Multiple Static Segmentation of Videos using a Convolution of Mixtures of Gaussian processes
Authors
Jensen, Jacob Jon ; Samuelsen, Niels Nørgaard ; Nielsen, Christoffer Samuel
Term
4. term
Education
Publication year
2015
Submitted on
2015-06-08
Pages
72
Abstract
I dette projekt undersøger vi multiple statiske segmenteringer af videoer—at lave flere supplerende opdelinger af pixels i hvert billede, så forskellige strukturer fremhæves. Vi bygger på en nyligt foreslået metode til multiple segmentering af billedstakke og videreudvikler den. Vi erstatter den oprindelige ikke-temporale probabilistiske model med en model baseret på gaussiske processer (Gaussian processes), en statistisk ramme der kan modellere glatte ændringer over tid. Inden for denne ramme afprøver vi forskellige middelværdi-funktioner (mean functions) for den gaussiske proces og viser, at en stykkevis lineær middelværdi kan give gode segmenteringer. Vores forsøg viser, at metoden kan finde flere meningsfulde strukturer i videoer, herunder segmenteringer vi ikke kunne opnå med den tidligere metode. Den er dog ikke bedre i alle tilfælde. For at øge effektiviteten anvender vi hovedkomponentanalyse (PCA) til at reducere dataenes dimensionalitet og opnår gode resultater med væsentligt kortere køretider.
This project studies multiple static segmentation of videos—creating several complementary ways to partition the pixels in each frame so that different structures are highlighted. We build on a recently proposed method for multiple segmentation of image stacks and develop improvements. We replace the original non-temporal probabilistic model with one based on Gaussian processes, a statistical framework that can model smooth changes over time. Within this framework, we test different mean functions for the Gaussian process and show that using a piecewise linear mean function can produce good segmentations. Our experiments indicate that the method can uncover several meaningful structures in videos, including segmentations we could not obtain with our previous approach. However, it does not yield better results in every case. To improve efficiency, we apply principal component analysis (PCA) to reduce the dimensionality of the data and achieve good results with much shorter running times.
[This abstract was generated with the help of AI]
Documents