eprintid: 10198472
rev_number: 9
eprint_status: archive
userid: 699
dir: disk0/10/19/84/72
datestamp: 2024-11-14 15:30:21
lastmod: 2024-11-14 15:30:21
status_changed: 2024-11-14 15:30:21
type: thesis
metadata_visibility: show
sword_depositor: 699
creators_name: Morreale, Luca
title: Neural Surface Representations
ispublished: unpub
divisions: UCL
divisions: B04
divisions: F48
note: Copyright © The Author 2024. Original content in this thesis is licensed under the terms of the Creative Commons Attribution 4.0 International (CC BY 4.0) Licence (https://creativecommons.org/licenses/by/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms.
abstract: Classic shape representations define 3D models as a set of discrete elements, generally as triangles or quads. This description is widely adopted across graphics pipelines, from rendering, compression, and shape editing, to shape correspondence.  Recently, neural shape representations have emerged as effective tools to describe complex geometries and structures. Models are encoded in network weights, that are then queried through a 3D or 2D point. However, most of these representations are task-specific and thus must be converted back to meshes to be used across applications. The need for a flexible representation, adaptable across the different neural pipelines, that lends itself to optimizations, remains.

This thesis explores the idea of neural surface representations as a map. First, we define surfaces as 2D-3D map encoded into neural network weights, dubbed Neural Surface Map. We adopt it as a building block to define a comprehensive mapping framework. Indeed, by composing multiple maps, we establish and optimize correspondences between shapes. This framework sidesteps the intricacies of optimization encountered by conventional methods while achieving continuous and bijective maps. Then, building on this foundation, we relax constraints within the mapping framework. In particular, we eliminate the need for human supervision by extracting (noisy) labels from pre-trained models. These labels are used to distil inter-surface maps between highly non-isometric shape pairs.

Finally, by interpreting a shape as the composition of a coarse structure and detail, we extend the neural representation to enable shape manipulation, compression, and detailed transfer. Intuitively, the structure defines the pose of the model and details repeating information, such as wrinkles. This representation lends itself to interactive shape manipulation, such as feature enhancement, while compressing detail into the network weights.
date: 2024-10-28
date_type: published
oa_status: green
full_text_type: other
thesis_class: doctoral_open
thesis_award: Ph.D
language: eng
primo: open
primo_central: open_green
verified: verified_manual
elements_id: 2327334
lyricists_name: Morreale, Luca
lyricists_id: LMORR34
actors_name: Morreale, Luca
actors_id: LMORR34
actors_role: owner
full_text_status: public
pages: 145
institution: UCL (University College London)
department: Computer Science
thesis_type: Doctoral
citation:        Morreale, Luca;      (2024)    Neural Surface Representations.                   Doctoral thesis  (Ph.D), UCL (University College London).     Green open access   
 
document_url: https://discovery.ucl.ac.uk/id/eprint/10198472/1/PhD_Thesis_final.pdf