Title

Uncertainty-Aware Disentangled Representations

Abstract

Abstract

In many computer vision tasks, not every part of an object of interest is always visible because of challenges like occlusion, viewpoint and pose variation. One approach to these kinds of challenges is working with part-based representations. In this thesis, we tackle the problem of obtaining part-based representations using disentanglement while estimating the uncertainty of each part to assess its availability. Parts are disentangled using a part-related supervised task and by using an adversarial loss, unrelated information is removed. Uncertainty of parts are estimated using loss attenuation over the same part-related task. We try several methods to integrate uncertainty values into both the training procedure and the decision making process during test time, making the model more robust to unavailable parts. The experiments are conducted over a toy dataset and the person re-identification task (namely, the Market-1501 dataset) which can benefit from a part-based representation.

Zoom Link:
https://zoom.us/j/98389755595?pwd=UVM4eDVrSzR6MFZkL2ZYWUFhaDh3dz09

Supervisor(s)

Supervisor(s)

SEZAI ARTUN OZYEGIN

Date and Location

Date and Location

2021-09-09 10:00:00

Category

Category

MSc_Thesis