Symphony 3D2024

Pix2Vox Machine Learning Model Training

Training and optimizing a machine learning model to generate 3D voxels from 2D images using auto-encoder architecture for custom-made hearing aids manufacturing

Project Context

End-to-end solution for 3D prints custom-made hearing aids. The goal was to develop a machine learning model capable of generating 3D voxel representations from 2D images, enabling automated 3D reconstruction for hearing aid manufacturing.

Technical Implementation

Model Architecture

  • Auto-encoder based architecture for 2D to 3D conversion
  • Improved voxel resolution from 32x32x32 to 128x128x128
  • Multi-epoch training optimization for best results

Data Processing

  • Generated images from different angles for data augmentation
  • STL file processing for training data creation
  • Comprehensive understanding of model behavior and limitations

Results & Achievements

Model Resolution Enhancement

Successfully upgraded voxel resolution from 32³ to 128³, significantly improving 3D reconstruction detail

Data Augmentation Pipeline

Implemented comprehensive data augmentation from STL files with multi-angle image generation

Training Optimization

Optimized training process across multiple epochs for maximum performance

Performance Analysis

Achieved ~60% accuracy in testing phase - identified areas for future improvement

Technologies & Tools

Python
Machine Learning
Auto-encoder
Data Augmentation
STL Processing
Neural Networks

Challenges & Key Learnings

Model Architecture Optimization

Learned to modify and optimize auto-encoder architectures for specific 3D reconstruction tasks, including resolution enhancement techniques.

Data Augmentation Strategies

Developed expertise in creating effective training datasets from 3D models, including multi-angle image generation and STL file processing.

Performance Analysis & Iteration

Gained valuable experience in model evaluation and identifying areas for improvement when results don't meet production requirements.