Monday, January 17, 2011

Corner Detection

GoodFeatureToTrack

  • Default use Shi-Tomasi 94 method ("Texturedness" section): Store per-pixel 2 eigenvalues of a 2x2 matrix. It's a corner if both values above some threshold.
  • Option to use Harris method: Store Corner and Edge Response value for each pixel. It is a corner if the value is a local maxima of the surroundings (8-connected).

* Only works for single channel images *

GFTT sample (my own)

  •  Shi-Tomasi method return more keyPoints than Harris. Decrease the 'k' parameter increase the number of keyPoints detected, but still no match for Shi-Tomasi.
  • Increase the 'k' parameter to 0.3 detects no keyPoints, using stanley.png and person2.bmp.
  • Pass 'HARRIS' to FeatureDetector::create() to get Harris corner method
  • Perhaps make a generic feature param class that FeatureDetector::create() would accept.
  • Cannot detect any points on human face except teeth and eye-glasses, using person1.jpg and person2.bmp.

Readings
  • Good Features to Track, Shi, Tomasi
  • A combined corner and edge detector, Chris Harris & Mike Stephens (1988)

No comments:

Post a Comment