PointGrow: Autoregressively Learned Point Cloud Generation with - - PowerPoint PPT Presentation

▶

Dec 15, 2022 927 likes •1.23k views

PointGrow: Autoregressively Learned Point Cloud Generation with Self-Attention Anonymous Authors Key Ideas Generate realistic point cloud from scratch or conditioned on semantic contexts Recurrent sampling operation Augment with

SLIDE 1

PointGrow: Autoregressively Learned Point Cloud Generation with Self-Attention

Anonymous Authors

SLIDE 2

Key Ideas

Generate realistic point cloud from scratch or conditioned on semantic

contexts

Recurrent sampling operation
Augment with dedicated self-attention to capture long-range inter-point

dependencies

Learn a smooth manifold of image conditions

SLIDE 3

Recurrent Point Generation

Estimate conditional

distribution of point given all preceding points

Use discrete

softmax to decide next point

handle irregularity
f point cloud
Encode diverse

local structure

SLIDE 4

PointGrow

Assign a probability to each point cloud by factorization
Unconditional:
Conditional:

SLIDE 5

Context Awareness Operation

Fetching and averaging pooling

SLIDE 6

Capture Long-Range Dependencies

SLIDE 7

Self-Attention Fields

Distance between query point context feature to its accessible points

(inaccessible ones marked as infinity)

SLIDE 8

Magic Show

SLIDE 9

Learning Representations and Generative Models for 3D Point Clouds

Anonymous Authors

SLIDE 10

Key Ideas

This is the first deep generative model for point clouds
A new autoencoder + GAN architecture for point clouds
A compact representation with good reconstruction quality is learned
Point cloud metrics study

SLIDE 11

Network Configurations

AE
Encoder: 1-D convs + feature-wise maximum (symmetric permutation invariant)
Decoder: FCs
Loss: earth mover’s distance / Chamfer distance
AE Raw + GAN
AE Latent + GAN
AE Latent + GMM (works best with CD)

SLIDE 12

Representation Magic

Unseen shape reconstruction
Part editing: simple additive algebra
Interpolating shapes
Shape analogies
Shape completions
Shape classification
3D point cloud generation

SLIDE 13

Show, Attend and Translate: Unsupervised Image Translation with Self-Regularization and Attention

Anonymous Authors

SLIDE 14

Key Ideas

Learn mapping from input image in source domain to output image in

target domain

Pair training data is costly in this case (unsupervised needed)
Translated image is perceptually similar to original and appears to be

drawn from new domain

Attention module guides translation to focus on subject of interest

SLIDE 15

Method

Adversarial loss + Perceptual loss + Attention

SLIDE 16

Model

SLIDE 17

Magic Show

SLIDE 18

Geomstats: a Python Package for Riemannian Geometry in Machine Learning

Anonymous Authors

SLIDE 19

Key Ideas

A package specifically targeted to the machine learning community
It has numpy and tensorflow backend, GPU-compatibility
Keras version is also provided
Riemannian geometry education through a hands-on approach

SLIDE 20

Riemannian Manifold

Growing interest in using Riemannian geometry in machine learning
Input: can belong to or itself is Riemannian manifold (human pose)
Output: can belong to Riemannian manifold (predict camera pose)
Parameters: can be constrained on Riemannian manifold (Stiefel

manifold)

Low-dimensional manifold saves computations and memory

SLIDE 21

Use Cases

Hypersphere
Example: minimization of a scalar field on a sphere
Hyperbolic
Example: relevant to Gaussian space and hierarchical

representations

Symmetric Positive Definite Matrices
Example: connectivity graph, covariance, feature

constraints

Lie groups SO(n), SE(n)
Example: orientation and pose prediction (rigid

transformations), Riemannian geodesic distance