SLIDE 1
Volumetric and Multi-View CNNs for Object Classification on 3D Data - - PowerPoint PPT Presentation
Volumetric and Multi-View CNNs for Object Classification on 3D Data - - PowerPoint PPT Presentation
Volumetric and Multi-View CNNs for Object Classification on 3D Data Charles R. Qi*, Hao Su*, Matthias Niener, Angela Dai, MengyuanYan, Leonidas J.Guibas Rich Applications of 3D Augmented Robot Reality Perception 3D Representations for
SLIDE 2
SLIDE 3
3D Representations for Generic Object Classification
3DShapeNets by Z. Wu et
- al. CVPR 15
VoxNet by D. Maturana et
- al. IEEE/RSJ 15
MVCNN by H. Su et al. ICCV 15 DeepPano by B. Shi et al. IEEE/SPL 15
Volumetric Multi-Views
SLIDE 4
Volumetric CNNs Revisited
3DShapeNets by Z. Wu et
- al. CVPR 15
Volumetric CNNs
SLIDE 5
Multi-View CNNs Revisited
MVCNN by H. Su et al. ICCV 15
Multi-View CNNs
SLIDE 6
Shape Classification Results Revisited
3DShapeNets Wu et al. MVCNN Su et al. 70 75 80 85 90 95
77.3% 90.1%
SLIDE 7
Shape Classification Results Revisited
3DShapeNets Wu et al. MVCNN Su et al. 70 75 80 85 90 95
77.3% 90.1%
Big gap between volumetric and multi-view based methods
Why?
SLIDE 8
Cause 1: Architecture and Engineering
LeNet, 1998 AlexNet, 2012
SLIDE 9
Cause 1: Architecture and Engineering
LeNet, 1998 AlexNet, 2012 3DShapeNets, 2015
SLIDE 10
Cause 2: Resolution
224x224 Images Multi-View CNNs
MVCNN Su et al.
SLIDE 11
Cause 2: Resolution
224x224 Images 30x30x30 Volumes Volumetric CNNs
3DShapeNets Wu et al.
Multi-View CNNs
MVCNN Su et al.
SLIDE 12
Diagnosis of Causes: Variable Control
- Same resolution, study architectures
- Same architecture, look into resolutions
SLIDE 13
Sphere Rendering
Occupancy Grid 30x30x30 Polygon Mesh Image 224x224
SLIDE 14
Sphere Rendering
Occupancy Grid 30x30x30 Polygon Mesh Image 224x224 Same “3D Resolution”
SLIDE 15
Investigation into Architecture
3D CNN Multi-View Image CNN Different Architecture
Sphere Rendering Images Occupancy Grid Volumes
Same 3D Resolution (30x30x30)
SLIDE 16
CNNs with Same 3D Resolution Inputs
72 74 76 78 80 82 84 86 88
MVCNN with Sphere Rendering Images 3DShapeNets Wu et al. Shape Classification Accuracy
SLIDE 17
Novel 3D CNN Architectures
3D NIN with Subvolume Supervision Push Harder for Learning Better!
SLIDE 18
Novel 3D CNN Architectures
Anisotropic Probing Network
SLIDE 19
Results of Our Novel 3D CNNs
72 74 76 78 80 82 84 86 88
MVCNN with Sphere Rendering Images 3DShapeNets Wu et al. Ours 3D CNN Shape Classification Accuracy
SLIDE 20
72 74 76 78 80 82 84 86 88
MVCNN with Sphere Rendering Images 3DShapeNets Wu et al. Ours 3D CNN
Results of Our Novel 3D CNNs
Closed the Gap under same 3D Resolution
Shape Classification Accuracy
SLIDE 21
Investigation into Resolution
Multi-View Image CNN 3D CNN Multi-View Image CNN
Standard Rendering Images Sphere Rendering Images 30x30x30 Volume
Same Architecture Different 3D Resolution
SLIDE 22
Performance Trend wrt 3D Resolution
82 84 86 88 90 92 94 50 100 150 200 250
Accuracy (%) 3D Resolution
MVCNN-Sphere
SLIDE 23
Performance Trend wrt 3D Resolution
82 84 86 88 90 92 94 50 100 150 200 250
Accuracy (%) 3D Resolution
MVCNN-Sphere Our 3D CNN
SLIDE 24
Generalization to Real Scans
Real Scan Dataset 243 objects 12 categories
Shape retrieval on scan data
SLIDE 25