Pairwise shape studies in 3D deep learning
Recently, deep learning has achieved impressive success on modeling and understanding 3D shapes. It becomes a fundamental research question how the learning based methods are generalizable to a collection of shapes in various geometry.
This talk will discuss two scenarios where studying the interpolation between a pair of shapes helps to improve and understanding the generalization of 3D deep learning models. We show that interpolation on the raw geometry of two-point clouds helps to improve the performance of point cloud classification (Part 1), while hidden feature-level interpolation helps to understand how the latent-conditioned implicit neural representations generalize to representing different 3D shapes (Part 2). This talk is based on the following papers by the speaker:
Lecture slides: https://drive.google.com/file/d/1DIR4...
00:00 Intro
05:23 PointMixup: Augmentation for Point Cloud
17:04 Interpolation with one-to-one correspondence
20:26 Point Cloud Interpolations
21:09 Experiments and Discussion
22:46 3D Point Cloud Classification
25:29 Compare with baseline interpolations
26:58 Different Networks / Datasets
29:38 Conclusion
32:19 Neural Feature Matching in Implicit 3D Representations
40:59 Implicit function for 3D surface reconstructiouchen
42:27 Smooth interpolation
56:14 Hierarchy in layers
58:46 Discussion - Smooth interpolation
01:05:07 Different Networks / Datasets
01:07:39 [Yang al, FoldingNet] Point cloud interpolation: no meaningful point correspondence
01:13:58 Application: Mesh Deformation in existence with inconsistency in topology or semantic parts
01:17:33 Quantitative Results
01:19:56 Conclusion
[Chapters were auto-generated using our proprietary software - contact us if you are interested in access to the software]
This talk is based on the following papers by the speaker:
1. PointMixup: Augmentation for Point Clouds.
ECCV 2020
https://arxiv.org/abs/2008.06374
https://github.com/yunlu-chen/PointMixup
2. Neural Feature Matching in Implicit 3D Representations.
ICML 2021
http://proceedings.mlr.press/v139/che...
Presenter BIO:
Yunlu Chen is a PhD candidate at University of Amsterdam, advised by Dr Efstratios Gavves. His research focuses on 3D deep learning, including monocular depth estimation, RGB-D semantic segmentation, point cloud understanding and implicit neural representations.
His git: https://github.com/yunlu-chen
-------------------------
Find us at:
Newsletter for updates about more events ➜ http://eepurl.com/gJ1t-D
Sub-reddit for discussions ➜ / 2d3dai
Discord server for, well, discord ➜ / discord
Blog ➜ https://2d3d.ai
AI consultancy Abelians ➜ https://abelians.com/