Bayesian approach for integrating spatial and temporal information during scene analysis

Abstract

Statement of Work:The objective of the proposed effort is to develop a computational framework for constructing a system for autonomous scene analysis that will be able to overcome shortcomings of current approaches and specifically be able to learn object categories from a small number of observations. The proposed effort will use a probabilistic Bayesian approach to learn object categories and relations between objects and object parts. It will develop a framework for using both spatial and temporal contextual information and for integrating the two types of information during the learning and inference stages.Objective:The objective of the proposed effort is to develop a computational framework for constructing a system for autonomous scene analysis that will be able to overcome shortcomings of current SOA approaches and specifically will be able to learn object categories from a small number of observations and utilize spatial and temporal contextual information during the recognition process. The program officer will work with a graduate student and provide guidance in addressing various research issues related to the proposed work.Approach:The proposed effort will use a probabilistic Bayesian approach to learn object categories and relations between objects and object parts. It will develop a framework for using both spatial and temporal contextual information and forintegrating the two types of information during the learning and inference stages. In addition, the proposed effort will use a Bayesian approach and a Dirichlet prior to reduce the dimensionality of the data by constructing features can capture both non-linear dependences and higher order statistics.Overall Merit and ONR Mission/Relevance:Current state-of-the-art (SOA) approaches for scene analysis critically depend on classifiers such as the SupportVectors Machine and Deep Learning architectures which are usually trained on small image regions using largeamounts of labeled data. Training on each object class is done from scratch regardless of how similar it is to otheralready learned classes. During the recognition stage, most systems still use a sliding window approach and to detectobjects at different scales the image has to be re-sized and the scanning repeated. This tunnel vision and brute forceapproach is very inefficient, brittle and inaccurate. One of the main limitations is that it doesn t take into considerationcontextual information.

Document Details

Document Type
DoD Grant Award
Publication Date
Sep 30, 2016
Source ID
N000141612904

Entities

People

  • Ernst Niebur

Organizations

  • Johns Hopkins University
  • Office of Naval Research
  • United States Navy

Tags

Fields of Study

  • Computer science

Readers

  • Computer Vision.
  • Distributed Systems and Data Platform Development
  • Neural Network Machine Learning.

Technology Areas

  • AI & ML