Boer, Jakob, de (2026) Exploring the possibility, challenges and (dis-)advantages of deep ensemble methods in Point Cloud Completion. Master's Thesis / Essay, Artificial Intelligence.
|
Text
MasterThesisJakobdeBoer-1.pdf Download (31MB) | Preview |
Abstract
Reconstructing a complete 3D point cloud from an incomplete one is a critical challenge in numerous vision and robotics applications. Various point cloud completion methods have previously been introduced, each with its own strengths and weaknesses. In an attempt to make different methods complement one another, ensemble methods for Point Cloud Completion (PCC) tasks are proposed. Ensemble learning methods aim to enhance overall predictive performance beyond that achieved by any single constituent algorithm. In this thesis, six different deep-learning baseline PCC algorithms (PoinTr [1], PCN [2], GRNet [3], FoldingNet [4], SnowFlakeNet [5] and TopNet [6]) are combined using three novel ensembling criteria (Point Aggregation, Uniform Sampling and Performance-based Sampling) in both homo- and heterogeneous ensembles. Homogeneous ensemble base models are diversified with both random initialization and bagging. Parts of the large space of ensemble configurations are validated on the Completion3D dataset and a novel agricultural completion dataset. It is observed that ensembles using the point aggregation criterion improve chamfer distance beyond that of individual base learners up to 41.0%. Both random initialization and bagging successfully diversify base models, while randomly initialized base learners outperform the bagged alternative in both individual and ensembled contexts. Similar ensemble mechanisms are observed on the agricultural completion dataset.
| Item Type: | Thesis (Master's Thesis / Essay) |
|---|---|
| Supervisor name: | Mohades Kasaei, S.H. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Master's Thesis / Essay |
| Language: | English |
| Date Deposited: | 02 Apr 2026 11:01 |
| Last Modified: | 02 Apr 2026 11:01 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/37278 |
Actions (login required)
 |
View Item |