REPORT 2Levi Smith
PROJECT STATUS
Reading papers Getting data set together
Clipping videos to form the training and testing data for our classifier
Project separation Christian will focus on action detection and
recognition My focus will be on shot type detection and
localization on the field
PAPERS
CRAM: Compact Representation of Actions in Movies Display concurrently the desired portions of the video Extracts actions of interest from 3D optical flow field Use action template to find similar actions within the given
video
Not good for group actions, such as those on the soccer field We do not want to display all of our events concurrently, but
some of the techniques could prove useful for action detection
PAPERS
An Effective Soccer Video Shot Detection Algorithm Only uses the frame color histogram to categorize shots Looks at amount of green pixels to verify if field is visible or
not
Would be advantageous to look at more features to detect and classify shots
PAPERS
Automatic Soccer Video Analysis and Summarization Shot boundary detection
Absolute difference between two frames in their ratios of dominant (grass) colored pixels to total number of pixels
Difference in color histogram similarity Shot classification
Utilize a Golden section composition rule, where they look at the amount of grass colored pixels in each region of the subdivided frame
Shot class can be determined from single keyframe or from a set of frames
PROJECT
Goal Extract a meaningful summary of the
sports video provided Method
Combine action recognition and shot detection/classification techniques
Assign probabilities to field locations for each localized action to assist in action classification
PROJECT OVERVIEW
Train classifier to detect shot boundary and classification
Localize the shot on the field Assign probabilities for each action to
locations on the field
SHOT DETECTION
Train a classifier, which will give us confidence levels Given a shot, classify it as one of a list of types
Panoramic, audience, zoomed in, corner, goal post, penalty box
Features CSIFT STIP HOG
Penalty Box Long shot
SHOT LOCALIZATION
Take a shot and localize it on the field by matching features
Field symmetry could present a challenge
ASSIGN PROBABILITIES
For each location on the field, assign a probability to each action to assist with classification
When given a new shot to classify, we will use this probability to increase our confidence in the action detection
Goal .8Foul .3
………
Goal .10Foul .5
………
Corner .8