Real-time feature extraction from video stream data for stream ...

ai.cs.uni.dortmund.de

Real-time feature extraction from video stream data for stream ...

7. Experiments and Evaluation

valid for todays’ video content of German broadcasting companies as well. Therefore I

manually analyzed the four news show videos (approximately 1 hour of video data) of

”Tagesschau”, included in my ”news” dataset. The results are shown in table 7.1.

Obviously todays’ news shows still have way more hard cuts than gradual transitions in

them. It looks like the rate has even been rising over time. Hence we simplifying focus

on the detection of hard cuts first.

7.1.1. Approach and Evaluation

As described in chapter 2.2, the simplest approaches for hard cut detection base on the

calculation of pair-wise differences between successive frames. This can be done in the

Stream Framework as well, by calculating the difference image, converting the difference

image to grayscale and checking, whether the average gray value of the difference image

exceeds a threshold T . This approach almost is a direct reimplementation of the approach

by [Nagasaka and Tanaka, 1992]. The difference is that T in my implementation is a

threshold on the average gray value of the difference image. Thus it does not need to be

adjusted to the resolution of the input video stream.














Figure 7.1.: Streams framework process to detect shot boundaries. A shot boundary is

declared, when two successive frames differ more than a given threshold. The difference

of two successive frames is calculated on the grayscale difference image of the frames.

Figure 7.1 shows the stream framework process for the described experiment. The threshold

t = 35 is chosen randomly. The resulting confusion matrix shows, that 97,7% (128

out of 131) of all cuts have been correctly classified as a cut (True Positives). Unfortunately

the number of False Positives (shots that were incorrectly labeled as cuts) is quite

high as well (precision = 71.5 %, recall = 97,7%).

72

More magazines by this user
Similar magazines