User Tools

Site Tools


10a.002.TAU_WP5 - Confidence Estimation in Neural Network for Illumination Estimation

Project - Summary

Confidence estimation in deep learning refers to the task of estimating a confidence/uncertainty score along with the model’s main prediction. Recently, there has been a lot of interest in solving this problem as it is crucial especially in applications where model failures carry serious effects, such as color constancy [1-2]. However, most of the interest has been directed toward classification and little work has been done a regression configuration [3-5]. In this project, the main goal is to focus on proposing novel approaches for confidence estimation in a regression setting and the target application is illumination estimation.

Project - Team

Team Member Role Email Phone Number Academic Site/IAB
Moncef Gabbouj PI 358 40 073 6613 Tampere University
Jenni Raitoharju Co-PI 358 50 447 8418 Tampere University
Alexandros Iosifidis Co-PI +4593508875 Tampere University
Firas Laakom Researcher 358 46 5219 0250 Tampere University
Jarno Nikkanen Project Mentor N/A Xiaomi

Project - Novelty of Approach

From an algorithmic perspective, in uncertainty estimation literature, more attention has been given to the classification task, while here we focus on the regression task. From an application perspective, previous uncertainty estimation techniques in illumination estimation focus on aggregating different Methods [6]. Here, we aim to develop approaches to estimate the uncertainty of the model directly. This can yield Novel and more robust illumination estimation techniques based on confidence estimation.

Project - Deliverables

1 Developing an approach for measuring similarity between a test sample and the training data

Project - Benefits to IAB

We believe that the proposed framework will be especially beneficial for Xiaomi since it can be used in the imaging models of their product. In addition, confidence estimation in deep learning is not restricted to the color constancy problem.

Project - Documents

projects/year10/10a.002.tau_wp5.txt · Last modified: 2022/10/03 15:03 by sally.johnson