RF-DETR dataset variant (LibreRFDETRm-ui.pt), and fix two RF-DETR bugs it exposed.
Why: #896. UI-DETR-1 (racineai, MIT) is an RF-DETR-M fine-tune for class-agnostic UI element detection, so it needs weights, not a new family.

• WEIGHT_VARIANTS = ("ui",) on RF-DETR, so LibreRFDETRm-ui.pt auto-downloads from LibreYOLO/LibreRFDETRm-ui.
• PUBLISHED_WEIGHT_VARIANTS limits variant downloads to the published (size, task, variant) triples; names like LibreRFDETRs-ui.pt or LibreRFDETRm-seg-ui.pt raise FileNotFoundError listing the valid name instead of hitting a 404.
• RF-DETR predict/val resize is now bilinear without antialiasing (cv2 float32), matching RF-DETR training and upstream predict() since rf-detr 1.9.0 (fix(inference): disable antialias in predict() resize to match training roboflow/rf-detr#1206). Before this, PIL bilinear antialiased on downscale and boxes drifted on large inputs.
• Single-output upstream RF-DETR heads (one-category training set) now convert as nc=1, not nc=80 with 79 placeholder names.
• Weights uploaded to https://huggingface.co/LibreYOLO/LibreRFDETRm-ui (tensors identical to upstream model state, strict schema validation passes) and added to the LibreYOLO Models collection. Auto-download from a clean directory verified.
• Results.plot() renders every predict output (boxes, masks, OBB, keypoints, classify top-5, points, OCR, semantic, panoptic, gaze). It raised NotImplementedError for all of them. Returns a BGR np.ndarray (pil=True for PIL) and takes conf, labels, boxes, masks, probs, line_width, img, show, save, filename. One renderer, drawing.draw_results, now backs plot(), predict(save=True), video saving, L2CS and every exported-model backend, so a plot matches the saved file.
• Results.orig_img (BGR) keeps in-memory and URL sources. Local files are reopened from path; collected video/GIF results keep no frames and plot() decodes frame_idx on demand, so directory and video predictions do not hold every image.
• Docs: nomenclature, weights/LICENSE_NOTICE.txt, upload-skill whitelist, CHANGELOG.

Parity vs upstream rfdetr 1.10.1 at conf 0.3 (5 UI screenshots, 6 COCO images; UI-DETR-1 and stock RF-DETR-M):

	UI boxes matched IoU>0.9	UI median IoU	COCO median IoU
before	26-60 per image	0.71-0.88	0.968-0.999
after, PyTorch	all	0.999-1.000	1.000
after, ONNX	all	0.999-1.000	1.000

ONNX vs PyTorch on UI-DETR-1: identical boxes at conf >= 0.3; differences only in boxes below conf 0.26.

Check (plot): draw_results is a move of _save_annotated_image's per-task branches; saved detect/segment/pose images are pixel-identical before and after the refactor (YOLO9, RF-DETR-seg, YOLO-NAS-pose). Classification saves now draw the top-5, and tracked results show track IDs.

Greptile notes: the "single-logit head mislabelled" finding is not a bug. RF-DETR uses sigmoid/focal classification with no background logit; a one-category upstream dataset gives a one-output head, and upstream rfdetr predicts class 0 for every UI-DETR-1 box, which this conversion matches box for box.

Known limits: in-memory image lists keep their sources (the caller already holds them); a collected remote video stream cannot be re-decoded, so pass stream=True to plot its frames.

Check: the resize change affects every RF-DETR detect/seg/obb predict, val and ONNX runtime (pose keeps its antialiased path; LW-DETR/DETR/Deformable DETR have their own preprocess and are untouched). RF-DETR-M COCO mAP50-95 on a 200-image val subset: 0.6195 -> 0.6179.

Not verified:

• Full COCO val (5k) mAP delta.
• The six Roboflow Universe training-dataset licenses were not opened individually; the authors declare the merged training data MIT.
• Local PR gate: 7677 passed; 3 ONNX parity failures (DEIM, RT-DETRv2, D-FINE-seg) also fail on clean dev on this machine and pass in dev CI.

