Spatial Inference Machines
Probabilistic graphical models provide powerful tools for modelling interactions between random variables in computer vision, bioinformatics, and other fields. Most factors are assumed to be conditionally independent from each other to make inference tractable and for regularization sake. The structure of modelled interactions is usually defined heuristically using some a priori assumptions (e.g. square grid for 2D images). For some problems like segmentation of sparse 3D point clouds there is no obvious choice for model structure. Thus, the problem of automatic determination of structure is practically important.
However, no prominent result in learning model structure is known so far, probably due to strict nature of graphical models. Recently, inference machines have been proposed as an alternative to traditional graphical models. They are not so mathematically sound, but are more flexible than graphical models. For example, inference can be parallelized naturally, which make them specifically alluring for practical interactive applications like robotics. We believe that their flexibility also allows to determine the model structure relevant for specific task.
The ultimate goal of the project is to develop a framework based on inference machines that is able to determine the model structure to best grab spatial dependencies. As the first step, we fix the possible spatial displacements and learn sparse coefficients that weight the contribution of the corresponding spatial factors. As future work, we plan to develop a method for automatic determining those displacements. We apply the model for semantic segmentation of 3D point clouds and 2D images.
Paper: R. Shapovalov, D. Vetrov, P. Kohli. Spatial Inference Machines. IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2013. [pdf]
Funding: Microsoft Research programs in Russia.
Have you spotted a typo?
Highlight it, click Ctrl+Enter and send us a message. Thank you for your help!