Non-Parametric Probabilistic Image Segmentation

Abstract

We

Unsupervised Learning of Categorical Segments in Image Collections

“…2), which can be viewed as a probabilistic formulation of the model of Russell et al [3], but here segmentation and recognition happen simultaneously. It can also be seen as an extension of the image segmentation model proposed by Andreetto et al [21] where image segments are represented by a distribution of visual words, on top of local appearance.…”

“…For example, a car can be of many different colors and appear in various image locations, however, its overall appearance, as described by the visual words, is the same in all images. We model the RGB distributions f k,m with the hybrid nonparametric model proposed by Andreetto et al [21], while for the φ k we use an LDA model, as proposed by Fei-Fei et al [12] and Sivic et al [24]. In other words, if we switch off the x n component, the model reverts to an LDA one, while, if we switch off the w n component, the model reverts to model of Andreetto at al.…”

“…In other words, if we switch off the x n component, the model reverts to an LDA one, while, if we switch off the w n component, the model reverts to model of Andreetto at al. [21] In this model, visual words are grouped by segments. This enables learning topics that are related to object parts rather than to whole scenes, as is the case in the standard "bag of words" representation over the entire image.…”

“…Second, recent literature on image segmentation successfully combines the notion that images are 'piecewise smooth' with the notion that segments shapes are more often than not 'simple'. These insights have been pursued both with parametric probabilistic models [18,19], with non-parametric deterministic models [20] and with probabilistic hybrid parametricnonparametric models [21]. The latter is a very simple probabilistic formulation which, as we shall see, combines gracefully with the popular LDA model for visual recognition.…”

“…While Russell et al's segmentation is independent for each image, in our work segmentation is carried out simultaneously and each segment's definition benefits from related segments being simultaneously discovered in other images. Conversely, Andreetto et al [21] segment an entire collection of images simultaneously, while discovering the correspondence between homologous segments. However, the characteristics that pair segments are restricted to size, shape and average color.…”