Adaptive Shadow Detection System (ASDS) is a context-aware, lightweight neural network framework designed to improve assistive navigation for visually impaired individuals.
The system reduces false positives caused by shadows and enables real-time deployment on mobile devices using efficient MobileNet-inspired architecture.
The complete research paper is available here:
- Lightweight MobileNet-inspired architecture
- Context-aware environmental classification
- Adaptive threshold learning
- Temporal smoothing for stability
- Physics-informed shadow validation
- 8-bit quantized mobile optimization
- Real-time performance (22 FPS on Snapdragon 778G)
The ASDS pipeline includes:
- Multi-scale Feature Extraction
- Context Analysis Module
- Shadow Prediction Engine
- Temporal Learning Module
- Mobile-Optimized Real-time Processing
| Metric | Value |
|---|---|
| Accuracy | 0.8892 |
| F1-score | 0.8564 |
| AUC | 0.9178 |
| FPS (Mobile) | 22 FPS |
| Model Size | ~7.3 MB |
- ISTD (Image Shadow Triplets Dataset)
- UCF Shadow Dataset
- Python 3.8
- PyTorch
- ONNX
- Mobile 8-bit Quantization
- Ubuntu 18.04
- Snapdragon 778G (Deployment)
- Paulami Sahu
- Kritika Varyani
- Mallu Vineela Reddy
Vellore Institute of Technology, India