Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Computing | en_US |
| dc.contributor.advisor | Yang, Bo (COMP) | en_US |
| dc.creator | Li, Jinxi | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/14403 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Learning 3D physics from multi-view videos | en_US |
| dcterms.abstract | Modeling dynamic 3D scenes with physical accuracy and predicting their future states is critical for applications in augmented reality, robotics, and content creation. While recent advances in implicit neural representations have achieved remarkable results in reconstructing dynamic geometry and appearance, they primarily excel at interpolating within observed time and fail to extrapolate future frames due to a lack of explicit physical modeling. Early physics-integrated methods, such as those using Physics-Informed Neural Networks (PINNs) and numerical simulations, face limitations including poor generalizability, heavy computational overhead, and training instability, hindering their practical deployment. | en_US |
| dcterms.abstract | This thesis addresses three core challenges in dynamic 3D scene modeling: ensuring physical plausibility in future predictions, improving efficiency in physics-aware modeling, and enabling explicit learning of underlying physical parameters. To tackle these challenges, we present three interconnected frameworks that progressively advance the state-of-the-art: In Chapter 3, we introduce NVFi, the first framework fusing dynamic radiance fields with mass and momentum conservation laws by PINNs loss functions. NVFi learns disentangled, physically consistent velocity fields from multi-view videos, eliminating unrealistic motion artifacts without relying on additional annotations. In Chapter 4, we propose FreeGave to address PINNs' inefficiencies. It incorporates a divergence-free module to enforce local physical constraints without global PDE sampling and uses learnable physics encoding to decompose complex motions, achieving efficient and stable physics-aware modeling for complex scenes. In Chapter 5, we develop TRACE, which shifts from implicit neural fitting to explicit physical parameter learning. By modeling each 3D Gaussian as a rigid particle with independent translation-rotation dynamics systems, TRACE directly learns core physical parameters and enables controllable future extrapolation while eliminating neural ODE integration overhead. | en_US |
| dcterms.abstract | Collectively, these contributions advance dynamic 3D scene modeling from physics-unaware interpolation to accurate, efficient, and controllable future frame extrapolation. By progressively integrating physical constraints, optimizing computational efficiency, and enabling explicit parameter learning, this thesis establishes a robust framework for machines to perceive and predict dynamic 3D world dynamics, laying the foundation for next-generation intelligent systems interacting with physical environments. | en_US |
| dcterms.extent | xvi, 120 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2026 | en_US |
| dcterms.educationalLevel | Ph.D. | en_US |
| dcterms.educationalLevel | All Doctorate | en_US |
| dcterms.LCSH | Three-dimensional modeling | en_US |
| dcterms.LCSH | Computer vision | en_US |
| dcterms.LCSH | Machine learning | en_US |
| dcterms.LCSH | Hong Kong Polytechnic University -- Dissertations | en_US |
| dcterms.accessRights | open access | en_US |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14403