cecile-supercool-tracker 0.0.1

Machine learning framework for building object trackers and similarity search engines
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79

use crate::track::Feature;
use std::ops::{Mul, MulAssign, SubAssign};

/// Euclidian distance between two feature vectors
///
/// When the features distances lengths don't match, the longer feature vector is truncated to
/// shorter one when the distance is calculated
///
pub fn euclidean(f1: &Feature, f2: &Feature) -> f32 {
    let mut acc = 0.0;
    for i in 0..f1.len().min(f2.len()) {
        let mut block1 = f1[i];
        let block2 = &f2[i];
        block1.sub_assign(block2);
        block1.mul_assign(block1);
        acc += block1.reduce_add();
    }
    acc.sqrt()
}

/// Cosine distance between two vectors
///
/// When the features distances lengths don't match, the longer feature vector is truncated to
/// shorter one when the distance is calculated
///  
pub fn cosine(f1: &Feature, f2: &Feature) -> f32 {
    let mut divided = 0.0;
    let len = f1.len().min(f2.len());
    for i in 0..len {
        let mut block1 = f1[i];
        let block2 = &f2[i];
        block1.mul_assign(block2);
        divided += block1.reduce_add();
    }

    let f1_divisor = f1
        .iter()
        .take(len)
        .fold(0.0_f32, |acc, a| acc + a.mul(a).reduce_add());

    let f2_divisor = f2
        .iter()
        .take(len)
        .fold(0.0_f32, |acc, a| acc + a.mul(a).reduce_add());

    divided / (f1_divisor * f2_divisor).sqrt()
}

#[cfg(test)]
mod tests {
    use crate::distance::{cosine, euclidean};
    use crate::track::utils::FromVec;
    use crate::track::Feature;
    use crate::EPS;

    #[test]
    fn euclidean_distances() {
        let v1 = Feature::from_vec(vec![1f32, 0.0, 0.0]);
        let v2 = Feature::from_vec(vec![0f32, 1.0f32, 0.0]);
        let d = euclidean(&v1, &v1);
        assert!(d.abs() < EPS);

        let d = euclidean(&v1, &v2);
        assert!((d - 2.0f32.sqrt()).abs() < EPS);
    }

    #[test]
    fn cosine_distances() {
        let v1 = dbg!(Feature::from_vec(vec![1f32, 0.0, 0.0]));
        let v2 = dbg!(Feature::from_vec(vec![0f32, 1.0f32, 0.0]));
        let v3 = dbg!(Feature::from_vec(vec![-1.0f32, 0.0, 0.0]));
        let d = cosine(&v1, &v1);
        assert!((d - 1.0).abs() < EPS);
        let d = cosine(&v1, &v3);
        assert!((d + 1.0).abs() < EPS);
        let d = cosine(&v1, &v2);
        assert!(d.abs() < EPS);
    }
}