3D point cloud analytics
As collection of real world data is tedious and can sometimes be difficult due to places being inaccessible by scanners, the study of incorporating synthetic data together with real world data will be conducted to find a potential solution. The purpose of this report investigates the act of performi...
Saved in:
| 主要作者: | |
|---|---|
| 其他作者: | |
| 格式: | Final Year Project |
| 語言: | English |
| 出版: |
Nanyang Technological University
2022
|
| 主題: | |
| 在線閱讀: | https://hdl.handle.net/10356/162913 |
| 標簽: |
添加標簽
沒有標簽, 成為第一個標記此記錄!
|
| 機構: | Nanyang Technological University |
| 語言: | English |
| 總結: | As collection of real world data is tedious and can sometimes be difficult due to places being inaccessible by scanners, the study of incorporating synthetic data together with real world data will be conducted to find a potential solution. The purpose of this report investigates the act of performing semantic segmentation on 3D point clouds and explores the differences between using real world and synthetic dataset for training. SPVCNN based semantic segmentation model will be used in this study as a baseline and is selected due to its good performance as compared to other architectures. Self-training for unsupervised domain adaptation used commonly in 2D semantic segmentation will be studied and attempted on 3D points clouds to bridge the differences between the source and the target domain . Results from the baseline model, compared and analysed with the self-training model shows improvement in prediction accuracy but mostly only for one of the dominant classes. A deeper look into the model predictions explains the reasons for such improvement and proposes potential solutions for enhancing the model even further. |
|---|