Muralikrishnan, Sanjeev;
(2025)
Towards 3D Generative Models With Sparse Guidance.
Doctoral thesis (Ph.D), UCL (University College London).
Preview |
Text
Muralikrishnan_10209667_Thesis.pdf Download (102MB) | Preview |
Abstract
This thesis presents three works that introduce frameworks for generating static and dynamic 3D content from very sparse datasets. They were presented at CVPR, 3DV, and ECCV. These works primarily learn to generate human/animal shapes in varied poses and/or shapes. In this context, pose refers to the orientation of parts relative to others according to a hierarchical kinematic tree defined on body joints, while shape refers to geometric surface details – like curvature and normal orientation – that define the identity of the human or animal. Chapter 2 (GLASS) introduces a method to iteratively grow pose spaces - the space of relative part orientations - seeded with 3–10 poses to rich spaces of up to 2500 pose variations (identity-preserving deformations), by injecting classical geometric priors into the neural learning process. Chapter 3 (BLiSS) extends this idea by injecting learned neural priors to model shape variation - identity-changing deformations. Starting with 200 artist-curated body shapes, BLiSS grows the shape space – the space of all identities – to 1000 through iterative growth. Chapter 4 (Temporal Residual Jacobians) boosts learning of spatio-temporal motion fields from very few motion examples by learning temporal signals that preserve motion consistency; the learned field, when applied to characters-in-the-wild, imparts complex motion to these stylistic characters, creating newly animated shapes from sparse motion datasets. Each chapter is self-contained, with its associated literature review included. This thesis demonstrates that learning generative 3D models greatly benefits from task-specific priors, enabling such models to be trained from super-sparse guidance.
| Type: | Thesis (Doctoral) |
|---|---|
| Qualification: | Ph.D |
| Title: | Towards 3D Generative Models With Sparse Guidance |
| Open access status: | An open access version is available from UCL Discovery |
| Language: | English |
| Additional information: | Copyright © The Author 2025. Original content in this thesis is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) Licence (https://creativecommons.org/licenses/by-nc/4.0/). Any third-party copyright material present remains the property of its respective owner(s) and is licensed under its existing terms. Access may initially be restricted at the author’s request. |
| UCL classification: | UCL UCL > Provost and Vice Provost Offices > UCL BEAMS UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science > Dept of Computer Science |
| URI: | https://discovery.ucl.ac.uk/id/eprint/10209667 |
Archive Staff Only
 |
View Item |
