Exploring low-labeled representation learning for learning and perception in realistic scenarios
Loading...
Author(s)
Lunayach, Mayank
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Most progress in computer vision has been on well-curated clean object-centric datasets. This thesis attempts challenging computer vision problems in realistic scenarios. For realism, the following criteria are followed in this work: 1) scene-centric and not object-centric data, and 2) limited or no labeled data. We pick one problem each in learning and perception. Specifically, in the chapter on learning, we: 1) propose Continual Wandering, a novel naturalistic online lifelong learning setting 2) introduce strong baselines suited for this challenging setting and 3) empirically evaluate the baselines and observe a trade-off between online performance and forgetting - methods having good online accuracy also had the highest forgetting. Then we study a case of realistic and low-labeled perception. Specifically, we deal with the problem of 6D pose, size, and estimation. To our knowledge, we are the first to propose an end-to-end method that does detection, localization, and 3D predictions in one forward pass in an in-the-wild setting without using any labeled 3D data. We get impressive qualitative results and comparable quantitative results when compared to baselines.
Sponsor
Date
2023-05-01
Extent
Resource Type
Text
Resource Subtype
Thesis