Pyro-SDIS-FINAL: A 3LC-Curated Wildfire Smoke Detection Dataset

This is Pyronear's pyro-sdis dataset after a full label-curation pass in 3LC. Same images, cleaner labels. Every bounding-box correction in this dataset was found and applied using 3LC's interactive, model-in-the-loop curation workflow.

Curated with 3LC

3LC is a tool for interactive, model-guided data debugging and editing. The entire curation for this dataset was done in 3LC, using a tight train → inspect → fix → retrain loop:

1. Train a small YOLO model on the current labels.
2. Collect per-sample metrics back into 3LC (per-image TP/FP/FN, IoU, prediction confidence, embeddings).
3. Surface label-noise candidates in the 3LC dashboard — sort/filter by disagreement between model predictions and ground truth, cluster by embedding, flag neighbor-frame inconsistencies.
4. Edit labels row-by-row directly in the 3LC dashboard (add, delete, move, resize boxes; drop bad samples).
5. Commit the edits as a new table revision and retrain — repeat until the noise is gone.

Because 3LC tracks every edit as a versioned table revision, the whole curation history is reproducible and auditable.

What 3LC surfaced and fixed

• Missing ground-truth boxes — smoke the model detected confidently but that annotators had missed.
• Spurious boxes — fog mislabeled as smoke, and hallucinated annotations.
• Overlapping-box artifacts — the source dataset's SAM-based auto-labeling occasionally stacked duplicate boxes on one plume; 3LC's per-image IoU metrics made these easy to find and remove.
• Neighbor-frame inconsistencies — adjacent frames from the same camera whose labels disagreed on smoke presence or box count (3LC flags these directly).
• Drifted box geometry — positions/sizes corrected where auto-labels had slipped off the actual plume.

The validation set needed heavy curation too

The validation split was as noisy as the training split and required extensive correction — hundreds of box edits, additions, and deletions. This matters: benchmarking against the original pyro-sdis validation labels gives an unreliable signal, because the ground truth itself is wrong in many places. A model can be penalized for correct detections that the noisy val labels don't include, or rewarded for matching bad labels. The curated validation set here provides a much more trustworthy evaluation target.

What's in here

• 29,537 train + 4,099 val images (same splits as source pyro-sdis)
• 38,778 curated smoke bounding boxes
• 1 class: smoke (source fog boxes dropped for the standard 1-class formulation)
• YOLO format: class_id cx cy w h normalized coords in .txt files, alongside .jpg images
• Source imagery: ~40 fixed cameras deployed with SDIS (French fire service) partners across France, captured throughout 2024

Shipped as pyro-sdis-FINAL.tar.gz. Extract with tar -xzf pyro-sdis-FINAL.tar.gz:

pyro-sdis-FINAL/
├── data.yaml          # nc=1, names=['smoke']
├── images/{train,val}/*.jpg
└── labels/{train,val}/*.txt

Quick start (Ultralytics)

from ultralytics import YOLO

model = YOLO("yolov8n.pt")
model.train(data="pyro-sdis-FINAL/data.yaml", epochs=30, imgsz=640)

To reproduce the curation loop (train → collect metrics → inspect → edit → retrain) on your own data, see 3LC and its Ultralytics integration.

Attribution

• Source dataset: pyronear/pyro-sdis by Pyronear, a non-profit building open-source AI for wildfire early detection. Original imagery captured by cameras deployed with SDIS (Service Départemental d'Incendie et de Secours) partners in France.
• Curation: performed with 3LC.

License

Apache-2.0, matching the source dataset.

Citation

Please cite the Pyronear team for the source pyro-sdis data (HF dataset), and note that curation was done with 3LC.