Junkeun Yi

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2024-72

May 9, 2024

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-72.pdf

Unsupervised Object-Centric Learning is an effective means of learning reusable representations that can explain data in meaningful partitions, through creating a separate representation for each ‘object’ in the scene. In recent literature, Slot Attention has shown to be an effective unsupervised method for learning object representations, showing effectiveness in tasks such as object discovery, composition, and novel view synthesis. However, while showing remarkable results when operating on synthetic datasets such as CLEVR and MultishapeNet, it has yet to show promising results when applied to natural images. Meanwhile, NeRFs have shown that given many views, a neural field can be created that contains a dense and high resolution representation of a scene. In this thesis, we propose a new multi-scene NeRF method which leverages the ability of Slot Attention to discover object representations in the input data to train a model capable of representing many scenes. We demonstrate that when different scenes share common objects in the foreground area, the slot attention method can discover reusable representations of these objects to be shared in the volumetric rendering process.

Advisors: Trevor Darrell

---

BibTeX citation:

@mastersthesis{Yi:EECS-2024-72,
    Author= {Yi, Junkeun},
    Title= {Object Discovery In Multi-Scene NeRFs},
    School= {EECS Department, University of California, Berkeley},
    Year= {2024},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-72.html},
    Number= {UCB/EECS-2024-72},
    Abstract= {Unsupervised Object-Centric Learning is an effective means of learning reusable representations that can explain data in meaningful partitions, through creating a separate representation for each ‘object’ in the scene. In recent literature, Slot Attention has shown to be an effective unsupervised method for learning object representations, showing effectiveness in tasks such as object discovery, composition, and novel view synthesis. However, while showing remarkable results when operating on synthetic datasets such as CLEVR and MultishapeNet, it has yet to show promising results when applied to natural images. Meanwhile, NeRFs have shown that given many views, a neural field can be created that contains a dense and high resolution representation of a scene. In this thesis, we propose a new multi-scene NeRF method which leverages the ability of Slot Attention to discover object representations in the input data to train a model capable of representing many scenes. We demonstrate that when different scenes share common objects in the foreground area, the slot attention method can discover reusable representations of these objects to be shared in the volumetric rendering process.},
}

EndNote citation:

%0 Thesis
%A Yi, Junkeun 
%T Object Discovery In Multi-Scene NeRFs
%I EECS Department, University of California, Berkeley
%D 2024
%8 May 9
%@ UCB/EECS-2024-72
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2024/EECS-2024-72.html
%F Yi:EECS-2024-72