Files in this item

FilesDescriptionFormat

application/pdf

application/pdfECE499-Fa2018-jagarlamudi.pdf (3MB)Restricted to U of Illinois
(no description provided)PDF

Description

Title:Structurally consistent, diverse, colorization using generative modeling and GTAV
Author(s):Jagarlamudi, Shreya
Contributor(s):Schwing, Alexander G.
Subject(s):Video Colorization
Gaussian Conditional Random Field
VAE
MDN
Abstract:Colorizing videos is an important task in the media industry and in object tracking due to the nature of temporal coherency that comes with color. Doing this automatically, consistently and diversely has been a challenging problem in computer vision. Existing work in this field only deals with part of the problem to achieve temporally incoherent results by stitching colorized images together. In this thesis, we use a conditional random field based, variational auto encoder to model structural consistency and diversity of videos and train it on our custom dataset generated from GTA V. Instead of predicting colors for each pixel individually, we learn both spatial correlations between pixels in a frame and temporal correlations between pixels across different frames of each video in order to keep colorization consistent across the video. The model also includes a mixture density network to create a distribution of diverse colorizations per gray-scale video input. We also allow optional user controllability for more realistic colorizations. We demonstrate that our model achieves spatially and temporally consistent, diverse colorizations of gray-level videos.
Issue Date:2018-12
Genre:Other
Type:Text
Language:English
URI:http://hdl.handle.net/2142/105407
Date Available in IDEALS:2019-09-04


This item appears in the following Collection(s)

Item Statistics