Monday, February 23, 2009

Epitomic Location Recognition

Main idea: It uses a generative model based on epitomic image analysis. This analysis is based in a probabilistic approach.
  • The appearance and geometric structure of the environment is captured into this representation.
  • It has the ability to model translation and scale invariance together with the fusion of diverse visual features yield enhanced generalization with economical training.
  • The recognition of a location class is achieved by convolving the query image and the learned epitome.
  • It doesnt estimate the accurately the camera position.
  • Occlusions, reflections or non-rigid motions are modeled as noise whose variance changes for different regions within the environment.
  • These epitomes are generative, probabilistic models and various sources of uncertainty are captured in the variance maps.
  • In this model, an image is extracted from a larger latent image, the epitome, at a location given by a discret mapping.
  • Every N x M image I is generated from a Ne x Me location epitome e.
  • In this approach Ne x Me translations and 3 scales are considered.
Inference and learning
  • Every image is independient and identically distributed given the epitome.
  • The goal is to find a single epitome e* which maximizes the probability of the observations. This is achivied by EM algorithm.
Visual features
  • raw RGB pixels
  • gist features
  • disparity maps
  • local histograms
It uses a stereo camera.

No comments:

Post a Comment