3D Shape Analysis Based Supervised Learning

The rapid development of digital geometry processing techniques such as 3D scanning and virtual reality has been greatly boosting both quantity and quality of 3D digital geometry.Processing,analyzing and semantically understanding of 3D shape,and highlevel understanding of 3D scene utilizing these obtained sematic information has been widely recognized by the graphics community.Machine learning has been successfully studied and applied in many fields of computer science community.Supervised machine learning techniques,such as convolutional neural network,extreme learning machine and support vector machine,become more and more popular in our research and industry.Inspired by all these insights,this thesis mainly focus on using these newly developed supervised machine learning methodology to study difficult issues in 3D shape analysis and understanding.We finish three works on 3D shape descriptor,3D shape segmentation and 3D shape feature learning in classification tasks as followings:1.AB3D: action-based 3D descriptor for shape analysis.High-level geometry processing has been a hot topic in graphics community.The functionality analysis of 3D models is an essential issue in this area.Existing 3D models often exhibit both large intraclass and inter-class variations in shape geometry and topology,making the consistent analysis of functionality challenging.Traditional 3D shape analysis methods which rely on geometric shape descriptors can not obtain satisfying results on these 3D models.We develop a new 3D shape descriptor based on the interactions between 3D models and virtual human actions,which is called Actions Based 3D Descriptor(AB3D).Due to the implied semantic meanings of virtual human actions,we obtain encouraging results on consistent segmentation based on AB3 D.Finally,we present a method for recognition and reconstruction of scanned 3D indoor scenes using our AB3 D.Experiments show that AB3 D is a promising shape descriptor towards functionality analysis of 3D shapes.2.3D shape segmentation and labeling via extreme learning machine.We propose a fast method for 3D shape segmentation and labeling via Extreme Learning Machine(ELM).Given a set of example shapes with labeled segmentation,we train an ELM classifier and use it to produce initial segmentation for test shapes.Based on the initial segmentation,we compute the final smooth segmentation through a graph-cut labeling constrained by the super-face boundaries obtained by over-segmentation and the active contours computed based on ELM segmentation.Experimental results show that our method achieves comparable results against the state-of-the-arts,but reduces the training time by two orders of magnitude,both for face-level and super-face-level,making our method scale well for large datasets.Based on such notable improvement,we demonstrate the application of our method for fast online sequential learning for 3D shape segmentation at face level,as well as realtime sequential learning at super-face level.3.Feature learning for 3D shapes from depth images.High level 3D shape features are crucial to shape analysis.However,hand-crafted shape features are labor-intensively designed and are unable to extract discriminative information from the data.Inspired from biological studies that 3D shape perception in human visual cortex is from combination of multiple depth cues,we propose a deep learning method,called the Multi-view Deep Extreme Learning Machine(MVD-ELM),for automatically learning shape features.Given a set of 2.5D depth images generated from 3D shapes on different views,we obtain multiple layers of 3D feature representations through the MVD-ELM.The visualization network gives us insights into the function of intermediate 3D feature layers and the operation of 3D shape classifier.Due to the rapid,abstract,rotation invariance properties of MVD-ELM,our networks can learn quite good 3D features which can be easily be used to obtain state-of-the-art 3D shape classification results,3D mesh segmentation,and object detection in RGBD images.Moreover,our method runs much faster than existing deep learning methods by approximately two orders of magnitude,which makes 3D features extracted by MVD-ELM more practically useful than other features.4.3D Feature Learning via Convolutional Auto-Encoder Extreme Learning Machine.3D shape features play a crucial role in graphics applications like 3D shape matching,recognition,and retrieval.Traditional 3D descriptors are hand-crafted features which are labor-intensively designed and are unable to extract discriminative information from existing large-scale 3D data.Convolutional neuron networks and auto-encoders are two most popular neuron networks in the field of deep learning.Based on the framework of extreme learning machines,we propose a rapid 3D feature learning method – convolutional extreme learning machine auto-encoder,which could automatically learn shape features from 3D shape datasets.Our method runs faster than existing deep learning methods by approximately two orders of magnitude.Experiments show that our method is superior to these existing hand-crafted features and other deep learning methods in 3D shape classification and 3D object detection.