Parameter Estimation for Optimal Object Recognition: Theory andApplication

  • Authors:
  • Stan Z. Li

  • Affiliations:
  • School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore 639798

  • Venue:
  • International Journal of Computer Vision
  • Year:
  • 1997

Quantified Score

Hi-index 0.00

Visualization

Abstract

Object recognition systems involve parameters such asthresholds, bounds and weights. These parameters have to be tunedbefore the system can perform successfully. A common practice is tochoose such parameters manually on an {\it add\ hoc} basis, which is adisadvantage. This paper presents a novel theory of parameterestimation for optimization-based object recognition where the optimalsolution is defined as the global minimum of an energy function. Thetheory is based on supervised learning from examples. {\it Correctness} and {\it instability} are established as criteria for evaluating theestimated parameters. A correct estimate enables the labeling impliedin each exemplary configuration to be encoded in a unique global energyminimum. The instability is the ease with which the minimum is replacedby a non-exemplary configuration after a perturbation. The optimalestimate minimizes the instability. Algorithms are presented forcomputing correct and minimal-instability estimates. The theory isapplied to the parameter estimation for MRF-based recognition andpromising results are obtained.