VisionAId is an offline-first Android application designed to assist people with visual impairment by turning a commodity smartphone into a real-time visual assistant. The system integrates six on-device deep learning models running through ONNX Runtime, including metric monocular depth estimation, instance segmentation, and face detection.
- A distinctive feature is a few-shot pipeline for personal objects that allows users to photograph items from multiple angles for later location via augmented-reality markers and spatial audio.
- The system uses INT8 quantization on a Samsung Galaxy S21 Ultra to reduce depth latency from approximately 1200 ms to 491 ms.
- A custom banknote detector achieves an mAP@50 of 0.986, while metric depth calibration maintains below 1 cm of error within 3 meters.
By operating entirely on-device with optional cloud support for narrative descriptions, VisionAId addresses the limitations of existing assistive apps that rely heavily on cloud connectivity or dedicated hardware.