hand detection zhong zhang. skin and motion detector
TRANSCRIPT
Hand Detection
Zhong Zhang
Skin and motion detector• A skin color likelihood distribution and a non-skin color distribution,
denoted as and respectively are proposed.• The probability of a pixel, whose color vector is [r,g,b], being skin is
defined using Bayes rule:
• Motion detector is based on frame differencing which works as follows:– Let denote the intensity value at pixel , at the i-th frame.– By comparing with and , we compute a motion indicator value .
Skin and Motion Detector
Top 1 candidate Skin indicator
Motion indicator Skin and motion indicator
ResultTop 1 candidate
Top 2 candidates
Top 3 candidates
Top 4 candidates
Top 5 candidates
Top 8 candidates
Top 20 candidates
mp ----sm
mp ----sm
mp ----sm
mp ----sm
mp ----sm
mp ----sm
mp ----sm
Gallaudet 100
53.94%----67.31%
74.33%----81.38%
79.05%----86.35%
81.49%----89.07%
82.66%----91.61%
85.29%----96.28%
93.12%----99.61%
Liz 100 29.64%----67.67%
39.41%----80.91%
44.48%----86.84%
47.75%----90.59%
50%----93.46%
52.57%----97.66%
59.23%----100%
Tyler 100 28.58%----64.37%
41.20%----76.58%
48.77%----82.71%
53.91%----86.03%
58.20%----89.57%
64.82%----96.31%
76.86%----99.96%
Mp: hand detection using multiple proposals. Sm: skin and motion detector. The detection is considered as correct if the distance between the center of the detection box and annotation box
is less than half of face box width. The box size is [35 35].
ResultTop 1 candidate
Top 2 candidates
Top 3 candidates
Top 4 candidates
Top 5 candidates
Top 8 candidates
Top 20 candidates
mp ----sm
mp ----sm
mp ----sm
mp ----sm
mp ----sm
mp ----sm
mp ----sm
Gallaudet 100 ([40 40])
34.7%----42.23%
50.87%----56.94%
54.17%----62.33%
55.26%----66.19%
55.87%----68.90%
56.57%----74.64%
60.49%----78.16%
Liz 100 ([30 30])
5.92%----43.94%
7.14%----57.26%
7.41%----63.28%
7.47%----67.44%
7.49%----70.79%
7.53%----75.22%
8.05%----78.96%
Tyler 100([35 35])
10.56%----42.94%
13.31%----54.77%
14.23%----61.35%
15%----65.65%
15.49%----69.53%
16.22%----76.17%
19.78%----80.33%
Mp: hand detection using multiple proposals. Sm: skin and motion detector. The detection is considered as correct if the overlap score between detection and annotation is larger than 0.5
Result
Result
Result
ResultTop 1 candidate Top 2 candidates Top 4 candidates Top 8 candidates Top 20 candidates
>=0.3
>=0.4
>=0.5
>=0.3
>=0.4
>=0.5
>=0.3
>=0.4
>=0.5
>=0.3
>=0.4
>=0.5
>=0.3
>=0.4
>=0.5
Gallaudet 100 ([40 40])
62.3%
55.2%
42.2%
79.2%
73.2%
57% 88.2%
83.9%
66.2%
96.4%
93.3%
74.6%
99.6%
97.2%
78.2%
Liz 100 ([30 30]) 62.9%
54.7%
43.9%
78.4%
70.9%
57.3%
88.1%
83.4%
67.4%
96.2%
93.3%
75.2%
99.9%
98.2%
79%
Tyler 100([35 35]) 62.8%
54.9%
42.9%
75.3%
68.9%
54.8%
85.5%
81.5%
65.7%
95.7%
93.7%
76.2%
99.8%
98.9%
80.3%
The detection is considered as correct if the overlap score between detection and annotation is larger than a threshold. In the this table, this threshold can be 0.3, 0.4 and 0.5