TRECVID2005 Overview

TRECVID官方链接

1. Shot boundary detection Approaches in brief

The City University of Hong Kong

1. 使用spatio-temporal(SD)slides，shot的变化类型（切变，溶解）会在SD中表现出一定的特征。
2. 使用Gabor特征做运动纹理特征。
3. 使用SVM进行二值分类。（计算开销大）
   Because of image processing and use of support vector machines (SVM), the approach is computationally expensive. The CLIPS-IMAG, LSR-IMAG, NII approach was essentially a rerun of their 2004 system,which may offer some insight into the relative difficulty of the 2005 test data compared to that from 2004.
4. 切变在运动补偿后通过图像之间的比较检测得到，渐变通过计算图像的一阶和二阶temporal？导数的范数得到。性能：实时。

Fudan University

1. 帧与帧之间的相似性，可变阈值，SVM分类。
2. 使用了HSV和CIE 颜色空间。
3. 将短暂的渐变认为是切变。

The team at FX Palo Alto

built on pre-
vious years with intermediate visual features derived
from low-level image features for pairwise frame sim-
ilarities over local and longer-distances. The system
used the similarities as input to a k-nearest neigh-
bor (kNN) classifier, and added information-theoretic
secondary feature selection to select features used
in classifier. Feature selection/reduction yielded im-
proved performance but not as good as expected be-
cause of sensitivity to the training data.

Hong Kong Polytechnical University

computed
frame-frame similarities over different distances and
generated distance maps, which have characteris-
tics for cuts, gradual transitions, flashes, etc. Per-
formance was about equal to real-time.

The researchers at IBM

built upon previous CueVideo
work at TRECVID. The system was the same as
2005, except it used a different video decoder to
overcome color errors. Switching the video decoder
yielded improved performances. They noticed that
the TRECVid 2005 video encoding had no B-frames.

The Indian Institute of Technology

system fo-
cused on hard cuts only. It addressed false posi-
tives caused by abnormal lighting (flashes, reflections,
camera movements, explosions, fire, etc.) A 2-pass
algorithm - first computed similarity between adja-
cent frames using wavelets, then focused on candi-date areas to eliminate false positives. Computation
time was about the same as real-time.

The team at KDDI

developed a system that worked in the com-
pressed domain and so was fast. Luminance adaptive
thresholds and image cropping yielded good results.
They extended last year’s work by adding edge fea-
tures from discrete cosine transform (DCT) image,
color layout, and SVM learning. LaBRI from the

University of Bordeaux

used last year’s approach in
the compressed domain, computed motion and frame
statistics, then measured similarity between compen-
sated adjacent I-frames. Performance was good on
hard cuts, and fast; but not so on gradual transitions

RMIT

created a new implementation of their
sliding query window approach, computing frame
similarities among X frames before/after based on
color histograms. They experimented with differ-
ent (HSV) color histogram representations.; Feature
selection/reduction yielded improved performances.
Performance was not as good as expected because of
sensitivity to the training data;

The system devel-

oped at the University of Delft represented video as
spatio-temporal video data blocks and extracted pat-
terns from these to indicate cuts and gradual transi-
tions. The approach was efficient and is likely to be
expanded to include camera motion information.
At

Tsinghua University

researchers
re-
implemented previous years’ very successful ap-
proaches, which had evolved to a set of collaboration
rules for various detectors. The new system is
a unified framework with SVMs combining fade-
in/out detectors, gradual transition detectors and
cut detectors, each developed in previous years;
Despite individual detectors performing separately,
overall performance was very fast.

The University of Modena / University of Central Florida

team
used frame-frame distances computed based on
pixels, and based on histograms. They examined
frame difference behaviors over time to see if it
corresponded to a linear transformation. The system
was not optimized for speed.

University of Iowa’s system built on previous

years’ with a cut detection followed by gradual tran-
sition detection. Frame similarities were computed

based on color histograms, on aggregated pixel dis-

tances, and on edges. There are still some issues of
combining gradual transition and cut detection logic.
The approach taken by the University of Marburg
was based on frame similarities measured by motion-
compensated pixel differences and histogram differ-
ences for several frame distances. An unsupervised
ensemble of classifiers was then used. SVM classifiers
were trained on 2004 data. Performance was good
and quite efficient.

University of Rey Juan Carlos

concentrated on cut detection by shape and by a com-
bination of shape and color features. Shape used
Zernike moments; color used histograms from last
year. Combination methods used various logical com-
binations. The system did well on precision for cuts.
The University of Sao Paolo approach appears to be
fast but not yet among the best. No details on the
system were provided to date.