Seminar Series by Sydney AI Meetup

Join the Sydney AI Meetup Seminar Series for expert talks on cutting-edge AI research, industry trends, and networking opportunities!

Time: 11:00 AM - 12:00 PM on 26 Feb 2026 (Sydney time)

Venue: Zoom: https://uni-sydney.zoom.us/j/83784288641?from=addon

Title: Reconstructing the Unseen: From Detail Recovery to Structure-Guided 3D Generation

Abstract: A central challenge in 3D vision is the reconstruction of what cannot be directly observed. Real-world observations are inherently incomplete: objects are partially visible, captured at limited resolution, or observed from restricted viewpoints. Traditional reconstruction methods are therefore constrained by available measurements, while recent generative approaches often produce visually plausible results without sufficient geometric grounding. This talk presents a research trajectory that addresses the problem of reconstructing the unseen by progressively integrating generative reasoning into 3D modeling.

The first part of the talk examines recovering unseen details within reconstruction, where generative priors are used to synthesize close-up geometry that is not clearly captured in the initial model, extending reconstruction beyond the limits of observation. The second part introduces a structural abstraction based on point clouds, representing only reliably observed geometry while learning to infer missing structure through generative completion. The final part advances this idea by embedding explicit 3D priors directly into the generative process, enabling the generation of complete 3D assets that preserve observed geometry while plausibly reconstructing unseen regions.

Together, these works illustrate an emerging paradigm shift in 3D vision: reconstructing the unseen is no longer treated purely as a reconstruction problem, but as a structured generation problem guided by geometric constraints. This perspective unifies reconstruction and generation within a single framework and suggests a new direction for controllable, reliable, and scalable 3D modeling.

Bio: Lingqiao Liu is an Associate Professor in the School of Computer Science at The University of Adelaide, Australia, and an Academic Member of the Australian Institute for Machine Learning (AIML). He was a recipient of the Australian Research Council (ARC) Discovery Early Career Researcher Award (DECRA) and the University of Adelaide Research Fellowship in 2016.

His research spans machine learning, computer vision, and natural language processing, with the overarching goal of building practical machine learning systems that are data-efficient, robust, and generalisable in real-world environments. His earlier work focused on learning under limited supervision, including semi-supervised, unsupervised, few-shot, and zero-shot learning, as well as improving the generalisation capability of machine learning systems through domain generalisation and compositional learning. More recently, his research has shifted toward generative models across multiple modalities, with a particular focus on data-efficient learning and generalisable adaptation of generative models for different tasks and application scenarios.