sanjeevafk/glasseye-yolo

GlassEye YOLOv8n (BFDD + CUBIT Combined Checkpoint)

glasseye-yolo-bfdd-cubit-v1 is a fine-tuned YOLOv8n model built for autonomous building façade defect detection within the GlassEye inspection pipeline.

It is trained on a combined dataset of real-world façade defect imagery from BFDD (Building Façade Defect Dataset) and CUBIT (Concrete Ultrasound / Building Inspection dataset), unified under a single binary defect class.

Model Summary

• Base Architecture: YOLOv8n (yolov8n.pt, 3.2M parameters)
• Framework: PyTorch (torch 2.13+cu130) / Ultralytics 8.4.120
• Class Map: 0: defect (binary classification for façade surface and structural defects)
• Input Resolution: 320 × 320 px
• Model Checkpoint Size: 6.2 MB (best.pt)
• SHA-256: df78a4f90c776d8f122ec02972a193f91b1f4fc3916f1d78a4f51bb2e05445cf

---

Benchmark & Performance

Evaluated strictly across three separate test sets using standard IoU thresholds (IoU = 0.50, confidence = 0.20, imgsz = 320):

1. BFDD Held-Out Test Benchmark (Primary Ground-Truth)

149 untouched images, 1,033 mask-derived ground truth boxes.

Model Checkpoint	mAP@50	mAP@50-95	Recall
Synthetic Baseline (glasseye-yolo-v1)	0.0131	0.0099	0.0165
BFDD-only (glasseye-yolo-real-bfdd-v1)	0.0940	0.0429	0.1413
BFDD + CUBIT (glasseye-yolo-bfdd-cubit-v1)	0.1512	0.0707	0.2014

Improvement: BFDD+CUBIT achieves +60.8% mAP@50, +64.8% mAP@50-95, and +42.5% Recall over the BFDD-only model on untouched held-out building façade images.

2. CUBIT Test Benchmark

701 images, 5,085 derived polygon boxes.

Model Checkpoint	Full Test mAP@50	Full Test Recall	Far-Subset mAP@50 (≥10 frames dist)
Synthetic Baseline	0.0031	0.0069	0.0000
BFDD-only	0.0211	0.0350	0.0164
BFDD + CUBIT	0.1990	0.2061	0.1279

3. UAV2K Test Benchmark (Aerial Out-of-Domain)

200 images, 527 derived boxes (high-resolution aerial drone captures).

Model Checkpoint	mAP@50	Recall	True Positives
Synthetic Baseline	0.0000	0.0000	0
BFDD-only	0.0000	0.0000	0
BFDD + CUBIT	0.0132	0.0133	7

---

Training Details

• Dataset: data/bfdd_cubit_binary_v1 (Train: 1,299 images [600 BFDD + 699 CUBIT val-as-train]; Val: 89 BFDD images; Test: 850 images).
• Optimizer: AdamW
• Epochs: 32
• Batch Size: 4
• Image Size: 320 px
• Seed: 20260815 (deterministic)
• Augmentations: Horizontal flip (0.5), HSV-H (0.01), HSV-S (0.2), HSV-V (0.15), Mosaic (0.0).

---

Quickstart & Usage

1. Installation

pip install ultralytics torch huggingface_hub

2. Run Inference in Python

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

# Download model from Hugging Face
model_path = hf_hub_download(
    repo_id="sanjeevafk/glasseye-yolo-bfdd-cubit-v1",
    filename="best.pt"
)

# Load model
model = YOLO(model_path)

# Predict on an image or video
results = model.predict(
    source="path/to/facade_image.jpg",
    conf=0.20,
    iou=0.45,
    imgsz=320,
    save=True
)

for r in results:
    for box in r.boxes:
        print(f"Detected {model.names[int(box.cls)]} at {box.xyxy.tolist()} (conf: {float(box.conf):.3f})")

3. Ultralytics CLI

# Run detection directly
yolo predict model=best.pt source="https://example.com/facade.jpg" conf=0.20 imgsz=320

---

Project Context

This model is integrated into GlassEye, a deterministic façade-inspection pipeline featuring YOLO object detection, panel projection mapping, event logging, and advisory VLM second opinions.

• Live Application: glasseye-td75.onrender.com
• Source Repository: github.com/sanjeevafk/glasseye