omerfarooq223/FireWatch-AI

FireWatch AI — Dual Detection & Segmentation YOLO Models

This repository provides fine-tuned YOLOv8 weights powering the FireWatch AI safety monitoring and incident management framework.

Live inference microservice: omerfarooq223/firewatch-yolo-inference

---

Model Checkpoints

This repository hosts two distinct weights files:

1. detector-best.pt (~262 MB)

   • Task: Bounding-box Object Detection
   • Base Architecture: YOLOv8l (depth=1.0, width=1.0, nc=2)
   • Classes:
     • 0: fire
     • 1: smoke
   • Role: Localizing flame and smoke bounding boxes in surveillance video frames.

2. seg-best.pt (~6.7 MB)

   • Task: Polygon Instance Segmentation
   • Base Architecture: YOLOv8n-seg (depth=0.33, width=0.25, nc=1)
   • Classes:
     • 0: fire
   • Role: Extracting pixel-level flame boundaries and calculating fire surface area. (Does not segment smoke).

---

Evaluation Metrics (Validation Split)

Below are the exact metrics recorded in the training checkpoint logs. Metrics associated with the saved best.pt checkpoint are distinguished from individual peak-epoch measurements:

1. Detection Model (detector-best.pt — YOLOv8l)

• Training duration: 17 epochs completed (target was 50; training stopped at Epoch 17).
• Saved Checkpoint (best.pt / Epoch 17):
  • metrics/mAP50(B): 79.91% (0.79907)
  • metrics/mAP50-95(B): 54.23% (0.54227, peak fitness for the run)
  • metrics/precision(B): 78.84% (0.78842)
  • metrics/recall(B): 70.41% (0.70407)
  • val/box_loss: 1.08875
  • val/cls_loss: 0.87772
  • val/dfl_loss: 1.28927
• Peak Metric Comparison:
  • Box mAP@50 reached its individual peak at Epoch 16 with 79.92% (0.79921, with Precision: 78.88%, Recall: 70.41%, mAP50-95: 54.15%), before best.pt was finalized at Epoch 17 on overall fitness.

2. Segmentation Model (seg-best.pt — YOLOv8n-seg)

• Training duration: 20 epochs completed. The saved best.pt corresponds to the final epoch evaluated (Epoch 20, fitness 1.40684).
• Saved Checkpoint Metrics (Epoch 20):
  • Mask Metrics (Contour Segmentation):
    • metrics/mAP50(M): 92.29% (0.92287)
    • metrics/mAP50-95(M): 62.29% (0.62285)
    • metrics/precision(M): 90.50% (0.90495)
    • metrics/recall(M): 86.26% (0.86260)
  • Bounding Box Metrics:
    • metrics/mAP50(B): 96.50% (0.96503)
    • metrics/mAP50-95(B): 78.40% (0.78399)
    • metrics/precision(B): 92.11% (0.92114)
    • metrics/recall(B): 89.65% (0.89654)
  • val/box_loss: 0.66318 | val/seg_loss: 1.44122 | val/cls_loss: 0.50684 | val/dfl_loss: 0.92107

---

Dataset Details

The detector and segmentation models were trained on separate dataset configurations:

A. Detection Dataset (detector-best.pt)

• Upstream Benchmark: Trained on the FASDD (Flame And Smoke Detection Dataset) — FASDD_CV computer vision benchmark subset (Wang et al., 2023).
• Benchmark Size: 95,314 total images (full official FASDD_CV benchmark across all splits combined).
• Reference Upstream Benchmark Splits:

  Note: The figures below represent the official published FASDD_CV benchmark partition ratios provided for reference. The exact split counts ingested by the specific training notebook script were managed via data.yaml and are not recorded in the checkpoint dictionary.

  • Reference Training Partition (~50%): ~47,657 images
  • Reference Validation Partition (~33.3%): ~31,771 images
  • Reference Test Partition (~16.7%): ~15,886 images
• Annotations: Dual-class bounding box labels for fire and smoke.

