C. Semantic Concepts
This method was based on video retrieval on the basis
of semantic concepts present in the video. This method of
video retrieval consists of the following steps.
C.4 Clustering:
After obtaining the DCT coefficients (feature vectors)
for all the selected frames of the shot, we separated the
feature vectors belonging to particular concepts. For
example, for the concept ‘mountain’, we searched all the
videos for the shots labeled ‘mountain’. We obtain the
feature vectors from all the shots having the concept and ran Figure 4
the ‘kmeans clustering’, algorithm on the feaure vectors.
This clustering was done in 100 dimensional space in which 3. RESULTS
all feature vectors were points. We divided each concept In this part of the report we discuss results from the three
into three clusters. The three centroids were saved for each approaches of video retrieval.
concept.
A. Content Based Video Retrieval
C.5 Automatic Annotation The results are based on first seasons of TV shows
Next task was to automatically annotate the videos, for Merlin and Sons of Anarchy (which were used to make the
this a shot to be annotated was first converted to feature data set).
vectors (DCT coefficients for all the frames (or selected
frames)). Then the clustering was done on the resultant A.1 Quantization Levels
feature vectors, as described in the previous section. The In table 1, results are shown for different quantization
centroids obtained were matched to the centroids of all the levels. As discussed earlier, we were only using the
concepts (previous section), and the shot was labeled with histograms to compute the feature vectors, hence these
the concept for which the score was minimum. We used the vectors do not have any information regarding the concept
sum of absolute difference as the distance measure. Below of video, objects in video, etc. Hence, feature vectors of two
are the results for automatic annotation: entirely different TV shows (on the basis of concepts and
objects in videos) can give high similarity as shown in table.
Video Name: Formula 1 2014 Australian Grand Prix Official Race The table was computed by giving ‘Merlin S01E11’ as a
Edit [1080p] - Video Dailymotion.avi
Given Labels for shot 2: people query. The similarity measure used was absolute difference.
Given Labels for shot 3: race, road
Given Labels for shot 4: race, people If we assume 256 quantization level retrieval to be the
perfect case, we can see that as the quantization levels
Computed Labels for shot 2: road, people, building, race, person increase the error (w.r.t. to 256 quantization levels) in the
Computed Labels for shot 3:road, race, people, building, person retrieval of videos also increase. For example we can see
Computed Labels for shot 4:building, person, people, race, road from the table that for 128 quantization levels although the
top 10 retrieved videos remain the same but there ranks
The 9 frames from the shot 4 of the video are shown in change i.e. positions 4&5 are switched, positions 6 & 7 are
figure 4. The frames from the shot show that the annotation also no longer the same and so on. But if we are only
is not bad, as the concept of building was missed in labeling, concerned with top 10 matches (irrespective of rank), this
but automatic annotation did not miss it. can work.
For quantization levels of 64 and 8, new videos are matches with certain correlation, it is considered to be a
introduced in the top 10, which shows more error. Here similar video, but for CBVR every sampled frame should
again the top 3 videos are the same (not considering rank). match considerably with every frame of the reference video
So depending on the application we can use any of the for a considerable match.
quantization levels.
References
A.2 Global Features
Table 2 gives a comparison for global and local feature [1] Yang Cai, Linjun Yang, ‘Large Scale Near Duplicate
vectors. The table was computed for 256 quantization levels web Video Retrieval: Challenges and Approaches’
and using absolute difference as similarity measure.
As can be seen from table 2, in the top 10 retrieved [2] Alexandre Karpenko, Parham Aarabi, ‘Tiny Videos: A
videos, only 2 are different, shown with red color, although Large Data Set for Nonparametric Video Retrieval and
rank can differ. But global features with same quantization Frame Classification’
levels require much less space and are time efficient. Hence
depending upon the application we can used global or local
features.
B. Tiny Videos
C. Semantic Concepts
To retrieve the video, user searches for a word and the
videos in which the shot labeled with the entered word are
shown on screen. The videos are shown in the order such
that the video with maximum number of shots labeled with
the concepts is shown in the top and so on.
The results for searching different key words are shown
in table 5.
Sr. Local Feature Vector Global Feature Vectors No Corr = 0.5 Corr = 0.6 Corr = 0.7 –
No 0.999
1 'Merlin S01E01’ ‘Merlin S01E01’ ‘Merlin S01E01’
1 'Merlin S01E11’ ‘Merlin S01E11’ 2 'Merlin S01E02’ 'Merlin S01E03’
2 'Merlin S01E05’ 'Sons.of.Anarchy.S01E10’
3 'Merlin S01E03’ 'Merlin S01E06’
3 'Merlin S01E01’ 'Sons.of.Anarchy.S01E09’
4 ‘Sons.of.Anarchy.S01E03’ 'Sons.of.Anarchy.S01E05’ 4 'Merlin S01E04’ 'Merlin S01E07’
5 ‘Sons.of.Anarchy.S01E10’ 'Merlin S01E05’ 5 'Merlin S01E05’ 'Merlin S01E11’
6 'Sons.of.Anarchy.S01E07’ ‘Sons.of.Anarchy.S01E07’ 'Merlin S01E06’ 'Sons.of.Anarchy.S01E01’
6
7 'Merlin S01E10’ 'Sons.of.Anarchy.S01E08’
7 'Merlin S01E07’ 'Sons.of.Anarchy.S01E05’
8 'Sons.of.Anarchy.S01E08’ ‘Sons.of.Anarchy.S01E03’ 8 'Merlin S01E08’ 'Sons.of.Anarchy.S01E09’
9 'Sons.of.Anarchy.S01E05’ 'Sons.of.Anarchy.S01E06’ 'Merlin S01E09’ 'Sons.of.Anarchy.S01E11’
9
No Corr = 0.5 Corr = 0.9 No. KeyWord: Key Word: Vehicle Key Word: arrow
1 Merlin S01E01 ‘Merlin S01E01’ Mountain
2 Merlin S01E02 ‘Merlin S01E02’
1 Sons of Anarchy Sons of Anarchy S01E02 MerlinS01E01
3 Merlin S01E03 S01E01
4 Merlin S01E04 2 MerlinS01E01 Sons of Anarchy S01E01 MerlinS01E02
5 Merlin S01E05
3 MerlinS01E02 MerlinS01E01 MerlinS01E03
6 Merlin S01E01
7 Merlin S01E02 4 MerlinS01E03 MerlinS01E02 Sons of Anarchy
8 Merlin S01E03 S01E01