Lê, Eric-Tuan Minh Claude; (2025) 3D parsing, estimation and learning using weak image supervision. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text 3D_parsing_estimation_and_learning_using_weak_image_supervision.pdf - Accepted Version Download (178MB) | Preview

Abstract

The goal of this thesis is to provide methods that will allow an ML-based 3D system to meet several requirements: it should be fast, offer high-level of control, be differentiable, and produce pixel-accurate reconstructions from images to 3D. We also consider a broad range of tasks, from analyzing existing 3D data to generating new 3D content with precise, parametric control while addressing the unique challenges of 3D analysis and synthesis, such as limited and costly training data, handling self-occlusions, modeling 3D deformations and enabling camera control in virtual scenes. To address these needs we leverage the complementary strengths of 3D representation to tackle various 3D tasks. We present contributions across a wide range of 3D representation methods. We begin with a pure point-based segmentation model, using point clouds as input to explore how 2D-inspired concepts can improve accuracy and efficiency in labeling and processing. Building on this, we focus on reconstructing 3D shapes by parsing point clouds into sets of primitive shapes, creating a more structured representation. Next, we introduce a hybrid model that bridges point clouds and mesh-based representations, enabling differentiable image rendering of point-based scenes. Shifting toward image-centric tasks, we propose a mesh-based model for reconstructing human body parametric meshes with strong 2D alignment from single RGB images. Building on the advantages of parametric approaches, we further extend our work by combining morphable models with a 3D synthesis pipeline. This approach allows us to learn a 3D template as an implicit representation for any object category, relying solely on 2D image supervision through backpropagation. Together, these contributions bring us one step closer to more efficient 3D machine learning frameworks that would harnesses the unique benefits of 3D data while achieving the ease of use and flexibility that have been successful in 2D models.

Download activity - last month

Export as CSVJSONXML

Download activity - last 12 months

Export as XMLJSONCSV

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item