Abstract:We developed graph analysis tools processing on Maximal Disjoint Ball Decomposition (MDBD) for shape analyzing tasks, and the consistency of results across any valid geometric representations is guaranteed by the nature of MDBD.We designed a DNN based on graph CNN that achieves a comparable classification accuracy with that of other leading classifiers on a widely used benchmark dataset, but with a significantly lower number of parameters.We proposed a graph kernel based on random walks that exploits the hierarchical structure of MDBD and has linear time complexity.We implemented a graph clustering technique (Adversarially Regularized Variational Graph Autoencoder + KMeans), and showed that we can segment geometric models within an unsupervised machine learning framework.By combining the graph kernel and the model segmentation technique, we established a substructure matching method which leads to novel, powerful, and representation-agnostic feature recognition algorithms that can hardly be built with existing shape descriptors.Practically all existing shape analysis and processing algorithms have been developed for specific geometric representations of 3D models. However, the product development process always involves a large number of often incompatible geometric representations tailored to specific computational tasks that take place during this process. Consequently, a substantial effort has been expended to develop robust geometric data translation and conversion algorithms, but the existing methods have well known limitations.The Maximal Disjoint Ball Decomposition (MDBD) was recently defined as a unique and stable geometric construction and used to define universal shape descriptors based on MDBD's associated contact graph. In this paper, we demonstrate that by applying graph analysis tools to MDBD in conjunction with graph convolutional neural networks and graph kernels, one can effectively develop methods to perform similarity, segmentation and solid feature recognition from geometric models regardless of their native geometric representation. We show that our representation-agnostic approach achieves comparable performance with state-of-the-art geometric processing methods on standard yet heterogeneous benchmark datasets while supporting all valid geometric representations.Download : Download high-res image (224KB)Download : Download full-size image

Multi-Scale Local Shape Analysis and Feature Selection in Machine Learning Applications

Local Deep Feature Learning Framework for 3D Shape.

Adaptive Multiscale Feature Extraction in a Distributed System for Semantic Classification of Airborne LiDAR Point Clouds

LIMAN: Local Information based Multi Attention Network for 3D Shape Recognition

Discriminative Localized Sparse Representations for Breast Cancer Screening

Point Set Multi-Level Aggregation Feature Extraction Based on Multi-Scale Max Pooling and LDA for Point Cloud Classification

Supervised Multi-scale Locality Sensitive Hashing

Automatic Shape Model Construction in Multiscale Space

Localizing Anomalies via Multiscale Score Matching Analysis

Multi-Scale Analysis Of Color And Texture For Salient Object Detection

Multi-View Scaling Support Vector Machines for Classification and Feature Selection.

Multi-Scale Neighborhood Feature Extraction and Aggregation for Point Cloud Segmentation

Robust Multi-scale Anatomical Landmark Detection in Incomplete 3D-CT Data

A Multiscale and Hierarchical Feature Extraction Method for Terrestrial Laser Scanning Point Cloud Classification

Machine learning-assisted multi-scale modeling

Local Model Feature Transformations

Graph-Based Shape Analysis for Heterogeneous Geometric Datasets: Similarity, Retrieval and Substructure Matching

Connected Shape-Size Pattern Spectra for Rotation and Scale-Invariant Classification of Gray-Scale Images

Local extreme learning machine: local classification model for shape feature extraction