Patch-based Object Recognition Using Discriminatively Trained Gaussian Mixtures

Abstract

We present an approach using Gaussian mixture models for part-based object recognition where spatial relationships of the parts are explicitly modeled and parameters of the generative model are tuned discriminatively. These extensions lead to great improvements of the classification accuracy. Furthermore we evaluate several improvements over our baseline system which incrementally improve the obtained results which compare favorable well to other published results for the three Caltech tasks and the PASCAL evaluation 05 tasks.

Retrieval of Images with Objects of Specific Size, Location, and Spatial Configuration

“…The work in [11] incorporates spatial layout by introducing a Gaussian location model per visual word and encoding only the absolute spatial information. Utilizing localized grids into the feature representation is also a common approach to integrate spatial information [2,10,13,24]. These methods often result in high-dimensional representation and rely on a predefined partitioning of the image which is independent of its content.…”

Integration of active learning in a collaborative CRF

“…Methods based on local descriptors [10,11], transform images into a large collection of local features invariant to geometric and photometric changes. Other models combine segmentation and recognition into a single process by coupling patches with compatible part configurations and pixels with coherent low-level features [3].…”

Image Retrieval and Annotation Using Maximum Entropy

“…These two methods were also used in the PASCAL visual object classes challenge 2006. The third method [17] we submitted to the PASCAL challenge could not be applied to this task due to time and memory constraints.…”