2026 · Computer Vision
Aerial Dumpsite Detection
A Faster R-CNN baseline for finding illegal waste dumpsites in aerial imagery.
- Aerial imagery
- Faster R-CNN
- Dumpsite detections
Working baseline; demo and benchmarks in progressGitHub ↗
Problem
Illegal dumping is costly to monitor manually across large areas.
My role
Author of the repository.
Approach
- Fine-tune torchvision Faster R-CNN (ResNet50-FPN) on two public datasets: Global Dumpsite Test Data and AerialWaste (as negatives).
- Config-driven CLI for train, infer and benchmark; detection metrics (IoU, precision/recall, mAP) implemented from first principles.
- FastAPI + Gradio demo scaffolded; ONNX export verified.
Challenges
- True negatives are written as empty label files rather than fake full-image boxes.
Results
- No benchmark accuracy is claimed: the README presents it as a first baseline with a roadmap.
Stack
- Python
- PyTorch
- torchvision
- Albumentations
- FastAPI
- Gradio