LassoNet: Deep Lasso-Selection of 3D Point Clouds

Zhutian Chen, Wei Zeng, Zhiguang Yang, Lingyun Yu, Chi-Wing Fu, and Huamin Qu IEEE VIS 2019
Paper Code Pipeline Dataset Examples

Abstract

Selection is a fundamental task in exploratory analysis and visualization of 3D point clouds. Prior researches on selection methods were developed mainly based on heuristics such as local point density, thus limiting their applicability in general data. Specific challenges root in the great variabilities implied by point clouds (e.g., dense vs. sparse), viewpoint (e.g., occluded vs. non-occluded), and lasso (e.g., small vs. large). In this work, we introduce LassoNet, a new deep neural network for lasso selection of 3D point clouds, attempting to learn a latent mapping from viewpoint and lasso to point cloud regions. To achieve this, we couple user-target points with viewpoint and lasso information through 3D coordinate transform and naive selection, and improve the method scalability via an intention filtering and farthest point sampling. A hierarchical network is trained using a dataset with over 30K lasso-selection records on two different point cloud data. We conduct a formal user study to compare LassoNet with two state-of-the-art lasso-selection methods. The evaluations confirm that our approach improves the selection effectiveness and efficiency across different combinations of 3D point clouds, viewpoints, and lasso selections.

Pipeline and Network Architecture

missing
Figure 3. Pipeline. LassoNet consists of three stages: In Interaction Encoding stage, we associate point cloud with viewpoint and lasso through 3D coordinate transformation and naive selection; In Filtering and Sampling stage, we reduce the amount of points for network processing through intention filtering and farthest point sampling. Lastly, we build a hierarchical neural network in Network Building stage.
missing
Figure 4. Overview of network building. The DNN network is built upon (a) PointNet ref_pn, and we employ a hierarchical structure that generates more local and global features using (b) abstraction and (c) propagation components.

Dataset

missing
Figure 5. Exemplar annotation records for point clouds in ShapeNet (left) and S3DIS (right): target and interfering points are colored in yellow and blue respectively, while lassos are in red color.
Dataset #Point Clouds #Targets #Records
ShapeNet 2,332 6,297 19,432
S3DIS 2,72 4,018 12,944
Statistics of lasso-selection records.

The table presents statistics of lasso-selection records. In total, we have collected 19,432 lasso-selection records for 6,297 different parts of target points in ShapeNet point clouds, and 12,944 records for 4,018 different parts of target points in S3DIS point clouds.

The dataset can be downloaded from our github repo>.

Examples

Reference

Charles R. Qi, Hao Su, Kaichun Mo, Leonidas J. Guibas. "PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation." In Proc. IEEE CVPR. 2017.