Project Description
Deep architecture for comparing 3D models.
The project builds upon various prior research methods and components for a content-based retrieval system utilized in managing manufacturing datasets. The Shokoufandeh group has developed novel methods in computer vision to partially automate reconstructing 3D objects from a series of lower dimension data sets. Such computational systems should be capable of object browsing, query processing, and interaction of 3D object composition/decomposition mechanisms. Special emphasis will be devoted to the task of 3D object classification and RGB-D Object-to-CAD retrieval using deep machine learning. As part of these activities, SMREU students will become familiar with various deep learning approaches for 3D object reconstruction, classification, segmentation and RGB-D object-to-cad retrieval methods. Voxelization will be used to convert unstructured geometric data to regular 3D grid. 3D convolutional layers will be used to extract the features. Voxnet architecture will be implemented using Keras and Tensorflow. The figure provides an overview of an ample pipeline for comparing two 3D models using a sample deep neural network architecture.
Research Goals
- Learn about various deep learning approaches
- Familiarize with 3D object classification, segmentation and RGB-D object-to-cad retrieval methods.
- Convert unstructured geometric data to regular 3D grid
Learning Goals
- How to develop a deep learning model
- How to conduct 3D object classification, segmentation and RGB-D object-to-cad retrieval
- How to use 3D convolutional layers to extract the features
Group Conducting Research
Shokoufandeh Lab: https://drexel.edu/cci/about/directory/S/Shokoufandeh-Ali/