Javascript must be enabled for the correct page display

Classifying objects from unseen viewpoints using novel view synthesis data augmentation

Eker, Thijs (2021) Classifying objects from unseen viewpoints using novel view synthesis data augmentation. Master's Thesis / Essay, Artificial Intelligence.


Download (2MB) | Preview
[img] Text
Restricted to Registered users only

Download (125kB)


Deep learning-based computer vision tools have been applied in a wide variety of situations with great success. However, the application of deep learning is limited by the availability of large annotated datasets. Many applications are highly specialized and often large datasets in the specialized domain are very costly to obtain. We propose a data augmentation technique in which novel view synthesis models are used to broaden certain characteristics of the training dataset. In our experiments, we focused on widening the viewpoint distribution of training dataset to include images taken from viewpoints with a higher elevation. Our results show that using the data augmentation method, we can drastically improve the recognition accuracy of an off-the-shelf model on synthesized datasets. Furthermore, we propose a symmetry-based loss to enhance the symmetry of the objects in the images generated by neural radiance field-based (NeRF) GANs and novel view synthesis models. We implemented the symmetry-based loss for different models and obtained mixed results.

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: 19 Oct 2021 13:22
Last Modified: 04 Nov 2021 12:01

Actions (login required)

View Item View Item