Dear Alan, here is the input from TZI (University of Bremen). Best regards, Andrea ------- The Center for Computing Technologies (TZI), University of Bremen, Gremany, participated in the video analysis track in the shot detection task and in the feature extraction task (features indoors and outdoors). The shot detection approach is based on histogram differences. It is divided into two steps - feature extraction and shot boundary detection. Firstly, the histogram differences are calculated for the entire video in real time. Secondly, shot boundaries are detected. The advantage of this approach is the possibility to set adaptive thresholds for the shot boundary detection considering all extracted features of the complete video sequence. The adaptive threshold is set to a percentage of the maximum of all calculated difference values of the video. In the case of gradual changes, often multiple shot boundaries are detected. Therefore multiple detected shot boundaries that follow each other within a short temporal interval are grouped together and a gradual change is detected beginning with the first and ending with the last shot boundary in the interval. For the feature extraction task it is examined whether it is possible to classify indoor and outdoor shots by their color distribution. In order to analyze the color distribution, first order statistical features are used, which are extracted from the histograms of the three color channels (RGB) and the grey level histogram. The features calculated from each histogram are average, variance, and amount of peaks, normalized to an interval [0...1]. In order to classify the shots into indoor and outdoor shots, a feed forward neural net with backpropagation learning was trained. At the input layer the 12 statistical features mentioned above are presented. The output layer consists of two neurons that take on values between 0 and 1 measuring the probability for the features indoors or outdoors to be present in the shot. Two hidden layers each with 20 neurons are initialized with random weights. In order to train the neural net, some videos from the feature development collection were chosen. The shots are classified manually to generate 323 training data sets, 178 for indoors and 145 for outdoors. In order to classify the shots from the feature extraction test collection, a set of n key frames is extracted from each shot. Every k-th frame of a shot is used as a key frame, but in order to be more independent of inaccuracies during the shot detection and of gradual changes (e.g., wipes, fades, or dissolves) a number of frames around the shot boundaries is skipped. In order to classify a shot, the set of n key frames is presented to the neural net. For each of the two output neurons a list is obtained containing n values, one for each key frame. The median for each list is calculated to obtain the final probabilities for the shot to be indoors or outdoors. In order to measure the accuracy of the classification result, the difference between the median values of the indoors and the outdoors neuron is calculated. If the difference exceeds a threshold the shot is classified to contain the feature with the higher probability. The difference is also used for the ranking. "Prof. Alan Smeaton" wrote: > TREC Video colleagues > > A while ago I sent around a message asking for a short summary of what > you did in TREC video track this year, to help Paul and I write the > overview paper for the notebook proceedings. This overview paper will > describe the track, the task, data, evaluation measures and a summary of > results, and will also have a short soundbite on what each group did. > > I've had input from 6 groups (Imperial/Bristol, CLIPS-IMAG, CWI, NUS > Singapore, MSR Asia and ourselves), and many thanks to all for that, but > there are 10 other groups who did something in this year's video track > who have yet to get back to me. > > I haven't hassled you about this because I knew you were all writing > your overview papers for next Monday, but right *NOW*, I need a couple > of paragraphs from each of the 10 groups yet to reply to me. So, hit > the "REPLY" button on your email client, and type up 4 or 5 lines. > > Many thanks. > > - Alan > > From: Prof. Alan Smeaton [mailto:asmeaton@computing.dcu.ie] > Sent: Thursday, October 17, 2002 3:21 PM > To: Multiple recipients of list > Subject: Input required for TREC Video Track overview paper > > 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. > > - Alan -- <><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><> Dipl. Inform. Andrea Miene University of Bremen - FB 3 - Computer Science Center of Computing Technology - Image Processing Department PO Box 330 440, Universitaetsallee 21-23, D-28359 Bremen Tel: +49+421-218-7827/7090, Fax: +49+421-218-7196 WWW: http://www.tzi.org/~andrea <><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>
S/MIME Cryptographic Signature