The thesis will stand in the Texmex project-team. The candidate will be inte- grated in a comfortable and stimulating environment. Rennes is a dynamic middle-sized town with a student population rate above 30%.
We address the problem of video representation and recognition. This problem receives increasing interest, due to the emergence of personal capture devices such as digital cameras and video-cameras, resulting in the production of huge personal and corporate multimedia databases.
The objective of this thesis is to improve the representation of videos, in order to obtain the maximum retrieval accuracy with respect to different applications, in particular copy detection and video interpretation. Another (contradictory) requirement is a reduced indexing and search complexity. This requires to design compact but discriminative video representations.
For this purpose, the student will consider different steps of the video description and indexing pipeline. First, how to convert the signal into a vector representation, or into a set or sequence of vectors ? Many approaches exist [1,2,3,4], in particular the so-called bag-of-features representation and its derivatives [1,3,4]. However, they mainly consider videos as sequences of images, and the existing extensions (e.g., [4]) are not satisfactory. Second, following a recent PhD thesis in the Texmex pro ject-team, the student will investigate how to efficiently compare such representations, by taking into account, in particular, the specificity of the temporal axis.
Video description and retrieval, Computer Vision, Multimedia