Primary supervisor

Fucai Ke

Co-supervisors

This project addresses the challenge of multi-view visual reasoning, where a model must integrate sparse observations from multiple viewpoints to build a coherent understanding of the underlying 3D world. Unlike standard single-image reasoning, multi-view reasoning requires the model to reason over view consistency, spatial layout, object relations, and hidden structure across observations.

Building on our recent work on VIEW2SPACE, this project aims to develop fast and effective modeling strategies for spatial reasoning in multi-view settings. The goal is to design a system that can represent the world in a structured and compact way, support robust inference across viewpoints, and enable downstream visual agentic reasoning.

 

Aim/outline

  1. Review recent literature on multi-view reasoning, spatial reasoning, embodied reasoning, and visual world modeling.
  2. Learn how to run and evaluate existing methods on relevant benchmarks, including MINDCUBE, VIEW2SPACE, and MMSI.
  3. Analyze the key failure modes of current VLMs in multi-view spatial understanding, such as weak view aggregation, poor cross-view consistency, and limited world-state representation.
  4. Design a spatial understanding model that can effectively represent the world and reason over multi-view evidence.
  5. Train and adapt VLM-based models for spatial reasoning tasks, with a focus on robust visual agentic behavior.
  6. Evaluate the resulting system on benchmark tasks and analyze generalization, efficiency, and reasoning quality.

URLs/references

VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations, ECCV2026 — the main reference paper for the project.

https://arxiv.org/abs/2603.16506

For any inquiries regarding this project, please contact Fucai Ke via fucai.ke@monash.edu

Required knowledge

  • Python programming and PyTorch
  • Vision-Language Models and multimodal learning
  • Spatial reasoning or 3D/4D visual understanding
  • Basic familiarity with benchmark evaluation and model training
  • Strong analytical skills for experimentation and ablation studies