Skip to content

Comparison of point cloud classification performance using 2D projection based and transformer based methods. Also explores adversarial attacks on these classes of models. Specifically, one-point and noise attacks.

License

Notifications You must be signed in to change notification settings

devashish-gupta/Pointy

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

22 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CS 4644/7643: Deep Learning | Spring 2023

Georgia Institute of Technology, Atlanta, GA

Final project by,

Devashish Gupta: devashish-gupta@gatech.edu
Divyaansh Singh: dsingh313@gatech.edu
Sanjay Josh: sjosh3@gatech.edu
Vastav Bharambe: vbharambe6@gatech.edu

Refer to driver.ipynb for all the main sections

Point Cloud Classification

Introduction

Point cloud data is an important type of data used in various applications such as autonomous vehicles, robotics, and virtual reality. Point cloud classification refers to the task of assigning a label to a given point cloud based on its features.

Dataset

In this notebook, we will be using the ModelNet40 dataset for testing our models. ModelNet10 is a widely-used benchmark dataset for evaluating 3D object classification algorithms. It contains 12,311 CAD models from 10 different categories, with each model represented as a point cloud.

ModelNet

Models

We will be implementing and comparing three different models for point cloud classification: SimpleView, 3DCTN, and a modified custom version of 3DCTN. SimpleView is a simple yet effective model that uses a fully connected network to classify point clouds. 3DCTN, on the other hand, is a more complex model that uses a 3D convolutional neural network (CNN) to classify point clouds. We will also be modifying the 3DCTN model to see if we can improve its performance.

Adversarial attacks and visualization

Additionally, we will also be testing each model against adversarial attacks, including one-point attacks and imperceptible noise addition attacks, to evaluate their robustness.

About

Comparison of point cloud classification performance using 2D projection based and transformer based methods. Also explores adversarial attacks on these classes of models. Specifically, one-point and noise attacks.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published