Three-Class Building Damage Classifier
Drone-image classifier for undamaged / damaged / collapsed buildings, comparing 12 model variants and a Grad-CAM-guided crop pipeline.
About this project
Pattern Recognition final project. The dataset combines drone images from the 2023 Turkey earthquake (UAV-TEBDE / Mendeley + Kaggle) and a Yushu, China research set into 9.6k train / 1.7k val / 4.7k test images across three classes. Both YOLOv8n-cls and YOLOv11n-cls were trained at three center-crop sizes (100 / 90 / 80%), softmax probabilities of each pair were averaged into soft-voting ensembles, and a Grad-CAM-guided crop dataset was generated from the best individual model (YOLOv8 @ 80%, highest macro-F1) and used to retrain both architectures. Because the test set is heavily class-imbalanced (damaged 2820, collapsed 1140, undamaged 802), macro-F1 — not accuracy — was the primary metric throughout.
- 12 model variants: 2 architectures × 3 crop sizes × (single + ensemble + CAM-crop)
- Grad-CAM-guided crop pipeline — threshold 0.4, 0.15 padding, fallback on empty mask
- Soft-voting probability averaging with alphabetical→canonical class index remapping
- MD5 deduplication + train/test leakage verification before training
Key features
Two YOLOs
v8n-cls + v11n-cls fine-tuned
Soft-vote ensemble
Probability averaging
Grad-CAM crop
Attention-guided dataset rebuild
Class-imbalanced
Macro-F1 over accuracy
Screenshots
Confusion matrices across center-crop models
Grad-CAM comparison on collapsed class
Accuracy vs macro-F1 across all variants
Per-class error rate heatmap
Tech stack
Other projects
Minecraft Mob Detection
Real-time object detection of 15 Minecraft mobs from live screen capture using YOLOv8s fine-tuned on a custom merged dataset.
No Car No Problem — Ride-sharing platform
Full ride-sharing app with customer/driver/admin/support roles, real-time chat, geospatial ride matching, and driver verification workflow.