Spatial Reasoning in 3D Point Cloud Search from Geometry to Semantics and Beyond: A Survey

PhD Qualifying Examination


Title: "Spatial Reasoning in 3D Point Cloud Search from Geometry to Semantics
and Beyond: A Survey"

by

Mr. Jiawei LI


Abstract:

The ability to understand complex 3D scenes is a fundamental prerequisite for 
spatial intelligence in applications like robotics and augmented reality. A 
key measure of this understanding is the capacity for a system to perform 
semantic search in point clouds, locating specific objects based on natural 
language descriptions. This survey provides a systematic review of 3D point 
cloud search technologies, tracing their evolution to identify the primary 
challenges that remain in achieving robust spatial reasoning. We first 
propose a new taxonomy that organizes the field into two major branches: 
Topological Spatial Search, which operates on pure geometric constraints, and 
Semantic Spatial Search, which requires a deeper understanding of objects and 
their spatial relations. We then review the foundational techniques, from 
classic geometric primitives to modern semantic search based approaches such 
as 3D visual grounding techniques that enable object identification. Our core 
analysis reveals that while current state-of-the- art methods excel at 
grounding objects based on their intrinsic attributes, they exhibit 
systematic failures when faced with queries where spatial relationships are 
the primary distinguishing factor. We diagnose the root cause of this 
fragility as the entanglement of semantic processing and geometric validation 
within monolithic, implicit reasoning models. Based on the insight, we 
advocate for a paradigm shift towards decoupled reasoning frameworks, such as 
a retrieve-then-rerank pipeline, which separates semantic selection for 
objects from explicit reasoning on their spatial relationships. By providing 
this structured analysis and a clear path forward, this survey aims to guide 
future research toward building the next generation of 3D data management 
systems capable of truly understanding our real-world environment.


Date:                   Thursday, 26 September 2025

Time:                   2:00pm - 3:00pm

Venue:                  Room 4472
                        Lifts 25/26

Committee Members:      Prof. Xiaofang Zhou (Supervisor)
                        Dr. Dan Xu (Chairperson)
                        Dr. May Fung