Exploration and Assessment of Parametric and Non-Parametric Networks for 3D Point Cloud Classification
Alternative Title: Classifying 3D Common Household Objects Using Deep Learning Approaches
By Jeffrey Li and Ning-Yuan Li
Last Updated: December 16, 2023
Processing 3D point cloud data is important for applications in self-driving cars, robotics, and virtual and augmented reality. Qi et al. (2017) introduced the PointNet architecture which performs classification tasks after being trained on point cloud data. The group based their design decisions on three properties of point clouds, namely permutation invariance, transformation invariance, and interactions among points. On the other hand, Zhang et al. (2023) offers a newer non-parametric approach in solving the same problem. Non-parametric building blocks are stacked across multiple stages to construct a pyramid hierarchy. We will be exploring and evaluating both models. Firstly, we focused on implementing both networks and testing them on a smaller point cloud dataset based off ModelNet40. Secondly, we performed a series of tests on robustness on PointNet, as we augment the data during inference. Lastly, we will visually assess the internal encodings of the multistep hierarchical layers inside the non-parametric encoder of the Point-NN to understand how it captures spatial representations for classification tasks.
-
papers: contains the original papers that we base our project from
-
pointnet: contains relevant code for PointNet model
- data_loader.py: load point cloud data
- eval_cls.py: assess model accuracy on validation set
- models.py: implementation of PointNet in PyTorch, including variants
- utils.py: used for saving models during checkpoints
- train.py: training and validation loops
- Visualize-PointNet-Results.ipynb: visualizing plots and results from robustness testing
-
pointnn: contains relevant code for PointNN model
- fps.py: performs farthest point sampling on point cloud data
- knn.py: performs k-nearest neighbors classification on point cloud data
-
report: contains final report submission
We developed this primarily using Python, PyTorch, and CUDA GPU.
Packages to Install:
pip install matplotlib
pip install numpy
pip install torch
pip install argparse
pip install torchsummary
pip install torchviz
pip install scikit-learn
pip install tensorboard
Data and source of inspiration for project available here.
The classification network takes n points as input, applies input and feature transformations, and then aggregates point features by max pooling. The output is classification scores for k classes.
Non-Parametric Encoder (Left). Zhang et al. (2023) utilized trigonometric functions to encode raw Point Clouds points into high-dimensional vectors in PosE. The vectors then pass through a series of hierarchical non-parametric operations, namely FPS, k-NN, local geometric aggregation, and pooling, where they will be encoded into a global feature
Sample point cloud demonstrating a chair undergoing rotations (left to right): 0 degrees, 5 degrees, 30 degrees, 45 degrees, 90 degrees. The model incorrectly classifies the chair starting at the 30 degrees rotation.
Sample point cloud demonstrating a chair undergoing sampling (left to right): 10000 points, 7500 points, 5000 points, 2500 points, 1000 points, 500 points, 100 points. The model correctly classifies the chair at all different samplings.
PointNet Training and testing curves over 250 epochs.
| Number of Samples | Accuracy | Degree of Rotation | Accuracy |
|---|---|---|---|
| 10000 | 98.3% | 0° | 98.3% |
| 7500 | 98.3% | 5° | 98.1% |
| 5000 | 98.3% | 30° | 73.6% |
| 2500 | 98.1% | 45° | 43.2% |
| 1000 | 97.5% | 90° | 25.3% |
| 500 | 97.0% | ||
| 100 | 94.6% |
Sample point clouds undergoing four FPS iterations (left to right): 10000 points, 5000 points, 2500 points, 1250 points, 625 points.
Sample point clouds in the nth stage of the multistage hierarchy undergoing FPS and k-NN with k = 90 (from left to right): the point cloud before FPS, the point cloud after FPS, and k-NN where red indicates the nearest neighbors (cluster) of a selected point in the cloud (after FPS).
Point cloud processing is important in several fields. In this project, we examined and assessed two different methods for classifying point clouds on a point cloud dataset. The PointNet architecture provides a way to learn global and local point features while also achieving permutation invariance, whereas the Point-NN network offers a way to capture features, like spatial patterns and structures, through a non-parametric approach. Through our experiments, we built and implemented both networks to understand the different mechanisms that allow them to perform classification tasks. Though PointNet achieves higher accuracy when evaluated on the test dataset, we acknowledge the ability for Point-NN to achieve reasonable accuracy with no training and substantially less computational resources. Additionally, we performed robustness tests, specifically data corruption and rotations, in order to evaluate PointNet's ability to classify objects despite changes to orientation and structure. From our tests, we found that PointNet is capable of recognizing global structures of objects despite missing points, but it suffers when large rotations are applied to the point clouds. Lastly, we visually analyzed the internal representations of the multistep hierarchical layers inside the non-parametric encoder to understand how Point-NN captures meaningful representations in point cloud data.
We have several directions in which we can expand upon our existing work. Firstly, there are several relevant and more advanced architecture capable of processing point cloud data, while also performing other vision tasks, such as segmentation and object detection, that we can further explore. PointNet is a stepping stone to more exciting networks, such as PointNet++, DGCNN, and Point Transformers, in which several changes and design decisions have been made to account for the drawbacks of the original PointNet. Secondly, Point-NN's utility can be expanded upon as a plug-and-play module to boost existing learnable 3D models without further training. Applying Point-NN's ability to capture spatial representations will enhance the shortcomings to baseline models such as PointNet. Thirdly, acquiring access to greater computing resources will allow us to build upon our existing work through being able to keep and run the Transformation-Network as well as the full ModelNet40 point cloud dataset.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent