Compositionality, MDL Priors, and Object Recognition
Abstract
Images are ambiguous at each of many levels of a contextual hi(cid:173) erarchy. Nevertheless, the high-level interpretation of most scenes is unambiguous, as evidenced by the superior performance of hu(cid:173) mans. This observation argues for global vision models, such as de(cid:173) formable templates. Unfortunately, such models are computation(cid:173) ally intractable for unconstrained problems. We propose a composi(cid:173) tional model in which primitives are recursively composed, subject to syntactic restrictions, to form tree-structured objects and object groupings. Ambiguity is propagated up the hierarchy in the form of multiple interpretations, which are later resolved by a Bayesian, equivalently minimum-description-Iength, cost functional.
Cite
Text
Bienenstock et al. "Compositionality, MDL Priors, and Object Recognition." Neural Information Processing Systems, 1996.Markdown
[Bienenstock et al. "Compositionality, MDL Priors, and Object Recognition." Neural Information Processing Systems, 1996.](https://mlanthology.org/neurips/1996/bienenstock1996neurips-compositionality/)BibTeX
@inproceedings{bienenstock1996neurips-compositionality,
title = {{Compositionality, MDL Priors, and Object Recognition}},
author = {Bienenstock, Elie and Geman, Stuart and Potter, Daniel},
booktitle = {Neural Information Processing Systems},
year = {1996},
pages = {838-844},
url = {https://mlanthology.org/neurips/1996/bienenstock1996neurips-compositionality/}
}