While multimodal AI excels at bridging text and 2D vision, aligning language with 3D geometry remains a fundamental open challenge. Can the structural gap between irregular 3D point clouds and sequential language embeddings be bridged without paired supervision? This thesis investigates the limits of cross-modal 3D-text alignment through the lens of Optimal Transport and geometric representation learning. Contrasting supervised linear and non-linear mappings against unsupervised Gromov-Wasserstein formulations, we reveal why relational optimal transport fails across modalities: 3D geometric manifolds and text latent spaces fundamentally break structural isometry. Our findings expose the mathematical boundaries of unsupervised cross-modal learning, demonstrating that explicit anchor supervision is not merely helpful, but indispensable to unite language and 3D space.
Mentre l'intelligenza artificiale multimodale connette con successo testo e immagini 2D, allineare il linguaggio naturale con la geometria tridimensionale rimane una sfida ancora aperta. È possibile colmare la profonda distanza strutturale tra nuvole di punti 3D e rappresentazioni linguistiche senza dati appaiati di supervisione? Questa tesi esplora i limiti dell'allineamento cross-modale 3D-testo attraverso il Trasporto Ottimo e l'apprendimento geometrico delle rappresentazioni. Mettendo a confronto proiezioni supervisionate con formulazioni non supervisionate basate su Gromov-Wasserstein, il lavoro svela le ragioni del fallimento dell'allineamento puramente relazionale: gli spazi geometrici 3D e gli spazi latenti testuali violano l'assunto di isometria strutturale. I risultati tracciano un chiaro confine matematico nell'apprendimento cross-modale, dimostrando come una supervisione esplicita tramite ancore sia indispensabile per unire linguaggio e spazio tridimensionale.
Aligning 3D and Text Latent Spaces: An Empirical Study of Alignment Methods
HYSA, ENDI
2025/2026
Abstract
While multimodal AI excels at bridging text and 2D vision, aligning language with 3D geometry remains a fundamental open challenge. Can the structural gap between irregular 3D point clouds and sequential language embeddings be bridged without paired supervision? This thesis investigates the limits of cross-modal 3D-text alignment through the lens of Optimal Transport and geometric representation learning. Contrasting supervised linear and non-linear mappings against unsupervised Gromov-Wasserstein formulations, we reveal why relational optimal transport fails across modalities: 3D geometric manifolds and text latent spaces fundamentally break structural isometry. Our findings expose the mathematical boundaries of unsupervised cross-modal learning, demonstrating that explicit anchor supervision is not merely helpful, but indispensable to unite language and 3D space.| File | Dimensione | Formato | |
|---|---|---|---|
|
Hysa_Endi.pdf
accesso aperto
Dimensione
2.39 MB
Formato
Adobe PDF
|
2.39 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/116549