Akash Gokul

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2022-254

December 1, 2022

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-254.pdf

Instance discrimination pretraining has become an effective means of learning transferable visual representations. To date, this paradigm has focused on learning image-level representations. This is not only suboptimal for downstream tasks such as object detection, but can also lead to representations which do not capture the object(s) in the scene. In this thesis, we present two extensions of the instance discrimination paradigm to the object-level. First, we present a method which finds objects in a scene and enforces representational invariance at the object-level (Chapter 2). Next, we apply object-level knowledge to medical images by incorporating anatomical priors into the pretraining pipeline (Chapter 3). These methods provide improvements in downstream performance, efficiency, and interpretability when compared to state-of-the-art instance discrimination pretraining. We conclude with a broader analysis of object-level representation learning and instance discrimination pretraining (Chapter 4).

Advisors: Trevor Darrell

---

BibTeX citation:

@mastersthesis{Gokul:EECS-2022-254,
    Author= {Gokul, Akash},
    Title= {Object-Level Representation Learning for Natural and Medical Images},
    School= {EECS Department, University of California, Berkeley},
    Year= {2022},
    Month= {Dec},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-254.html},
    Number= {UCB/EECS-2022-254},
    Abstract= {Instance discrimination pretraining has become an effective means of learning transferable visual representations. To date, this paradigm has focused on learning image-level representations. This is not only suboptimal for downstream tasks such as object detection, but can also lead to representations which do not capture the object(s) in the scene. In this thesis, we present two extensions of the instance discrimination paradigm to the object-level. First, we present a method which finds objects in a scene and enforces representational invariance at the object-level (Chapter 2). Next, we apply object-level knowledge to medical images by incorporating anatomical priors into the pretraining pipeline (Chapter 3). These methods provide improvements in downstream performance, efficiency, and interpretability when compared to state-of-the-art instance discrimination pretraining. We conclude with a broader analysis of object-level representation learning and instance discrimination pretraining (Chapter 4).},
}

EndNote citation:

%0 Thesis
%A Gokul, Akash 
%T Object-Level Representation Learning for Natural and Medical Images
%I EECS Department, University of California, Berkeley
%D 2022
%8 December 1
%@ UCB/EECS-2022-254
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-254.html
%F Gokul:EECS-2022-254