Interleaved Scene and Object Recognition

• Applicant: Dr. Julia Vogel
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: from 2005 to 2008
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 5457134

Humans have the impressive ability to recognize even noisy or blurred objects in scenes by using global context information. For instance, we expect the grayish blob in the image to be a car rather than a wild animal when the image has been taken in the streets of a metropolitan area. The importance of context information for humans has also been proven through psychophysical experiments. Due to its applicability for content-based image or video retrieval, research in computational image understanding, i.e. the automatic description of scene class, objects in the scene, and object and scene interrelations, has recently gained increased attention. Up to now, these different recognition tasks have mainly been treated separately. However, the findings from psychophysics suggest that the combination of scene and object recognition will be also beneficial for computational scene description. The goal of this research project is to develop a computer vision system for interleaved scene classification and object categorization. The idea is to combine bottom-up knowledge, e.g. from low-level features, with top-down knowledge, e.g. information about the scene class, in an iterative manner. The hypothesis is that object classification accuracies will improve due to the use of scene context and vice versa, even if the information from low-level features is of low quality. The combination can be achieved best using statistical methods. Especially graphical models and in particular Bayesian networks will be considered since they are well suited for modeling spatial, temporal, and semantic interrelations.

• DFG Programme: Research Fellowships