Javascript must be enabled for the correct page display

Uncertainty Estimation for Object Detection

Beckersjürgen, Yannik (2019) Uncertainty Estimation for Object Detection. Master's Thesis / Essay, Artificial Intelligence.


Download (11MB) | Preview
[img] Text
Restricted to Registered users only

Download (119kB)


In this thesis, we investigate whether additional uncertainty estimates can be added to the output of a neural network that is used for object detection. We are mainly interested whether dropout, a stochastic regularization technique, can be utilized for this purpose. When applied to the decision referral task in which one wishes to identify and exclude the most difficult inputs from the dataset, dropout does not perform better than the standard network output. In addition, using dropout to estimate bounding-box regression uncertainty performed worse when compared to modeling the error with a Laplace distribution. However, we found that an adjusted form of dropout performed well for detecting a change in data distribution between daylight and night images. We further investigated whether regression uncertainty or an autoencoder reconstruction error could be used to identify novel object categories of interest, which was found not to be the case. Overall, we find that prior research on classification tasks does not hold for our object-detection context.

Item Type: Thesis (Master's Thesis / Essay)
Supervisor nameSupervisor E mail
Supervisor (outside RUG):
Supervisor outside RUG nameSupervisor outside RUG E mail
Hillerström, FiekeUNSPECIFIED
Degree programme: Artificial Intelligence
Thesis type: Master's Thesis / Essay
Language: English
Date Deposited: 19 Jun 2019
Last Modified: 19 Jun 2019 08:45

Actions (login required)

View Item View Item