Visionariez: Autofocus Smart Glasses

Team 7

Image: https://eccvision.com/

​

Presbyopia:

age-related loss of the eyes' ability to focus actively

Image: https://www.nei.nih.gov/

​

Pain Points

​

Image: https://www.nei.nih.gov/

​

Limited Fields of Vision

Switching Glasses

Forgetting Glasses

Requirements

​

Image: https://www.nei.nih.gov/

​

Wearable

Real-Time

Wide Focal Range

Ease of Use

Autonomous Mode

Software Pipeline

Object Detection

Depth Estimation

User inference for manual adjustment

User manually inputs preferred zoom in/out level in app

Data processing and inference on Raspberry Pi 4B

Microprocessor sends control signal to motor drivers

Thread I

Thread II

Model Training

1. Customized Dataset
1. Manual classification and Annotation
2. Model Accuracy:
• Training Accuracy: 78.625%
• Validation Accuracy: 50.24%

Model Deployment

1. Latest ArmV8+ Architecture not compatible with the Pytorch Library
1. ❎ Download the latest wheels
2. ❎ Build Pytorch from Scratch

​

• Convert all the pytorch models to the tensorflow lite
• Improve the latency
• Encapsulate the model details

​

• Performance:
• 0.86s to run one frame once the model is loaded up and thread is initialized
• Accuracy is not good enough - Depth model running on the spy cameras is not reliable
• Corner cases

Software Improvement

1. Improve the model accuracy
1. Enlarge the dataset
2. Simplify the inputs
3. Add more layers

​

• Add a frame buffer structure -> Deque with 30 frames depth
• Every 10 frames, call the function to adjust the glass magnification level
• The magnification is set to the level corresponding to the majority model output value

​

• Asynchronous FIFO?
• Pre-processing model keeps processing new frames and puts the result to the front
• Vision model(Our own) deques the frame buffer and process at its own speed
• Not serialized process anymore

System Architecture

Manual Mode

Manual Mode

How it works?

Meeting Customer Requirements:

​

• Wide Focal Range
• Real-Time
• Cost Effective

​

Wide Focal Range ✓

• Diopter Range of -4.00 to 4.00
• Increments of 0.5 Diopter
• Improve precision, accuracy through gear meshing

Real-Time ✓

• 1-2 Seconds

Cost Effective ✓

• $350 vs. $150-$500+

Meeting Customer Requirements:

​

• Wearable
• Ease of Use

​

Wearable ✓

• Lightweight
• Tethered Connection
• Explore bluetooth
• In-house battery (overnight charging)

Ease of use ✓

• Calibration Process/Adjustment Period
• Clear, customer tested app
• Improve mechanical precision

Thank You! Any Questions?