Purchase this article with an account.
Meng Li, Zhiyong Yang; Statistics of natural objects and object recognition. Journal of Vision 2010;10(7):993. doi: https://doi.org/10.1167/10.7.993.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Natural visual scenes and objects entail highly structured statistics, occurring over the full range of variations in the world. Representing these statistics by populations of neurons is a paramount requirement for natural vision. The function of the visual brain, however, has long been taken to be the representation of scene features. It is thus not clear how representing individual features per se could deal with the enormous feature variations and co-occurrences of other features in the natural environment.
Here, taking object recognition as an example, we explore a novel hypothesis, namely, that, instead of representing features, the visual brain instantiates a large set of hierarchically organized, structured probability distributions (PDs) of natural visual scenes and objects and that the function of visual cortical circuitry is to perform statistical operations on these PDs to generate the full range of percepts of the natural environment for successful behaviors. To explore the merits of this hypothesis, we develop a large set of hierarchically organized, structured PDs of natural objects. First, we find all possible local structures of natural objects. Two object structures are deemed the same if they can be transformed to each other by an affine transform (displacement, rotation, scaling) and smooth nonlinear transforms. For each object structure, we then develop a PD that characterizes the natural variations of the structure. Finally, by applying this procedure at a set of spatial scales, we obtain a large set of object structures, each of which is associated with a PD. Using these object structures and associated PDs, we develop hierarchical, structured probabilistic representations of natural objects. Object recognition is performed as statistical inference. Tests on several large databases of objects show that the performance of this model is comparable to or better than some of the best models of object recognition.
This PDF is available to Subscribers Only