Closes #896.

Opened by an agent.

Code provenance

• Original code written for this PR.
• The antialias-free resize follows the behavior of roboflow/rf-detr commit 09ca8f54 (v1.9.0, Apache-2.0); no code was copied.
• Weights (hosted on HF, not in this repo): racineai/UI-DETR-1, HF revision 0f0dda5, MIT; fine-tuned from RF-DETR-M (roboflow/rf-detr, Apache-2.0) with a DINOv2 backbone (Apache-2.0).

Confidence Score: 5/5

The PR appears safe to merge; no new actionable issue or outstanding previous finding remains.

Findings

1.  UI weights cannot auto-download. ▶

Summary

The PR adds UI-DETR-1 as an RF-DETR-M weight variant, changes RF-DETR inference resize to match upstream behavior, and gives prediction results a shared renderer for plotting and saving.

• The changes since the previous review retain decoded sources for gaze and OCR results from in-memory images, URLs, and streamed video.
• Collected video results release source frames and decode them on demand.

_{Reviews (4) · Last reviewed commit: "Keep the source image on gaze and OCR re..."}

[greptile-apps]

UI weights cannot auto-download. If the Hugging Face mirror remains unavailable as stated in the PR, loading LibreRFDETRm-ui.pt without a local copy sends the user to a URL that returns 404. The advertised auto-download workflow fails, so users cannot obtain the new variant that way. Publish the weights before advertising it as downloadable.

[EHxuban11]

Handoff (end of 2026-09-24)

State of this PR at head 32ec7879 (branch 896-ui-detr, based on current dev):

Done

• LibreRFDETRm-ui.pt is uploaded to https://huggingface.co/LibreYOLO/LibreRFDETRm-ui (MIT, 5-file contract, strict schema validation passes, tensors bit-identical to upstream model state) and added to the LibreYOLO Models collection. Auto-download from a clean directory verified.
• RF-DETR predict/val resize now matches upstream rfdetr >= 1.9.0 (no antialias). Parity vs rfdetr 1.10.1: all boxes IoU > 0.9 on UI screenshots and COCO images, PyTorch and ONNX.
• Results.plot() renders every task through drawing.draw_results, shared with predict(save=True), video saving, L2CS and all export backends.
• Local Greptile (greptile review --agent --branch dev, after git branch -f dev origin/dev) last scored 3/5 with its one finding fixed in 32ec7879. The remaining notes are listed in the description as a false positive (single-logit head) and known limits.
• Local PR gate: 7677 passed. test_detr_cpu_export_matrix ONNX parity for DEIM / RT-DETRv2 / D-FINE-seg fails on macOS on clean dev too (passes in CI).

Next steps

1. Wait for the unit-test jobs and the Greptile bot on 32ec7879 (both pending at handoff). The two inline Greptile comments on model.py:108 are from the first commit and already addressed: the weights are uploaded, and unpublished variant names raise FileNotFoundError.
2. Fix any new valid bot finding by pushing a commit. Then it's a maintainer merge.

Open decisions (not in this PR)

• predict() treats numpy arrays as RGB when color_format="auto". The ecosystem convention is BGR (cv2.imread). Changing it changes predictions for array inputs.
• 18 LibreYOLO HF repos ship no LICENSE file (e.g. LibreRFDETR{n,s,m,l}-cls, LibreYOLO9{t,s,m,c}-cls, LibreYOLO9P2s-visdrone, LibrePPOCR{t,l}-ocr).

Done outside the repo (HF, no code change)

• Replaced the YOLOX (Megvii) LICENSE that was wrongly copied into 20 HF repos with each upstream's verbatim LICENSE: LibreRFDETR{n,s,m,l} and their -pose repos (roboflow/rf-detr), LibreRTDETR{r18,r34,r50,r50m,r101,l,x} (lyuwenyu/RT-DETR), LibreDEIM{n,s,m,l,x} (ShihuaHuang95/DEIM). Only the YOLOX repos mention Megvii now.