Leijenaar, Remco (2025) Self-Supervised 3D Representation Learning with Asymmetric Dual Self-Distillation for Point Clouds. Master's Thesis / Essay, Artificial Intelligence.
|
Text
mAI2025LeijenaarRF.pdf Download (12MB) | Preview |
|
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
Toestemming.pdf Restricted to Registered users only Download (159kB) |
Abstract
Recognizing tree species from 3D LiDAR scans remains a challenge in large-scale forest inventory systems, particularly due to the limited availability and diversity of annotated training data. This thesis explores how self-supervised representation learning (SSRL) can address this limitation by learning directly from raw, unlabelled 3D point clouds. We present AsymDSD, an Asymmetric Dual Self-Distillation framework that combines masked modeling and invariance learning through prediction in latent space. Key design elements include an asymmetric architecture, masked query isolation, multi-mask sampling, and a point cloud adaptation of multi-crop. AsymDSD is pretrained on over 52 billion LiDAR points covering 1300 hectares of forest and evaluated on a novel dataset of 12,000 trees spanning 29 genera in both forest and urban settings. Results show that SSRL halves the tree species classification error rate compared to training from scratch. Additionally, AsymDSD achieves state-of-the-art results on the ScanObjectNN benchmark (90.53%), improving to 93.72% when pretrained on 930k objects.
| 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: | 21 Jul 2025 14:28 |
| Last Modified: | 21 Jul 2025 14:28 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/36440 |
Actions (login required)
 |
View Item |