Efficient selection of candidates in video content search

This work presents a framework for video search and retrieval without metadata, based on the sequences of detected events. These sequences encompass time information in the form of interval durations and are the basis of our method. By examining Bernoulli trials we elaborate on the confidence in locating the query clip within the video content. Both the query clip and the video content to be searched are considered as imperfect images of inaccessible reference sources, since some of the events may be miss detected or events may be the outcome of false alarms. Tolerances on the certainty of event presence and placement in time are inherently foreseen. A comparison of event sequences is thus performed between two imperfect images shifted in time by an amount which has to be determined. By counting event mismatches a decision on the presence and location of a query clip within the video content base can be taken. Totally avoiding semantic or content analysis, our approach achieves high speed at the cost of accuracy. It is thus ideally suited for a drastic and very efficient narrowing down of the candidate space. It can be used as a first preselection layer outputting potentially matching candidates for more elaborate but costly content search in the video, audio or even (transcribed) text domain.