TRECVID 2007 by the Brno Group

TRECVID 2007 by the Brno Group

conference paper

en

High Level Feature Extraction     1. The runs:             - A_brU_1 - features extracted from each frame; SVM per-frame classifier trained on frames in each                 shot; simple decision tree judging shots based on per-frame results             - A_brV_2 - same as A_brU_1, but SVM trained on all training data (the first run divided the training                 data to training and cross-validation datasets), with SVM configured from the previous run     2. Significant differences between the runs:             - As expected, the second run performed generally better, in some cases notably better (which is slightly                 surprising, because besides the amount of training data, there was no change)     3. Contribution of each component:             - The low-level features appear to be good enough, though their number is relatively large and having                 more time we would experiment with reduction of the feature vector size (now 572 low level features)             - We considered using some mid-level features based on existing solutions the group has, such as face                 detection, car detection, etc., but for time constraints did not employ these in the feature vector             - The per-frame classification seems to suffer greatly from mis-annotated frames (whole shots are                 considered to share the same annotation information in our system) and could be the weakest point of                 the system             - The per-shot decision making seems to be sufficient given the data coming from the per-shot                 classification     4. Overall comments:             - see further in the paper Shot Boundary Detection We describe our approach to cut detection where we use AdaBoost boosting algorithm to create a detection classifier from a large set of features which are based on few simple frame distance measures. First, we introduce the reasons which led us to use AdaBoost algorithm, then we describe the set of features and we also discuss the achieved result. Finally, we present the possible future improvements to the current approach.

High Level Feature Extraction     1. The runs:             - A_brU_1 - features extracted from each frame; SVM per-frame classifier trained on frames in each                 shot; simple decision tree judging shots based on per-frame results             - A_brV_2 - same as A_brU_1, but SVM trained on all training data (the first run divided the training                 data to training and cross-validation datasets), with SVM configured from the previous run     2. Significant differences between the runs:             - As expected, the second run performed generally better, in some cases notably better (which is slightly                 surprising, because besides the amount of training data, there was no change)     3. Contribution of each component:             - The low-level features appear to be good enough, though their number is relatively large and having                 more time we would experiment with reduction of the feature vector size (now 572 low level features)             - We considered using some mid-level features based on existing solutions the group has, such as face                 detection, car detection, etc., but for time constraints did not employ these in the feature vector             - The per-frame classification seems to suffer greatly from mis-annotated frames (whole shots are                 considered to share the same annotation information in our system) and could be the weakest point of                 the system             - The per-shot decision making seems to be sufficient given the data coming from the per-shot                 classification     4. Overall comments:             - see further in the paper Shot Boundary Detection We describe our approach to cut detection where we use AdaBoost boosting algorithm to create a detection classifier from a large set of features which are based on few simple frame distance measures. First, we introduce the reasons which led us to use AdaBoost algorithm, then we describe the set of features and we also discuss the achieved result. Finally, we present the possible future improvements to the current approach.

TRECVID 2007, Brno, High Level Feature Extraction, Shot Boundary Detection

2008

01.03.2008

National Institute of Standards and Technology

Gaithersburg

978-1-59593-780-3

Proceedings of TRECVID 2007

NEUVEDEN

NEUVEDEN

1

6

6