quartz/content/notes/05-feature-description-and-matching.md
2023-03-14 13:29:38 +13:00

2.6 KiB
Raw Blame History

title tags
05-feature-description-and-matching
lecture
cosc342

last lecure: taking photographs and sticthcing them together. uses feature detection. identity interest points in image sequences. into homography matrix, map from one image into another

Blob Features

  • corner point dont keep scale
  • an alternative to corner features
    • have scale as well as location
  • can be created using the "difference of gaussian method"
    • Two Gaussians of different width
    • Subtract wide from narrow
    • Bright blobs high positive response
    • Dark blobs high negative response
  • blob detection example corners>blobs

FEATURE DESCRIPTION

  • Features are matched on the basis of some descriptor
  • This is usually a list of numbers, represented as a vector
  • Often a high dimensional vector
  • SIFT descriptors, for example, are 128-dimensional
  • Matching descriptors should be close to each other
  • This distance should be low even if the image changes
  • Translation and rotation in the image plane
  • Changes in viewing angle and distance (and therefore scale)
  • Changes in illumination, brightness, and contrast

[!INFO] we will use sift descriptors alot.

A SIMPLE FEATURE DESCRIPTOR

  • Could use pixels near the feature
  • This is easy to do
  • Works well in some cases
  • Example is greyscale, but generalises easily to colour images
  • Take an window ( odd)
  • -dimensional feature vector
  • Compare with Euclidean distance
  • Often easier to use squared distance

Feature invariance

[!INFO] does not handle rotation, changing scale or brightness

SIFT: SCALE-INVARIANT FEATURE TRANSFORM

  • Translation invariance is easy
  • Scale invariance comes from using Blob features
  • Descriptor is computed from a window around the feature
  • The size of the blob determines the size of the window
  • Brightness invariance comes from using image gradients
  • Relative brightness of pixels is fairly constant
  • Rotation invariance by estimating feature orientation
  • Window is oriented to the dominant image gradient

[!INFO] not just direct comparision, compared on multiple, permutations

[!QUESTION] how is this different from doing different perumtations of grayscale vlues

[!INFO]

SIFT FEATURE DETECTION/DESCRIPTION

  • Detect blob features and determine their scale
  • Compute a Histogram of Gradients around the blob
  • Peak(s) in the Histogram determine the orientation
  • A square region is used to compute the descriptor
  • Squares size from the blob size; orientation from HoG peaks