Classification of Large-Scale Shapes with Local Dissimilarities 24.04.2019 Classification of Large-Scale Shapes with Local Dissimilarities Xizhi Li University of Bremen, Germany cgvr.informatik.uni-bremen.de CGI’17, June 2017, Yokohama, Japan
Motivation 24.04.2019
3D model classification basics 24.04.2019 How to extract feature from this model ? 2D image ? Implicit function ?
Previous works 24.04.2019 LightField descriptors combine with Deep Learning [Bai & Zhou, 2016]: Spectral-based shape analysis – ShapeDNA [Reuter, 2009]: Histogram-based shape analysis [Rusu, 2008] The first
Our contribution Extended the work of Wahl [wahl, 2003] Add more shape descriptors to strength its discrimination Parallel version of algorithm to adapt to large-scale model classification Combine machine learning algorithm with histogram-based model classification Our algorithm is robust dealing with noise and incomplete models compared with contemporary algorithms
Our shape descriptors Definition: AHD descriptor 24.04.2019 Definition: AHD descriptor α= - ηis the Gaussian curvature β= - κis local normal perturbation γ= δ=
Our shape descriptors Parallel version Pseudo-Code: —————————— 24.04.2019 Parallel version Pseudo-Code: —————————— \definecolor{blau}{rgb}{0.15,0.15,1} ———————— \color{blau} % besser zu lesen als ub_blue d \leftarrow F(M) = F( x_0+1, y_0 + \tfrac{1}{2} ) = n_1 + \tfrac{n_2}{2} ————— \color{blau} t_1 = (\bar{A}_a \ominus O_a) \odot d'_a ————————— t_\mathnormal{l} = \min\{ t_i^{\mathnormal{l}}, 1\} A = \mathnormal{I} - \mathnormal{ \Lambda }
GPU VS CPU
Practice01:Asteroid classification 24.04.2019 Experiment data origin from 3D asteroid catalogue website. We selected 20 raw asteroids and utilizing possion disk sampling to extend the origin model into 1000 models. Add random noise on the surface of each model. (to simulate the real situation) The first is raw asteroid, the second one is sampling result. The third is noise asteroid.
Analysis PCA transform integrated all object‘s histogram into 3D space Choose the optimum machine learning algorithm to segment histogram space.
Results:Asteroid classification
Practice02:NTU database classification Experiment data origin from NTU database. We selected 1,218 3D models are composed into 10 classes. Some of the models are incomplete or integrated by several single parts.
Results:NTU database classification
Conclusions Local shape descriptors enable our hybrid-shape descriptors to classify models with local dissimilarities The parallel version of our algorithm can be adjusted to large-scale point clouds classification Random forest can be used to improve histogram-based signal classification Future challenges: Shape classification is a balance between computation and accuracy. Histogram-based algorithm cannot achieve very high accuracy but robust to incomplete while noise model classification. In the future, we can improve both hands. Incorporate with reinforcement learning
Thank you ! Q&A