1. The document discusses deep learning applications for 3D data including classification, segmentation, shape deformation, and remeshing.
2. PointNet and PointNet++ are introduced for 3D point cloud classification and segmentation using deep neural networks.
3. Non-Euclidean methods for processing manifolds include spectral and spatial convolutional neural networks on graphs and manifolds.
4. The RoButcher project aims to use 3D data and deep learning for autonomous robotic cutting of meat in abattoirs.
Ivan Sahumbaiev "Deep Learning approaches meet 3D data"
1.
2. Agenda
1. Introduction to 3D data
a) Data collection
b) Data representation
2. Deep Learning applications based on 3D data
a) Classification problem
b) Segmentation problem
3. Remeshing
4. Non-Euclidean methods for manifold processing
a) Spectral
b) Spatial
5. Libraries and datasets
6. RoButcher project
7. Q&A
2
7. 7
PointNet
[Charles R. Qi ] PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation
Properties
1. Permutation invariance
2. Rigid transformation invariance
3. No local context
8. 8
PointNet++
PointNet++, Qi et al., NeurIPS, 2018
Properties
1. Permutation invariance
2. Rigid transformation invariance
3. Hierarchical features learning
3. Acces to the local context
24. 24
Spacial patches
j
• Local system of coordinates Uij around vertex i
(geodesic polar)
• Spatial convolution with filter g
𝑥𝑖
′
∝
𝜃=1
Θ
𝑔 𝜃
𝑗∈𝑁(𝑥)
𝑤 𝜃 𝒖𝑖,𝑗 𝑥𝑗
Geometric deep learning on graphs and manifolds using mixture model CNNs
28. 28
Spatial convolutions
Shape correspondence quality obtained by different methods on the FAUST humans dataset
Geometric deep learning on graphs and manifolds using mixture model CNNs
29. ModelNet (2015)
(ModelNet10: 4899
models, 10 categories)
ModelNet40: 12311
models, 40 categories
PartNet (2019)
A Large-scale
Benchmark for Fine-
grained and
Hierrarchical Part-level
3D Object
Understanding
ABC (2018)
A Big CAD
Model
Dataset For
Geometric
Deep
Learning
MPI FAUST
300 human meshes
10 objects un 30
poses
~7k vertices
29
Datasets
30. 1. Open3d
2. Pytorch3d
3. Nvidia-Kaolin
4. Pytorch-Points
5. Pytorch-Geometric
Python C++
1. LibIgl (has python bindings)
2. CGAL
3. Point cloud library (PCL)
4. Geometry Central
5. Open3d
30
Libraries
Easy to install
Contains visualizer tools
Compute some descriptors, normals,
reconstructions
Algorithms
o Pytorch integration easier
o Need to compile
Contains visualizer tools
Compute descriptors, normals,
reconstructions
Algorithms
Speed
31.
32. 32
A Robust, Flexible and Scalable Cognitive Robotics
Platform
RoBUTCHER – a European funded Horizon 2020
innovation project that aims to replace the
conventional line production in abattoirs, with parallel
production in autonomous «meat factory cells».
In the RoBUTCHER concept the robotic system will
be able to understand and plan cutting trajectories
based on the carcass that is presented. To achieve this
it will use a combination of detailed computed
tomography (CT) data, real-time 3D imagery and
human-expert cutting data for neural network training
toward cutting trajectory planning.
https://robutcher.eu/
34. 34
PointNet++
Data collection and Segmentation
Data
collection
Point cloud
segmentaiton
Point cloud to
CAD morphing
Extraction
correspondences
35. 35
• Point clouds from cameras is not
manifold.
• Reconstructing the mesh is hard.
Point cloud morphing and Correspondence
Data
collection
Point cloud
segmentaiton
Point cloud to
CAD morphing
Extraction
correspondences
36. 36
Point cloud morphing and Correspondence
Data
collection
Point cloud
segmentaiton
Point cloud to
CAD morphing
Extraction
correspondences