Subset selection modeling for crowded point clouds
Progress in 3D sensing devices makes the capturing of 3D data possible and there is also a growing number of 3D shape repositories available online. These data are usually represented in the form of point clouds. However, processing a point cloud can be challenging, since the data are huge, which af...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis-Master by Coursework |
| Language: | English |
| Published: |
Nanyang Technological University
2020
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/140898 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | Progress in 3D sensing devices makes the capturing of 3D data possible and there is also a growing number of 3D shape repositories available online. These data are usually represented in the form of point clouds. However, processing a point cloud can be challenging, since the data are huge, which affect the computational cost, power consumption and communication load significantly. The situation becomes more serious when raw point clouds data contain outliers or noise through acquisition from 3D sensors. Therefore, a point cloud sampling method which can generate a simplified point cloud that is optimized for subsequent tasks is craving. This dissertation deeply understands and analyzes the latest point clouds processing approaches and proposes an adaptive sampling method which adaptively adjusts the points sampled by Learning to Sample beyond the entire point cloud and guarantees a minimal degradation in performance of downstream classification tasks.
We compare the performance of several data point subset selection methods, including random sampling, FPS, S-NET, SampleNet as well as the proposed Adaptive Sampling based on Learning to Sample, over the point clouds classification task with various sample sizes. We adopt ModelNet40 as the dataset and PointNet++ as the task network. To further verify the robustness, we replace a certain number of points with random noise to evaluate the performance of Adaptive Sampling in presence of outliers. Experimental results indicate that Adaptive Sampling based on Learning to Sample is a potential method for subset selection of point clouds while optimizing the objective function of a downstream task. |
|---|