Performance Comparison of Different Invariant Feature Detectors

button3Performance Comparison of Different Invariant Feature Detectors

By: Mohd Faid Bin Yahya (PhD Student)

 

Feature detector is valuable for detecting complex object of interest in 2 dimensional (2D) or 3 dimensional (3D) images. This article reviews the performance of several invariant feature detectors. The invariant feature detectors include the SIFT technique, Harris-Laplace, Hessian-Laplace, SURF, Harris-Affine, and Hessian-Affine. The datasets are based from Bark, Bikes, Boat, Graffiti, Leuven, Trees, UBC, and Wall with 8 sequences of 6 images. Table 1 below shows the aforementioned comparison.

3a june15

Table 1. Performance comparison of invariant feature detectors [1]

 

From the table, SIFT is excellent in Bark, Leuven, and Wall datasets but could not replicate the performance in Bikes, Graffiti, and Trees datasets. As for Harris-Laplace technique, its best performance is only noticeable in UBC dataset but performed badly in all other datasets. In contrast, Hessian-Laplace technique excels in Bikes, Boat, Trees, and UBC. It’s not surprising that SURF which is the most popular invariant feature detectors performed admirably in with the most of five datasets. Harris-Affine and Hessian-Affine both scored 15 and 18 for the total of accumulated datasets. In conclusion, based from the conducted research, the best invariant feature detector would be the SURF technique while the least performance detector is Harris-Laplace.

 

Reference(s):

  1. E.R. Davies (2010). Computer and Machine Vision: Theory, Algorithms, Practicalities. 4th Edition, Elsevier.