Deep Representation Learning for 3D Object Recognition in Open-Ended Domains

Iftime, Vlad Cosmin (2020) Deep Representation Learning for 3D Object Recognition in Open-Ended Domains. Bachelor's Thesis, Artificial Intelligence.

Preview	Text AI_BA_2020_VladCosminIftime.pdf Download (1MB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (95kB)

Abstract

Open-ended category learning refers to the act of learning new object categories, after the training phase, without forgetting the previously learned categories. Due to the limited number of training data and also the difficulty in reprogramming the entire architecture of the system when a new category is presented, open-ended category learning is a solution that still needs improvement in service robotics. To that end, this project analyzes the effects of four different autoencoder neural networks on the representation learning part of the OrtographicNet, a network designed by H. Kasaei to tackle 3D object recognition in open-ended domains. The four types of autoencoders analyzed in this project are a simple dense autoencoder, a convolutional autoencoder, a variational autoencoder, and an adversarial autoencoder. The autoencoders were first trained to provide a unique object feature representation using a self-supervised representation learning approach. The feature representation obtained using only the encoder part of the autoencoders was then used for both the recognition and learning processes. After exhaustive testing using different recognition algorithms parameters, similarity measures distances, and pooling functions, the best autoencoder model for learning and recognizing 3D objects in open-ended domains is the variational autoencoder, due to its prior Gaussian distribution that forces the latent space of the autoencoder to obtain disentangled features.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Mohades Kasaei, S.H.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	13 Aug 2020 08:16
Last Modified:	13 Aug 2020 08:16
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/23078

Actions (login required)

View Item