Candidate: Sheyang Tang
Date: March 11, 2026
Time: 2:00pm
Location: online
Supervisor: Dr. Zhou Wang
Abstract:
3D visual content is ubiquitous across various fields such as digital humans, product visualization, film and games, and AR/VR. Despite rapid advances in 3D acquisition, modeling, and rendering, the quality of 3D experience is ultimately determined by human perception. This talk presents a unified view of perception-aligned representation learning for 3D visual content—learning representations shaped by human-centric signals so that perceptual goals become measurable and useful for downstream tasks.
First, for coloured 3D mesh quality assessment, we show that perceived quality depends on the interaction between geometry and texture, where one can amplify or mask artifacts in the other. We introduce HybridMQA, which integrates topology-aware geometric learning with appearance cues from renderings to explicitly model geometry–texture interplay, improving evaluation accuracy and providing interpretable perceptual cues.
Second, for controllable 3D generation with implicit neural representations (INRs), we study generative modeling in parameter space and propose a layer-wise, hierarchical representation that aligns semantic control with network structure. By modeling cross-layer dependencies, the approach enables coarse-to-fine, interpretable control and improves generation quality across 3D content and additional modalities.
Third, for aesthetic viewpoint suggestion in 3D scenes, we introduce a 3D aesthetic field that distills 2D aesthetic knowledge into a feedforward 3D Gaussian Splatting representation, enabling aesthetic prediction at novel viewpoints from sparse inputs. An efficient search pipeline then identifies aesthetically appealing views without dense captures or reinforcement learning, yielding consistent improvements in framing and composition.