Geometric invariance of PointNet
Le, Hoanh (2021)
Le, Hoanh
2021
Bachelor's Programme in Science and Engineering
Tekniikan ja luonnontieteiden tiedekunta - Faculty of Engineering and Natural Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2021-03-23
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202106075707
https://urn.fi/URN:NBN:fi:tuni-202106075707
Tiivistelmä
PointNet has become one of the de facto deep learning architectures for 3D tasks, from object classification to scene segmentation. One of its main components is two Joint Alignment Networks, which were designed to help PointNet to be invariant to geometric transformation such as rigid transformation. They attempt to canonicalize the input set and feature space before feeding them to the main network. However, their effects have not been studied extensively. In this work, we will evaluate PointNet’s performance in the presence or absence of Joint Alignment Networks under rotation transformation. We show that with a limited number of data, the use of Joint Alignment Networks does not increase Pointnet’s robustness against rotation transformation but can actually decrease it.
Kokoelmat
- Kandidaatintutkielmat [8714]