Re: Input required for TREC Video Track overview paper
Prof. Alan Smeaton wrote:
>
> Dear TREC Video Track participants
>
> Paul and I have to write an overview paper describing the track
> activities, for the TREC notebook paper, and we have the same deadline
> as you ... Nov 3.
>
> Could each of you who participated in the shot bound detection, feature
> detection, and/or search tasks send me a couple of paragraphs for this
> overview, outlining what you did, thanks.
Hi Alan,
Here is some meterial aboout the reference segmentation, most of it coming
from "http://www-nlpir.nist.gov/projects/t2002v/CSBR_mpeg7/README" (which
is not fully up to date since there seems to be nothing about the fusion
of segments of duration less than 2 seconds). I will try to send you also
something about our system and experimentations.
----------------------------------------------------------------------
Information about the reference segmentation:
The segmentation was performed using the system used by CLIPS-IMAG for
the TREC 2001 SBD task (variant "CLIPS-1"). More information about this
system ca be found from the CLIPS-IMAG TREC 2001 paper:
<http://trec.nist.gov/pubs/trec10/papers/clips-imag-bin.pdf>
From the TREC 2001 SBD evaluation, it appeared that this system was
quite good for cuts but not so (in the average) for gradual transitions.
Also, it is a highly recall oriented system (which means significant
over-segmentation). More precisely, the figures were :
cuts gradual all
Insertion rate: 10.5% 56.6% 26.6%
Deletion rate: 1.2% 29.3% 11.0%
Since the TREC 2002 data is older, we expect that there will be
less special effects and that the gradual transitions will be a bit
more accurately extracted (the CLIPS system actually detects only
dissolves in the gradual transtion category). From the 295 files,
the system detected 20868 cuts and 14231 dissolves which indicate a
ratio comparable to the one of the TREC 2001 SBD collection.
The CLIPS system used the MPEG-2 decoder from the MPEG Software
Simulation Group (http://www.mpeg.org/MPEG/MSSG/#source). The frame
numbering and therefore the shot bound timings are defined according
to the output of this decoder. Both frame numbering and shot bound
timing can vary slightly (a few frames) if one uses another decoder.
The information about transition type (CUT vs. GRADUAL) and duration
(for GRADUAL ones) are not available from the reference segmentation
since only contiguous shots are described. When gradual (dissolve)
transitions were found, the middle of the extracted transition was
chosen as the shot transition, leaving half of the transition duration
at the beginning of the next shot and at the end of the previous one.
This original segmentation containing 35394 shots was then modified
in order to eliminate segments with durations of less than 2 seconds,
these being considered as unxeploitable for many feature extraction
tools and/or for manual checking. This elimination was performed by
fusing short segments with adjacent ones. The fusion is done by
simply removing the weakest transition surrounding the segments.
By doing so the number of shots is reduced to 24263. A few erroneous
transitions are removed during the process but most of the removed
ones were correct ones. Many actual shots are shorter that 2 seconds.
The consequence of that is that many shots in the reference segmentation
actually contain several shots with transitions clearly visible.
The segmentation algorithm is not perfect anyway and some internal
transitions, even rather obvious ones, were already present in the
original segmentation.
----------------------------------------------------------------------
I am not fully satisfied with this comment. Please feel free to modify
it as you feel necessary. I also joined the shot duration histograms
before and after shot fusion.
Best regards.
Georges.
Email: Georges.Quenot@imag.fr
CLIPS-IMAG, 385, rue de la Bibliothèque, B.P. 53, 38041 Grenoble Cedex 9
Tel: (33-4) 76 63 58 55, Fax: (33-4) 76 44 66 75
Histogram of shot lengths, 35387 shots in 295 video documents
Part 1: in frames:
1: 1 0.00%
2: 1295 3.66%
3: 910 2.57%
4: 214 0.60%
5: 1163 3.29%
6: 220 0.62%
7: 195 0.55%
8: 183 0.52%
9: 150 0.42%
10: 358 1.01%
11: 140 0.40%
12: 190 0.54%
13: 184 0.52%
14: 185 0.52%
15: 432 1.22%
16: 203 0.57%
17: 190 0.54%
18: 175 0.49%
19: 171 0.48%
20: 311 0.88%
21: 221 0.62%
22: 181 0.51%
23: 199 0.56%
24: 201 0.57%
25: 255 0.72%
26: 188 0.53%
27: 149 0.42%
28: 156 0.44%
29: 143 0.40%
30: 240 0.68%
1-15: 5820 16.45%
16-30: 2983 8.43%
Part 2: in seconds (times 1.001):
1-2: 3774 10.66%
2-3: 2999 8.47%
3-4: 2516 7.11%
4-5: 2226 6.29%
5-6: 1908 5.39%
6-7: 1693 4.78%
7-8: 1401 3.96%
8-9: 1168 3.30%
9-10: 994 2.81%
10-20: 5178 14.63%
20-30: 1553 4.39%
30-40: 611 1.73%
40-50: 271 0.77%
50-60: 146 0.41%
60-70: 61 0.17%
70-80: 43 0.12%
80-90: 18 0.05%
90-100: 10 0.03%
100-200: 13 0.04%
200-300: 0 0.00%
300-400: 1 0.00%
400-500: 0 0.00%
500-600: 0 0.00%
600-700: 0 0.00%
700-800: 0 0.00%
800-900: 0 0.00%
900-1000: 0 0.00%
> 1000: 0 0.00%
Histogram of shot lengths, 24263 shots in 295 video documents
Part 1: in frames:
1: 0 0.00%
2: 0 0.00%
3: 0 0.00%
4: 0 0.00%
5: 0 0.00%
6: 0 0.00%
7: 0 0.00%
8: 0 0.00%
9: 0 0.00%
10: 0 0.00%
11: 0 0.00%
12: 0 0.00%
13: 0 0.00%
14: 0 0.00%
15: 0 0.00%
16: 0 0.00%
17: 0 0.00%
18: 0 0.00%
19: 0 0.00%
20: 0 0.00%
21: 0 0.00%
22: 0 0.00%
23: 0 0.00%
24: 0 0.00%
25: 0 0.00%
26: 0 0.00%
27: 0 0.00%
28: 0 0.00%
29: 0 0.00%
30: 0 0.00%
1-15: 0 0.00%
16-30: 0 0.00%
Part 2: in seconds (times 1.001):
1-2: 189 0.78%
2-3: 3223 13.28%
3-4: 2796 11.52%
4-5: 2389 9.85%
5-6: 2039 8.40%
6-7: 1788 7.37%
7-8: 1458 6.01%
8-9: 1226 5.05%
9-10: 1045 4.31%
10-20: 5329 21.96%
20-30: 1589 6.55%
30-40: 622 2.56%
40-50: 272 1.12%
50-60: 150 0.62%
60-70: 63 0.26%
70-80: 41 0.17%
80-90: 20 0.08%
90-100: 10 0.04%
100-200: 13 0.05%
200-300: 0 0.00%
300-400: 1 0.00%
400-500: 0 0.00%
500-600: 0 0.00%
600-700: 0 0.00%
700-800: 0 0.00%
800-900: 0 0.00%
900-1000: 0 0.00%
> 1000: 0 0.00%
Date Index |
Thread Index |
Problems or questions? Contact list-master@nist.gov