220 likes | 540 Views
Agenda. Overview ( 2 slides )Prior Work ( 3 slides )Continuous Measure of Coherence ( 4 slides )On-the-fly Clustering ( 2 slides )Next Organization Level: Themes ( 2 slides )ReferencesQ
E N D
1. Video Scene Segmentation Via Continuous Video Coherence Shufang Wu
http://www.sfu.ca/~vswu
vswu@cs.sfu.ca
Wednesday, July 3, 2002
2. Agenda Overview ( 2 slides )
Prior Work ( 3 slides )
Continuous Measure of Coherence ( 4 slides )
On-the-fly Clustering ( 2 slides )
Next Organization Level: Themes ( 2 slides )
References
Q & A
3. Overview (2-1)
4. A Method for Measuring Scene Boundaries
By calculating a short term memory-based model of shot-to-shot “coherence”
A One-pass On-the-fly Shot Clustering Algorithm
Application of the Above to the “Theme”
Theme: next higher level of video structure
Overview (2-2)
5. Agenda Overview ( 2 slides )
Prior Work ( 3 slides )
Continuous Measure of Coherence ( 4 slides )
On-the-fly Clustering ( 2 slides )
Next Organization Level: Themes ( 2 slides )
References
Q & A
6. Frame Dissimilarity
Normalized color histogram difference is adopted
Measure of dissimilarity, or distance
Shot Dissimilarity
Minimum dissimilarity between any two frames of two shots
Prior Work (3-1)
7. Scene Detection ( A discrete, graph-based algorithm)
Shots are clustered into sets (probable single camera positions) via a cluster similarity threshold d
Scene Transition Graph (forward temporal transitions)
A scene boundary is defined to be where:
the graph is particularly thin, that is, at a cut edge.
Problem
Scenes are often reused (graph becoming cyclic)
Improvement (Imposing a constraint on clustering)
If shots are temporally too far apart, never merge
Temporal threshold, T
Prior Work (3-2)
8. Limitations
The approach is discrete and even binary
The algorithm is sensitive to both cluster definition (via d) and scene definition (via T)
The algorithm is expensive
The discrete definition of scene is sensitive to small changes in shot similarity, and cannot represent or accommodate any ambiguity of parse.
Prior Work (3-3)
9. Agenda Overview ( 2 slides )
Prior Work ( 3 slides )
Continuous Measure of Coherence ( 4 slides )
On-the-fly Clustering ( 2 slides )
Next Organization Level: Themes ( 2 slides )
References
Q & A
10. Problem Addressing
Retain the definitions of frame & shot dissimilarity
Replace the discrete algorithm by a continuous measure
Modeling of the Recall of a Single Shot
Short term visual memory buffer of frame perception
Having a limited capacity (buffer size B)
Preserving the order of visual stimulus
Losing older frames throughout the buffer the same rate as perceived
The likelihood of a frame remaining in buffer at time t
Amount of frames of a shot of length T remaining in buffer when the final frame enters the buffer
Shot recall is defined to be proportional to their shot similarity
Continuous Measure of Coherence (4-1)
14. Agenda Overview ( 2 slides )
Prior Work ( 3 slides )
Continuous Measure of Coherence ( 4 slides )
On-the-fly Clustering ( 2 slides )
Next Organization Level: Themes ( 2 slides )
References
Q & A
15. Diameter of a Cluster
The maximum dissimilarity between any two shots within it
Shot Clustering Algorithm
From each existing cluster, Cj, select the shot that is maximally dissimilar from the incoming shot, Si
The cluster that would enlarge least is chosen (Cadd)
If the increase in diameter is less than a threshold, Si is added to Cadd, otherwise a new cluster is begun.
Threshold: Mean Shot Dissimilarity
Works reasonably well and is self-adjusting
16. Segmentation Plus Clustering
Coherence and clustering algorithms can be combined to produce output that looks something like a musical score
Example
Typical behavior
Small number of clusters in a scene (median value is 4)
The relative density of their use
17. Agenda Overview ( 2 slides )
Prior Work ( 3 slides )
Continuous Measure of Coherence ( 4 slides )
On-the-fly Clustering ( 2 slides )
Next Organization Level: Themes ( 2 slides )
References
Q & A
18. Scene Dissimilarity
Scene is represented by a cluster histogram
Shots in a cluster were all the same “color”
Dissimilarity: sum of absolute differences of the bins afterthe two cluster histograms are truncated and normalized
19. Thematic Results
A new theme is started if the incoming scene is sufficiently different from any existing ones
Threshold: mean scene-to-scene dissimilarity plus three standard deviations
Example
20. References J. R. Kender and B. L. Yeo, “Video Scene Segmentation Via Continuous Video Coherence,” CVPR '98, pp. 367-373, June 1998.
A. S. Bregman, Auditory Scene Analysis: The Perceptual Organization of Sound, MIT Press, 1990.
M. Yeung and B. L. Yeo, “Time-constrained clustering for segmentation of video into story units, ” International Conference on Pattern Recognition (ICPR’96), Vol. C, pp. 375-380, 1996.