Project 01
YOLOv8-OBB Real-Time Object Detection Service
End-to-end Computer Vision pipeline: a 2-class oriented bounding-box (OBB) detector for bottles vs cans, fine-tuned on Colab T4 GPU and served externally through a FastAPI REST API.
- Fixed dataset class imbalance — discovered a 93% / 7% bottle/can split and implemented a stratified re-split (80/10/10) from scratch
- Fine-tuned YOLOv8-OBB — AdamW + cosine LR + mosaic/mixup augmentation → confidence 0.82–0.94 in real-world tests
- Designed the FastAPI server — /predict (JSON) + /predict/visualize (PNG) endpoints with auto-generated Swagger docs
- Real-time post-processing — temporal smoothing (STABILITY_FRAMES=3) + area filter to suppress false positives from webcam streams
Outdoor · 2 bottles + Pepsi can
Outdoor · 2 bottles + can
Outdoor · Pepsi can 0.90
Outdoor · Water bottle 0.94
Cafe · Coca-Cola can 0.94
In car · Bottle 0.94 + can 0.86
Python
PyTorch
Ultralytics YOLOv8
OpenCV
FastAPI
Colab T4 GPU
