Visualizing Video Processing Performance

1
Visualizing Video Processing Performance

• David Mihalcik
• CMSC 838B

2
Problem

• Have a variety of video processing algorithms
• Activity detection.
• Object tracking.
• Cut detection.
• Have a method of creating ground truth and a
  method of giving numerical results (ViPER).
• But this does not give a good overview of the
  results.

3
Solution

• Life-lines type interactive view.
• Provides a good way to compare details of
  multiple evaluations.
• Shows the time-series data, especially the
  nominal data, in a straightforward way.
• Pairing with a line graph or similar chart may be
  useful.
• Perhaps could make a parent bar that looks like a
  bar or h/l/c chart.

4
Alternatives

• ROC curves
• Tables!
• Line graphs for time series.
• Matrices for objects.

5
Object Detection
- alg1
Correct
Missed
False
Correct
Missed
False
6
Activity Detection
- alg1
Movement
Loading
Loitering
Correct
Missed
False
7
Goals

• Evaluation Goals
• Which algorithm is better? When?
• What are the weaknesses of some algorithm?
• General Video Summary Goals
• Finding interesting pieces of a video.
• Making searching faster than real-time.
• For both Getting some summary information e.g.
  standard ATS evaluation techniques.

8
Evaluation

• Data sets
• Using TREC data
• Mostly things like activities
• Person/Face tracking data
• Surveillance cameras
• About 15 minutes of video total
• Questions
• Where does algorithm A fail?
• When does A fail that B does not?
• Which is most likely to succeed in a given
  situation.

9
Limitations

• Probably wont make it look too good.
• Probably wont be too interactive.
• No dynamic queries
• Would like to be able to modify parameters on the
  fly.
• Ill probably just have several different
  visualizations with different parameters for the
  same algorithm run. (Run ahead of time, visualize
  all runs as if they were different algs.).