fireviewer/rfdetr-medium-visible-flame-v2.5

RF-DETR Medium · visible flame specialist · v2.5

Quick card

Field	Reference
Use	Find visible flames inside probable-fire proposals.
Expected input	Proposed RGB ROI resized to 1088 px.
Example	A proposed ROI → flame boxes in crop pixels, then original-image reprojection.
Limits	Crop specialist; not an independently qualified full-image detector. Tiny-flame localization remains weak in the external diagnostic.
Contract	English pipeline · Reproducible loading

Choose a variant · Resource navigation · Update log

purpose and training

One-class visible_flame detector on real probable-fire crops, separating visible combustion from smoke. Initialize from rfdetr-medium-probable-fire-v2.5 selected E10 with a fresh optimizer. Input 1088 px; outputs are boxes in crop pixels, reprojected to original-image coordinates before global duplicate suppression. It is a crop specialist, not a separately qualified full-image detector.

Run 20261005-081147-visible-flame: 10 completed epochs, selected E8. RF-DETR 1.11.0 Medium, micro-batch 8 × accumulation 2 = effective 16, workers 2, seed 42; detector LR 3e-5, backbone LR 3e-6, 0.5-epoch warmup and cosine decay. Balanced rotation uses 25,920 crops per epoch, not the whole 221,168 prepared train crops: 16,848 real proposals (65%), 6,480 train-only missed-flame jitters (25%) and 2,592 wider contexts (10%). No replacement inside one epoch; deterministic rotation between epochs. Public v2.5-specialisation-visible-fire contains 137,102 train / 14,822 validation crops after rights filtering. It is a reproducibility subset, not the exact sampler sequence.

Cascade validation

8,288 original validation images and 17,450 real proposal ROIs; no GT fallback on validation. Stage 1 is frozen at its E10 checkpoint. Context is 30% per side, minimum crop side 128; stage-1 score ≥0.10, NMS 0.75, at most six proposals. Visible outputs are reprojected, globally suppressed with NMS 0.60 and capped at 100. AP uses permissive score ≥0.001 with standard COCO maxDets [1,10,100]; counts use the fixed comparison score 0.25 and IoU 0.50. Raw RF-DETR queries (~300) are candidates, not 300 plausible flames. No inference-context or threshold change was used to improve the reported epoch comparison.

Selected E8 metric E8	Value
mAP50–95 / AP50 / AP75	0.231496 / 0.435868 / 0.210436
AP-small / AR-small / AR100	0.101914 / 0.332917 / 0.507002
Tiny recall IoU50 (753 tiny GT)	0.583001
Multi-flame recall	0.653578
Precision / Recall / F1 @0.25	0.197784 / 0.711612 / 0.309536
TP / FP / FN	2,053 / 8,327 / 832
Smoke-only image FP rate (168 / 2,398)	0.070058
Other negative-image FP rate	0.155187
Matched box mean IoU / predicted-to-GT area ratio	0.753184 / 1.019856

Tiny = original box area / original image area ≤0.0025. COCO small is a different original-pixel criterion. Multi-flame recall aggregates matched GT on images containing multiple flames; it is not complete-scene recall. Mean IoU is conditional on matched positives. Precision remains low at the comparison threshold; these results do not certify unattended alerts. The observed 21.97 ms/ROI was a warm L4, three-batch validation timing excluding IO, preprocessing and other stages, not end-to-end latency.

Selection: 0.4 AP-small + 0.4 AR-small + 0.2 tiny recall −0.05 smoke-only image FP rate. Proposal-cache key a5cd7865f2ff51dcb80618369c20cb356bb1fcf642b3ff0794761fc5b8359529.

Original selected checkpoint SHA-256 840b5dd7b21b66805b6971a5ef3f8d2e9ecd894e33146dd3b1fefefdecfd941a; safe tensor SHA-256 47660fbbfcda5a4b5c1c5594fe16592fd9012354c83cf53a3095f37274c72485. All exported tensors are finite and round-trip identical. The earlier invalid trial without preserved native EMA callbacks is excluded.

Reproducible loading

The release contains tensor-only model.safetensors, a plain config.json, an explicit inference.py and checksum provenance. Download a revision-pinned snapshot, install the repository's requirements, and load it explicitly. No Hugging Face trust_remote_code is required. The optional smoke backbone fetches a pinned official Facebook Research Torch Hub source; inspect and cache that source for an offline deployment. RF-DETR 1.11 requires a native wrapper; the helper creates it locally from safe tensors and JSON, never from an external pickle.

from huggingface_hub import snapshot_download
import importlib.util
root = snapshot_download(REPOSITORY, revision=REVISION)
spec = importlib.util.spec_from_file_location('fireviewer_inference', root + '/inference.py')
fv = importlib.util.module_from_spec(spec)
spec.loader.exec_module(fv)
model, config = fv.load_model(root, device='cpu')  # or cuda

Set REPOSITORY to this repository ID and REVISION to a full commit SHA from its Files tab. The complete inference CLI and geometry are documented in PIPELINE_EN.md and PIPELINE_FR.md. This is an explicit PyTorch research interface, not a Transformers AutoModel export.

Scope, rights and limitations

These checkpoints are retained research results. The original TEST set was locked during selection; no final held-out TEST score is claimed. Exact-byte/decoded-image duplicate checks were performed upstream, but event-level independence and deployment-domain transfer remain unestablished. Public linked data are rights-filtered subsets of the retained training corpus, not a byte-for-byte reproduction of the full training set. Private MyData_Fire and undocumented With-Augmentation-Big-1 images are absent from public datasets. Publicly releasing a checkpoint does not establish deployment qualification.

The model/code release uses Apache-2.0; RF-DETR and DINOv2 upstream notices must be retained. Image rights are separate and source-specific. Predictions are visual estimates, with possible missed fire, smoke confusion, lighting artifacts and source annotation omissions. They are not a calibrated emergency alarm or geographic ignition estimate. See each linked dataset's notices before reusing images.

Official RF-DETR source.

LiteRT FP32 export

The selected weights above are also available as a verified LiteRT inference package. It includes real CPU PyTorch/LiteRT parity evidence, output/class/preprocessing contracts, an Android ModelConfig and a Cadryl StudioPack. Static batch 1; no new training or quantization; Android-device qualification remains open. Full English/French cascade and external benchmark.

Likes are always welcome! ❤️ If these resources help you, a little heart makes our day and helps others discover the work. Thank you!