Segmentation and Classification

1
HMM Model Selection Issues For Soccer Video
Mark Baillie, Joemon M. Jose and Keith van
Rijsbergen Department of Computing
Science University of Glasgow bailliem, jj
_at_dcs.gla.ac.uk http//ir.dcs.gla.ac.uk
Overview This poster describes an investigation
into the effect of HMM parameter selection on
system performance, for broad general audio
classes found in Soccer video.
Motivation There has been a concerted effort from
the Video Retrieval community to provide tools
that automate the annotation process of Sports
video programmes. A popular indexing framework is
the Hidden Markov Model (HMM). HMM is however
largely applied in an ad hoc manner. The main
thrust of our research is to provide an in-depth
investigation into HMM parameter selection and
what effect, if any, poor selection can have on
indexing performance, specifically when modelling
audio information. Such optimally selected models
are then combined into a unified HMM framework
for segmentation and classification of Soccer
video.

• HMM Selection Strategies
• The main goal of model selection is to choose the
  simplest possible model without a deterioration
  in accuracy. This is important given the
  difficulty and practicality of generating large
  and varied training sets. For example, complex
  models require large training sets, while simpler
  models will not encapsulate content.
• We investigate three model selection strategies
  and what effect each has on classification
  performance.
• The three selection strategies are
• Exhaustive Search
• Akaike Information Criterion (AIC)
• Bayesian Information Criterion (BIC)
• For all three strategies, the predictive
  likelihood score is used as a guide for HMM
  selection.

As the parameter increases so does the likelihood
score. For Exhaustive Search, a threshold is
required to identify the optimal model Both the
AIC and BIC selection criterion penalise the
predictive likelihood score, creating a maximum
peak. We assume this maxima to be the optimal
HMM.
As the number of hidden Markov states increase so
does the classification accuracy, levelling off
once a specific number of states is added. A good
model selection strategy will pick a HMM at this
point. A simpler Gaussian Mixture Model (GMM) is
displayed as a baseline.
Model Selection Results

• We compared selection strategies across a large
  test collection, using three broad content
  classes found in Soccer video.
• A simple GMM classifier was used as a baseline
• BIC selected the simplest HMMs
• No significant improvement across strategies
• No significant improvement over baseline GMM

• Segmentation and Classification
• Optimally selected HMMs using BIC were
  integrated into a Superior HMM framework
• A Soccer video topology was generated utilising
  domain knowledge
• Each state in the Superior HMM topology
  represented specific content found in a Soccer
  video file
• Able to incorporate the temporal flow of the
  video file into the segmentation process
• Segment and classify an entire video file into
  semantic units in a single pass

• Conclusions
• We highlighted the importance of model selection
  from experimentation
• Selecting too few or too many hidden states can
  produce poor classification accuracy
• Selection of complex models will resulting in
  training and over-fitting problems
• We found BIC to select the simplest optimal HMMs
  without effecting accuracy
• Optimally selected HMMs can then be integrated
  into a segmentation and classification system for
  Soccer video