SmokeMM

SmokeMM is a smoke semantic segmentation benchmark assembled for robust evaluation across diverse smoke scales, scenes, illumination, weather, and smoke-like distractors. This release contains the fixed splits used in the paper Multi-Agent Scene-Adaptive Semantic Prior Guidance for Robust Smoke Semantic Segmentation.

Contents

• 27,899 image-label pairs
• 18,061 smoke-positive samples
• 9,838 no-smoke samples
• Binary PNG labels: 0 is background and 255 is smoke
• Fixed train, validation, and test split manifests

Split	Total	Smoke	No smoke
train	20,681	13,632	7,049
validation	2,383	1,599	784
test	4,835	2,830	2,005

Layout

data/
  train/images/part-*/  train/labels/part-*/
  validation/images/  validation/labels/
  test/images/        test/labels/
splits/
  train.csv
  validation.csv
  test.csv

Every image and label is named by the same sample_id; the larger training split is divided into 5,000-file directory shards. The CSV manifests use paths relative to the repository root and also retain the source dataset, positive/negative flag, smoke area ratio, and scale category required to reproduce the reported evaluation protocol.

Sources and usage

SmokeMM consolidates samples from SmokeSeg, UnetSmoke, and D-Fire and adds the standardized binary segmentation labels and fixed evaluation protocol described in the paper. Upstream data terms and citations continue to apply to source imagery; therefore this repository uses license: other instead of assigning a new blanket license to all source images. D-Fire is distributed by its authors under CC0 1.0. Users should consult the corresponding upstream releases before redistribution or commercial use.