Shape Quantization and Recognition with Randomized Trees

Abstract

We explore a new approach to shape recognition based on a virtually in nite family of binary features (\queries") of the image data, designed to accommodate prior information about shape invariance and regularity. E a c h query corresponds to a spatial arrangement o f s e v eral local topographic codes (\tags") which are in themselves too primitive and common to be informative about shape. All the discriminating power derives from relative angles and distances among the tags. The important attributes of the queries are (i) a natural partial ordering corresponding to increasing structure and complexity (ii) semi-invariance, meaning that most shapes of a given class will answer the same way t o t wo queries which are successive in the ordering and (iii) stability, since the queries are not based on distinguished points and substructures.No classi er based on the full feature set can be evaluated and it is impossible to determine a priori which arrangements are informative. Our approach i s t o s e l e c t informative features and build tree classi ers at the same time by inductive learning. In e ect, each tree provides an approximation to the full posterior where the features chosen depend on the branch which i s t r a versed. Due to the number and nature of the queries, standard decision tree construction based on a xed length feature vector is not feasible. Instead we e n tertain only a small random sample of queries at each n o d e , constrain their complexity to increase with tree depth, and grow m ultiple trees. The terminal nodes are labeled by estimates of the corresponding posterior distribution over shape classes. An image is classi ed by sending it down every tree and aggregating the resulting distributions.The method is applied to classifying handwritten digits and synthetic linear and nonlinear deformations of three hundred Latex symbols. State-of-the-art error rates are achieved on the NIST database of digits. The principal goal of the experiments on Latex symbols is to analyze invariance, generalization error and related issues, and a comparison with ANN methods is presented in this context. 2

Spacetime Forests with Complementary Features for Dynamic Scene Recognition

“…Following previous work in spacetime texture analysis [7], the spatiotemporal responses, (2), are further combined to yield measures of dynamic information independent of spatial appearance, as follows. In the frequency domain, motion occurs as a plane through the origin [35].…”

“…Previously, RF classifiers have been applied to a variety of vision tasks, e.g., [4,19,23,36]. Detailed descriptions of RFs are available elsewhere, e.g., [2,5,6].…”

Image-based object recognition in man, monkey and machine

“…For instance, Bricolo et al (1997) employ features characterized by small brightness regions that can be located at any position within the image, thereby facilitating flexible object representations. Amit and Geman (1997) use similar features, but relate the spatial positions of such features in a manner that allows for a highly robust matching scheme. There have also been attempts to develop models that use less local representations of shape.…”

“…Consider that images may be described as collections of local measures of the image at different locations (e.g. the output of oriented receptive fields, Edelman, 1993;small pixel regions, Bricolo et al, 1997;Amit and Geman, 1997; qualitative and quantitative measures of shape, color, texture, etc.). At one extreme, it may be possible to represent these features in a completely unordered fashion, thereby losing all information regarding the spatial relation between one local feature and the next.…”

“…One appealing aspect of compositionality is that it allows input shapes to be matched to stored representations through randomized tree searches (Amit and Geman, 1997). That is, rather than attempting to match all of the features of the representation during recognition, a series of binary 'queries' are performed.…”