Semi-supervised Mesh Segmentation and Labeling asd Semi-supervised Mesh Segmentation and Labeling Jiajun Lv, Xinlei Chen, Jin Huang, Hujun Bao State Key Lab of CAD&CG, Zhejiang University
Motivation Recognition of Mesh Semantic Meanings Head Torso Upper arm asd Motivation Recognition of Mesh Semantic Meanings Head Torso Upper arm Lower arm Hand Upper leg Lower leg Foot
asd Motivation 3D Modeling
Related Work Geometric Structure Based Approaches asd Related Work Geometric Structure Based Approaches Drawbacks: No Suitable Geometric Feature Data-driven Supervised Approaches Drawbacks: Large Size of Training Dataset Unsupervised Co-Segmentation Approaches Drawbacks: Inferior to Supervised Method
Related Work Head Neck Torso Leg Tail Ear asd Related Work Head Neck Torso Leg Tail Ear Learning 3d mesh segmentation and labeling, KALOGERAKIS E., 2010 Unsupervised co-segmentation of a set of shapes via descriptor-space spectral clustering, SIDI O., 2011 Learning boundary edges for 3D-mesh segmentation, BENHABILES H., 2011 Joint shape segmentation with linear programming, HUANG Q., 2011
Semi-Supervised Learning asd Semi-Supervised Learning
Semi-Supervised Learning asd Semi-Supervised Learning
asd Method
Key Technical Points Per-mesh Conditional Random Fields Model asd Key Technical Points Per-mesh Conditional Random Fields Model Incorporation of unlabeled mesh information with an entropy term Learning Parameters with Virtual Evidence Boosting
Definition Unary Features Labeled Meshes Semi-supervised asd Definition Unary Features Labeled Meshes Semi-supervised Mesh Segmentation Pairwise Features Unlabeled Meshes
Per-mesh Conditional Random Fields Model asd Per-mesh Conditional Random Fields Model
Incorporation of unlabeled mesh information with an entropy term asd Incorporation of unlabeled mesh information with an entropy term Information Gain: Negative Conditional Entropy of the CRF on Unlabeled Meshes Greater Certainty
Learning Parameters with Virtual Evidence Boosting asd Learning Parameters with Virtual Evidence Boosting
asd Learning Parameters with Virtual Evidence Boosting----Belief Propagation Information about distribution of sending node Information about which values recipient node should prefer
Learning Parameters with Virtual Evidence Boosting----LogitBoost asd Learning Parameters with Virtual Evidence Boosting----LogitBoost 1. Unary Energy Term 2. Pairwise Energy Term
Results and Discussion asd Results and Discussion
asd Results The segmentation and labeling results of our semi-supervised mesh segmentation algorithm on the whole Princeton Segmentation Benchmark
asd Results Experimental results of the semi-supervised mesh segmentation method. For each kind of dataset, the left column three are the labeled training dataset, and the right column three are the segmented meshes.
asd Results Comparison of Supervised and Semi-supervised Approaches
asd Results Semi-supervised Approach with Different Labeled Training Meshes
asd Robustness Noise of the labeled set is inevitable, such as human mislabeling Entropy term acts as a regularizer, avoiding over- fitting to training data
asd Robustness Comparison of Supervised and Semi-supervised Approaches on Inconsistent Labeled Data
Complexity of The Method asd Complexity of The Method Training Complexity Each Belief Propagation: Each Boosting Iteration: Labeling Complexity All Belief Propagation: All Boosting Iteration: Time Consumption Training: 7-12hours Labeling: a few minutes
Limitation Manually tuning parameters Jagged and Disconnect Patches asd Limitation Manually tuning parameters Jagged and Disconnect Patches Mesh with weak features
Future Work Hierarchical Models asd Future Work Hierarchical Models Partially Labeled Semi-supervised Mesh Segmentation
Thank You For Your Attention ! asd Thank You For Your Attention !