B. Segmentation Dataset (seg-best.pt)

• Subset Configuration: Trained using a derived fire segmentation configuration referenced in the training metadata as /kaggle/working/clean_fire_seg.yaml.
• Subset Size & Splits: Not recorded. The derived dataset files, image manifests, and exact train/val sample counts were generated in an ephemeral Kaggle working environment and are not embedded in the model checkpoint.
• Target Scope: Single-class polygon annotations for fire only (smoke masks are not included).

---

Training Configuration

Parameter	Detection (detector-best.pt)	Segmentation (seg-best.pt)
Model Architecture	YOLOv8l (Large)	YOLOv8n-seg (Nano Segmentation)
Pretrained Checkpoint	Resumed / fine-tuned from last.pt	yolov8n-seg.pt
Input Image Size	640 × 640	640 × 640
Batch Size	16	16
Epochs Trained	17 epochs	20 epochs
Optimizer	Auto (SGD/AdamW, momentum: 0.937, weight_decay: 0.0005)	Auto (SGD/AdamW, momentum: 0.937, weight_decay: 0.0005)
Dropout	0.5	0.0
Hardware	Kaggle GPU (device: 0, GPU model: not recorded)	Kaggle GPU (device: 0, GPU model: not recorded)
Ultralytics Version	8.4.41	8.4.45
Run Name / Project	yolov8l-FASDD (Fire-Smoke-FASDD)	final_fire_segmentation

---

How to Run Inference

Installation

pip install ultralytics huggingface_hub opencv-python pillow

Python Inference Example

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

# 1. Download weights from Hugging Face Hub
det_path = hf_hub_download(repo_id="omerfarooq223/FireWatch-AI", filename="detector-best.pt")
seg_path = hf_hub_download(repo_id="omerfarooq223/FireWatch-AI", filename="seg-best.pt")

# 2. Instantiate YOLO models
det_model = YOLO(det_path)
seg_model = YOLO(seg_path)

image_path = "test_frame.jpg"

# 3. Detect fire and smoke bounding boxes
det_results = det_model.predict(source=image_path, conf=0.40, imgsz=640)[0]
for box in det_results.boxes:
    cls_id = int(box.cls[0])
    label = det_results.names[cls_id]
    conf = float(box.conf[0])
    xyxy = box.xyxy[0].tolist()
    print(f"Detected {label} ({conf:.2%}) at {xyxy}")

# 4. Extract polygon masks for flame regions
seg_results = seg_model.predict(source=image_path, conf=0.35, imgsz=640)[0]
if seg_results.masks is not None:
    for polygon in seg_results.masks.xy:
        print(f"Fire polygon coordinates count: {len(polygon)}")

---

Potential Limitations & Operational Considerations

The following items represent theoretical considerations and known architectural constraints; untested scenarios have not been quantitatively evaluated:

• Potential Optical Confounds: In vision-based monitoring systems, strong directional solar glare, low-angle sunrise/sunset backlighting, reflections off polished surfaces, or high-intensity industrial sodium lamps may pose risks of false positive activations. These environmental conditions have not been systematically tested.
• Atmospheric Particulates vs. Smoke: Thin or diffuse smoke plumes share visual features with dense fog, steam plumes, and suspended dust. Performance in heavy fog or steam environments has not been benchmarked.
• Low-Light / Nighttime Deployment: Detection capability in poorly lit conditions without active illumination or complementary thermal/infrared sensors remains unverified.
• Architectural Scope Constraint: The segmentation checkpoint (seg-best.pt) was trained strictly for flame instance segmentation (fire); it does not segment smoke regions. Use detector-best.pt when smoke monitoring is needed.

---

Licensing

• Model Weights: Ultralytics YOLOv8 weights are distributed under the AGPL-3.0 License.
• Application Framework: MIT License.
• FASDD Dataset: Distributed under academic research and non-commercial benchmarking terms.