thanhhoangnvbg/fire-vn-yolo11seg-v1

Fire VN YOLO11-Seg v1

Fire VN YOLO11-Seg v1 is a prepared fire and smoke segmentation dataset focused on Vietnamese visual contexts. It is intended for training and evaluating YOLO11 segmentation models for early fire/smoke detection in scenes such as urban alleys, residential areas, distant fire/smoke events, hard negative visual distractors, and normal ambient scenes.

This dataset package is distributed as a ready-to-train YOLO segmentation zip:

fire_vn_yolo11seg_v1.zip

Dataset Summary

• Format: YOLO segmentation
• Classes: smoke, fire
• Total images after training-only slicing: 15,036
• Total annotations: 30,054
• Access: public dataset repository
• Primary use case: fire/smoke detection and segmentation
• Primary target metric: mAP50(B) >= 0.70
• Secondary metrics: mAP50(M), recall, false-positive rate, small-object recall

Data Groups

The dataset was prepared from five source groups:

Group	Description	Purpose
01_positive_standard	Clear fire/smoke images	Learn core fire/smoke appearance
02_Alley_Context	Vietnam-specific alley/urban context	Improve robustness in local environments
03_Negative_Hard_Samples	Hard negatives that may look like fire/smoke	Reduce false positives
04_SAHI_Small_Objects	Distant or small fire/smoke objects	Improve early/small-object detection
05_Ambient_Context_Null	Normal ambient scenes	Reduce background false alarms

Split Statistics

Split	Total Images	Total Annotations
Train	13,246	25,486
Valid	957	2,214
Test	833	2,354

Detailed group distribution:

Split	Group	Source Images	Final Images	Annotations
Train	01_positive_standard	7,873	7,873	20,118
Train	02_Alley_Context	352	1,558	3,751
Train	03_Negative_Hard_Samples	100	486	0
Train	04_SAHI_Small_Objects	70	348	1,594
Train	05_Ambient_Context_Null	2,981	2,981	23
Valid	01_positive_standard	633	633	1,964
Valid	02_Alley_Context	44	44	171
Valid	03_Negative_Hard_Samples	20	20	0
Valid	04_SAHI_Small_Objects	20	20	79
Valid	05_Ambient_Context_Null	240	240	0
Test	01_positive_standard	543	543	2,056
Test	02_Alley_Context	44	44	210
Test	03_Negative_Hard_Samples	20	20	0
Test	04_SAHI_Small_Objects	20	20	86
Test	05_Ambient_Context_Null	206	206	2

Preparation Method

The source Roboflow exports were provided as five separate COCO-segmentation groups. The preparation pipeline:

1. Read each source group independently.
2. Ignored the original Roboflow train/validation/test split.
3. Created a deterministic balanced split with seed 42.
4. Preserved all source images in only one split to avoid leakage.
5. Applied slicing only to the training split.
6. Kept validation and test images unsliced for fair evaluation.
7. Converted the result into YOLO segmentation format.

Training-only slicing was applied as follows:

• 01_positive_standard: original images only
• 02_Alley_Context: original images plus controlled slices
• 03_Negative_Hard_Samples: original images plus capped negative slices
• 04_SAHI_Small_Objects: original images plus small-object-focused slices
• 05_Ambient_Context_Null: original images only

Directory Structure

After extracting the zip:

fire_vn_yolo11seg_v1/
  data.yaml
  images/
    train/
    valid/
    test/
  labels/
    train/
    valid/
    test/
  coco/
    train/
    valid/
    test/
  metadata/
    split_summary.csv
    source_manifest.csv

data.yaml:

path: .
train: images/train
val: images/valid
test: images/test
names:
  0: smoke
  1: fire

Usage With Ultralytics

Download and extract the dataset zip from this public repository, then train a YOLO segmentation model:

pip install ultralytics
yolo segment train model=yolo11s-seg.pt data=fire_vn_yolo11seg_v1/data.yaml imgsz=640 epochs=80

For the project training pipeline, use:

python scripts/train_yolo11_seg.py --profile kaggle_s --device 0 --val-test
python scripts/train_yolo11_seg.py --profile vertex_final --device 0 --val-test

Evaluation Guidance

Recommended metrics:

• mAP50(B) as the primary detection metric
• mAP50(M) for mask quality
• recall for fire and smoke
• false-positive rate on groups 03 and 05
• small-object recall on group 04

The test split should be treated as the internal benchmark for comparing model profiles and architectures trained within this project.

Limitations

• This is a project-specific dataset, not a general-purpose fire benchmark.
• The data is optimized for Vietnam-focused fire/smoke detection scenarios.
• The current test split is an internal benchmark from the same dataset construction pipeline; an additional external Vietnam fire benchmark is recommended for final real-world validation.
• Some source annotations were normalized during conversion to YOLO segmentation format.

Intended Use

This dataset is intended for:

• research and development of fire/smoke detection systems,
• YOLO11 segmentation training,
• benchmark comparisons between internal model variants,
• fine-tuning and improving early warning models for Vietnamese environments.

It should not be used as the only evidence for production deployment safety. Real-world camera validation and operational false-alarm testing are still required.