Novel learning-based techniques and benchmarks for 3D part grouping
File(s)
Author(s)
Cheng, Junfeng
Type
Thesis
Abstract
This thesis proposes a novel task called “3D part grouping.” Consider a case where we have a mixed set of 3D parts originating from different 3D shapes, and the goal of 3D part grouping is to assign these parts to their corresponding groups. This scenario is quite common in industrial environments and home settings, such as reorganising mixed industrial parts or carelessly mixed IKEA furniture parts. In the archaeological field, this situation is also common, as relic parts or pieces are often mixed together. In such cases, it would be ideal to have an algorithm that can automatically reorganise these parts.
To address this challenging task, we propose two datasets, namely MixPartNet and ArcPie. The former focuses on daily-life scenarios, while the latter targets archaeological cases. We also propose two levels of evaluation metrics, called Part-Level (PL) evaluation and Group-Level (GL) evaluation, to comprehensively evaluate algorithms, thereby fully benchmarking the 3D part grouping problem. In addition, we propose two frameworks, Gradient-Field-based Auto-Regressive Sampling (G-FARS) and 3D Probabilistic Graph Search (3DPGS), to address the task. The two frameworks follow completely different mathematical formulations for 3D part grouping. G-FARS incorporates a gradient-field-based graph neural network that auto-regressively samples selection vectors to identify all groups, while 3DPGS leverages global relational features and explicitly predicts a relationship graph, which can be used to infer potential groups. Both algorithms successfully address this problem with State-of-the-Art performance, outperforming other baselines by large margins.
The proposed 3D part grouping task, along with the datasets, evaluation metrics, and algorithms, has strong potential for real-world applications, such as industrial or home-use robots and automatic archaeological relic reconstruction. In addition, 3D part grouping provides a valuable platform to study the multi-3D-object processing ability of 3D processing algorithms, which further contributes to the 3D vision community.
To address this challenging task, we propose two datasets, namely MixPartNet and ArcPie. The former focuses on daily-life scenarios, while the latter targets archaeological cases. We also propose two levels of evaluation metrics, called Part-Level (PL) evaluation and Group-Level (GL) evaluation, to comprehensively evaluate algorithms, thereby fully benchmarking the 3D part grouping problem. In addition, we propose two frameworks, Gradient-Field-based Auto-Regressive Sampling (G-FARS) and 3D Probabilistic Graph Search (3DPGS), to address the task. The two frameworks follow completely different mathematical formulations for 3D part grouping. G-FARS incorporates a gradient-field-based graph neural network that auto-regressively samples selection vectors to identify all groups, while 3DPGS leverages global relational features and explicitly predicts a relationship graph, which can be used to infer potential groups. Both algorithms successfully address this problem with State-of-the-Art performance, outperforming other baselines by large margins.
The proposed 3D part grouping task, along with the datasets, evaluation metrics, and algorithms, has strong potential for real-world applications, such as industrial or home-use robots and automatic archaeological relic reconstruction. In addition, 3D part grouping provides a valuable platform to study the multi-3D-object processing ability of 3D processing algorithms, which further contributes to the 3D vision community.
Version
Open Access
Date Issued
2026-02-10
Date Awarded
2026-04-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
Stathaki, Tania
Publisher Department
Department of Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)