Parameter Study

The table below summarizes a series of experiments in which we varied each of the 9 key parameters given in the paper. We performed this test on the bear-and-dog video because it shows some interesting sensitivity to parameter values. In this video, two stuffed animals are rotated around their respective vertical axes, and they move in such a way that the 3d segmentation algorithm can sometimes fail to distinguish them over a small number of frames.

For the same of readability, we have labelled the columns of the table with Roman letters, except for those that were given a specific name in the paper. The columns are as follows...
Parameters:

a = max ratio of patch dimensions
b = min correlation during first pass of tracking
c = min correlation for trimming tracks (2nd pass)
d = min patches to support a camera
e = min cameras to support a patch
f = max reprojection error
ω = min overlap between two tracks
ε = max reprojection error for adding tracks to component
ν = min tracks to form component

Results:

g = number of pure bear components
h = numbers of pure dog components
i = number of mixed components (having both bear and dog patches, and generally coming out bad)
j = reprojection error for principal bear
k = # of cameras in principal bear
l = # of patches in principal bear
m = reprojection error for principal dog
n = # of cameras in principal dog
o = # of patches in principal dog

Each row of the table begins with a segmentation video. This shows the input video overlaid with boxes which indicate by color the component to which they belong. The dog is by far the harder object to model, so we have also included a fly-around video of the principal dog component that resulted from each test, to convey the subjective quality of the 3d reconstruction.

The optimal values for this shot are given in the first row. All other tests used the same values, except for the parameter being varied in that test. To avoid visual clutter, we only showed the varied parameter in subsequent rows.

a b c d e f ω ε ν g h i j k l m n o
6 0.8 0.85 6 6 1.5 4 1.25 55 1 1 0 0.30924 150 2661 0.395766 150 1147
4 2 2 1 0.323411 117 2687 0.359234 101 624
8 0 1 2 0.322172 150 2149 0.430795 150 964
0.75 0 1 2 0.308828 150 1927 0.429461 150 882
0.85 0 3 3 0.309155 143 2675 0.412121 99 509
0.8 1 1 2 0.44573 135 1926 0.415934 150 868
0.9 1 1 2 0.339318 93 1977 0.406593 137 1203
2 1 2 0 0.307453 150 2662 0.392321 110 732
10 1 2 0 0.307709 150 2658 0.393071 99 736
2 1 1 0 0.30924 150 2661 0.395766 150 1147
10 1 2 0 0.30924 150 2661 0.390998 100 729
1.25 1 3 0 0.292799 150 2662 0.360393 101 700
1.75 1 1 0 0.324893 150 2670 0.453823 150 1180
2 1 2 0 0.309283 150 2662 0.392887 103 737
6 1 2 0 0.307374 150 2658 0.391401 113 736
1 1 4 0 0.298957 150 2657 0.384026 104 737
1.5 2 2 1 0.343016 119 2678 0.40638 108 708
25 1 2 1 0.30924 150 2661 0.395766 150 1147

a	b	c	d	e	f	ω	ε	ν	g	h	i	j	k	l	m	n	o
6	0.8	0.85	6	6	1.5	4	1.25	55	1	1	0	0.30924	150	2661	0.395766	150	1147
4									2	2	1	0.323411	117	2687	0.359234	101	624
8									0	1	2	0.322172	150	2149	0.430795	150	964
	0.75								0	1	2	0.308828	150	1927	0.429461	150	882
	0.85								0	3	3	0.309155	143	2675	0.412121	99	509
		0.8							1	1	2	0.44573	135	1926	0.415934	150	868
		0.9							1	1	2	0.339318	93	1977	0.406593	137	1203
			2						1	2	0	0.307453	150	2662	0.392321	110	732
			10						1	2	0	0.307709	150	2658	0.393071	99	736
				2					1	1	0	0.30924	150	2661	0.395766	150	1147
				10					1	2	0	0.30924	150	2661	0.390998	100	729
					1.25				1	3	0	0.292799	150	2662	0.360393	101	700
					1.75				1	1	0	0.324893	150	2670	0.453823	150	1180
						2			1	2	0	0.309283	150	2662	0.392887	103	737
						6			1	2	0	0.307374	150	2658	0.391401	113	736
							1		1	4	0	0.298957	150	2657	0.384026	104	737
							1.5		2	2	1	0.343016	119	2678	0.40638	108	708
								25	1	2	1	0.30924	150	2661	0.395766	150	1147