Miculita, Andrei-Lucian (2023) Learning Optimal View Selection for Multi-View Object Representation. Master's Thesis / Essay, Artificial Intelligence.
|
Text
mAI_2023_MiculitaAL.pdf Download (2MB) | Preview |
|
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
toestemming.pdf Restricted to Registered users only Download (137kB) |
Abstract
Object recognition, pose estimation, and grasp affordance tasks for robots rely heavily on an accurate understanding of an object’s geometry. However, this can often be challenging due to the absence of a reliable object representation. In this thesis, we propose a novel approach for learning a descriptive object representation by using a good view selection policy. Our method, called Maximum Entropy Viewpoint Selection (MEVS), selects the most informative view of an object to classify it. We present two alternative approaches for MEVS: (i) a differentiable renderingbased approach, optimizing view entropy, and (ii) a point cloud embedding-based approach, using PointNet++ for predicting depth entropy. MEVS is adaptable to new situations and environments, making it particularly valuable for robots deployed in dynamic settings. To assess our approach, we conduct evaluations on two adaptations of the ModelNet10 dataset, in a multi-view pipeline based on a ResNet backbone. We find that the point cloud embedding-based method is suitable for real-time applications and can consistently outperform random view selection in finding the most informative views of objects it has not seen before.
| Item Type: | Thesis (Master's Thesis / Essay) |
|---|---|
| Supervisor name: | Mohades Kasaei, S.H. and Carloni, R. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Master's Thesis / Essay |
| Language: | English |
| Date Deposited: | 18 Sep 2023 08:49 |
| Last Modified: | 02 Nov 2023 09:37 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/31411 |
Actions (login required)
 |
View Item |