AbstractThis dissertation is about in-situ object recognition, meaning that specific objects (instances) can be learned from a few training examples that depict them within the place where such objects are commonly present or being used. Learning to recognize objects in-situ opposes to conventional approaches in deep learning of relying on large-scale class-level datasets of grouped instances, utilizing complex image acquisition setups or utilizing synthetic data.
We aim for a scalable, robust, and real-time system based on Convolutional Neural Networks (CNNs) that learn discriminative features from images depicting objects from an egocentric point of view. We are particularly interested in learning objects from a few examples taken directly by an agent or by a demonstrator, and where the CNN does not need a finetuning process for learning additional instances, motivated by the computational limitations in most autonomous platforms. We hope our approach will be helpful for robotic tasks such as object manipulation, human-robot interaction, semantic mapping, scene understanding, autonomous navigation, and contribute to FARSCOPE's vision on advancing the state-of-the-art of autonomy in robots and intelligent systems.
|Date of Award||28 Nov 2019|
|Supervisor||Dima Aldamen (Supervisor) & Walterio W Mayol-Cuevas (Supervisor)|
Deep In-Situ Learning for Object Recognition
Lagunes Fortiz, M. A. (Author). 28 Nov 2019
Student thesis: Doctoral Thesis › Doctor of Philosophy (PhD)