Paper:
Differ
FAST algorithm not like SIFT/SURF in two points
- FAST is focusing on looking at potential corners, fast enough in real-time
- FAST does not contain a descriptor design like SIFT
FAST algorithm has two versions in 2006 and 2010
- 2006: it just gives a proposal on how to search a corner point
- 2010: it improves the accuracy with machine learning
HERE WE ILLUSTRATE SOME IMPORT SENTENCE IN THE ARTICLE
SENTENCE 1:
From the article, we know the essential definition:
1 | "The original detector classifies p as a corner, |
- comment:
n: n = 12 in this case
Ip: Intensity of Pixel(Point) what means a gray value
t: threshold, filter is [0, Ip - t) & (Ip + t, 255]
SENTENCE 2:
Now the problem convert to “how to detect the contiguous pixels fast”, answer is detecting with diagonal points of the 16 pixels
the article said:
1 | "The high-speed test examines pixels 1 and 9. |
BUT THERE ARE SOME WEAKNESSES
if more details, please read the article & document
- not good when n != 12 (article)
- not optimal (article)
- data waste (document)
- too close (article)
HOW TO SOLVE THESE WEAKNESSES
1,2,3: use machine learning ID3 algorithm to create a decision tree
4: use the nonmax suppression to refine the result
ID3 algorithm:
- train from some images contain keypoint
- create a table of point1, intensity, a new bool value indict if it is a keypoint
- use the formula Hg = H(P) − H(Pd) − H(Ps) − H(Pb) calculate the gain
- once the “decision tree” is created, can be used in FAST
nonmax suppression:
sum the area intensity around every p, then keep the max in a distance