SEECS Colloquium

Temporal video segmentation partitions a video sequence into shots by detecting scene changes. It is the first step towards indexing, parsing and characterization of video data. Lot of work has been done on abrupt scene changes using static thresholds on changes in image features. Detection of gradual transition like dissolves (fade-in or fade out) are more difficult to handle. We propose a fuzzy theoretic framework for handling both types of transitions. The ambiguities and uncertainities involved in selection of thresholds are taken care of by fuzzy rules. The process of combining various feature differences is simplified to straight forward application of standard fuzy implications. We have improved upon our hand-crafted rule based fuzzy system using evolutionary learning based paradigm.

In videos and films, one source of semantic and semiotic information is its narrative structure and presetation form adopted by the director. We propose an evolutionary learning based fuzzy theoretic scheme for determining elements of form and narrative structure. To start with we identify the pace of the sequence (as slow, moderate or fast), editing style( sharp or smooth) using shot lengths, number of abrupt/gradual transitions etc. Then we analyse camera motion used (panning, zooming), and also figure out scene activity ( local or global, e.g. anchor shot ,talk show,football shots etc.)

These syntactic features are then integrated over longer time scale to evolve fuzzy rule based system for characterizing the video sequences into generic categories like sports , news, feature films. The aim is to label a video sequence using the film theory principles but without using the domain knowledge.

The next step is to extract various domain specific visual cues wherever available, such as those observed in the sports clips. We have experimented with football, cricket, tennis, golf and athletics(sprint events) clips and have been successful to a large extent in classifying these into respective categories.

Prof Biswas obtained a B.Tech degree in Electrical Engineering from IIT Madras in 1968 and Mtech and Phd from IIT Delhi in 1970 and 1974. He joined the Process control group in electrical engineering department of IIT Delhi. He later moved on to the computer science and engg deptt in 1984 where is currently a professor.

His research areas are image and video compression, soft computing based video segmentation and characterization, anisotropic diffusion based 3D segmentation, content based image retrieval, 3D visualization. He organised the international confernce ICVGIP 98 held at Delhi.

He had been collaborating with prof Mike Brady in Univesity of Oxford in the area of robot vision, and now with Prof Bajaj in UT Austin on Interrogative Synthetic Environments. He is on editorial board of Journal of network Applications and Pattern Analysis and Applications.