Scribd
Upload
16 views4 pages

360-Improving Intrinsic Parameters

The document discusses the process of improving the representation of intrinsic parameters in camera calibration by transitioning from non-homogeneous to homogeneous coordinates. It explains the simplification of the intrinsic matrix by removing unnecessary components and focusing on five degrees of freedom: focal length, aspect ratio, skew, and offsets. Additionally, it highlights the assumption of ideal conditions, such as square pixels and no skew, which can reduce the degrees of freedom to just the focal length during lightweight calibration.

Uploaded by

saeb2saeb
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
16 views4 pages

360-Improving Intrinsic Parameters

The document discusses the process of improving the representation of intrinsic parameters in camera calibration by transitioning from non-homogeneous to homogeneous coordinates. It explains the simplification of the intrinsic matrix by removing unnecessary components and focusing on five degrees of freedom: focal length, aspect ratio, skew, and offsets. Additionally, it highlights the assumption of ideal conditions, such as square pixels and no skew, which can reduce the degrees of freedom to just the focal length during lightweight calibration.

Uploaded by

saeb2saeb
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 4

1

00:00:00,160 --> 00:00:02,530


This is pretty ugly, and we'd like to make it nicer, and

2
00:00:02,530 --> 00:00:04,610
we're going to do that through two ways.

3
00:00:04,610 --> 00:00:09,180
First, so here we have those uglier equa, equations and the first thing you'll

4
00:00:09,180 --> 00:00:15,450
notice is kind of like before, we're dividing the x's and the y's by z.

5
00:00:15,450 --> 00:00:16,170
All right.

6
00:00:16,170 --> 00:00:18,930
And so that should tell you that see I've

7
00:00:18,930 --> 00:00:22,160
wrote up here intrinsic parameters in non-homogeneous coordinates.

8
00:00:22,160 --> 00:00:23,270
Well, guess what?

9
00:00:23,270 --> 00:00:26,750
We're going to move to homogeneous coordinates by putting this whole thing in

10
00:00:26,750 --> 00:00:28,710
a matrix formulation.

11
00:00:28,710 --> 00:00:32,470
So now we can express the whole thing in homogeneous coordinates.

12
00:00:32,470 --> 00:00:37,720
Notice that here we have z times u, z times v, z, so later when we convert back

13
00:00:37,720 --> 00:00:41,390
from homogeneous to non-homogeneous, we divide by z, and we get what we want.

14
00:00:41,390 --> 00:00:45,910
We have the x y z one over here, and we have this matrix in the middle.

15
00:00:45,910 --> 00:00:50,840
So we can rewrite this as, sort of, this very simple equation where we have
16
00:00:50,840 --> 00:00:54,310
a three-dimensional point in the camera frames.

17
00:00:54,310 --> 00:00:56,447
So remember, we've gone from some world,

18
00:00:56,447 --> 00:01:00,810
arbitrary world frame to the three-dimensional frame of the camera.

19
00:01:00,810 --> 00:01:06,380
And we go from that to the homogeneous pixel representation,

20
00:01:06,380 --> 00:01:08,390
like that, in the image.

21
00:01:08,390 --> 00:01:12,950
And the matrix that takes them from the camera to the image,

22
00:01:12,950 --> 00:01:15,720
that's the intrinsic matrix.

23
00:01:15,720 --> 00:01:20,020
Okay, so that matrix represents the intrinsic parameters, all right.

24
00:01:20,020 --> 00:01:23,750
Now fortunately, we can make it look even nicer than this.

25
00:01:23,750 --> 00:01:27,491
The first thing to notice is that the last column of K,

26
00:01:27,491 --> 00:01:32,160
when I write K as a three by four, the last column of K is zeros.

27
00:01:32,160 --> 00:01:34,400
And that doesn't really do very much, so we can get rid of it.

28
00:01:34,400 --> 00:01:36,790
And then we can do even more.

29
00:01:36,790 --> 00:01:39,630
Here we have our kinder, gentler intrinsics.

30
00:01:39,630 --> 00:01:42,290
We can use a simpler notation.

31
00:01:42,290 --> 00:01:44,920
Like I said, we're going to remove that last column.

32
00:01:44,920 --> 00:01:47,920
And we've gotten rid of the explicit thetas and things.

33
00:01:47,920 --> 00:01:51,380
And you'll notice that we have the five degrees of freedom.

34
00:01:51,380 --> 00:01:56,415
We have f, which is focal length, a, which is aspect ratio, s,

35
00:01:56,415 --> 00:02:01,440
which is for skew, and cx and cy, those are the offsets.

36
00:02:01,440 --> 00:02:04,210
By the way, remember I said we can have two different scales, right?

37
00:02:04,210 --> 00:02:07,590
A scale for one, and, for u and a scale for v?

38
00:02:07,590 --> 00:02:12,890
Or what we can have is a focal length and a relative scale between the two.

39
00:02:12,890 --> 00:02:16,080
Normally, we tend to think of it that way, as a focal length.

40
00:02:16,080 --> 00:02:18,290
That's the overall focal length of the image.

41
00:02:18,290 --> 00:02:22,700
And then, if there is a non-uniform relationship between the width and

42
00:02:22,700 --> 00:02:25,570
the height, we include that as an aspect ratio.

43
00:02:25,570 --> 00:02:28,960
And that's why there are five degrees of freedom, okay.

44
00:02:28,960 --> 00:02:32,660
Now, it turns out, this can get even easier, all right.

45
00:02:32,660 --> 00:02:36,150
And the way it gets really easy is we assume a certain niceness of

46
00:02:36,150 --> 00:02:38,060
the universe, okay.

47
00:02:38,060 --> 00:02:43,627
The niceness of the universe that we might assume is, if we have square pixels,

48
00:02:43,627 --> 00:02:49,294
if there's no skew, and if the optical center is actually in the middle, okay.

49
00:02:49,294 --> 00:02:54,690
Then we have no a, we have no s, we have no cx, we have no cy.

50
00:02:54,690 --> 00:02:57,220
All we have left is f.

51
00:02:57,220 --> 00:02:59,110
F is the only degree of freedom left.

52
00:02:59,110 --> 00:03:02,390
So when you're doing a calibration, sort of a lightweight calibration,

53
00:03:02,390 --> 00:03:06,180
what you'll do is you'll just search for f, assuming that your optic axis is in

54
00:03:06,180 --> 00:03:06,180
the middle, assuming there's no skew, and assuming that your pixels are square.

You might also like