sanjeevafk/glasseye-dataset

GlassEye 3-Way Unified Building Defect Dataset (640px)

The GlassEye 3-Way Unified Building Defect Dataset is a curated, multi-domain benchmark corpus designed for training computer vision models to detect structural building defects (cracks, spalling, efflorescence) across close-up façade photography and high-altitude aerial drone surveys.

• —Primary Repository: GlassEye GitHub
• —Model Checkpoint: `sanjeevafk/glasseye-yolo`
• —Image Count: 2,899 total training images + 289 held-out validation images + 200 untouched aerial drone test images
• —Resolution: 640px native longest-edge resolution
• —Format: Standard YOLO format (images/ + labels/ .txt files) wrapped in glasseye_3way_dataset_640.zip

Dataset Composition & Source Synthesis

This dataset unifies three complementary civil engineering and building inspection sources to bridge the domain gap between close-up ground photography and high-altitude drone captures:

Directory & Label Structure

The primary archive glasseye_3way_dataset_640.zip unpacks into standard Ultralytics YOLO format:

glasseye_3way_dataset_640/
├── data.yaml
├── train/
│   ├── images/  (2,899 .jpg files)
│   └── labels/  (2,899 .txt files)
├── val/
│   ├── images/  (289 .jpg files)
│   └── labels/  (289 .txt files)
└── test/
    ├── images/  (200 untouched UAV2K drone holdout images)
    └── labels/  (200 .txt files)

Class Mapping

• —Class ID `0`: defect (unified single-class structural anomaly covering cracks, spalling, and weathering)

Bounding Box Format (labels/*.txt)

Standard YOLO normalized coordinates: <class_id> <x_center> <y_center> <width> <height>

Usage & Quickstart

1. Download & Unpack in Python / Colab

from huggingface_hub import hf_hub_download
import zipfile

# Download dataset zip
zip_path = hf_hub_download(
    repo_id="sanjeevafk/glasseye-dataset",
    filename="glasseye_3way_dataset_640.zip",
    repo_type="dataset"
)

# Extract archive
with zipfile.ZipFile(zip_path, 'r') as zip_ref:
    zip_ref.extractall("./data")

2. Fine-Tune YOLO Model

from ultralytics import YOLO

# Load base model
model = YOLO("yolov8n.pt")

# Train on GlassEye dataset
results = model.train(
    data="./data/glasseye_3way_dataset_640/data.yaml",
    epochs=50,
    imgsz=640,
    batch=16,
    device=0
)

Citation & License

• —License: MIT License
• —Maintainer: Sanjeev (sanjeevafk)
• —Project: GlassEye Autonomous Aerial Drone Inspection System (Physical AI Hackathon 2026)