Iterative Multiple Bounding-Box Refinements for Visual Tracking

. 2022 Mar 3;8(3):61. doi: 10.3390/jimaging8030061

Algorithm 1 Online tracking algorithm.

Input: Pre-trained CNN (

w_{1}

,…,

w_{5}

)
Initial target

b_{0}^{*}

Sequence of V frames

{F_{0}; F_{1}; \dots; F_{t}; \dots, F_{V}}

Output: Estimated sequence of target bounding-boxes

{b_{t}^{*}}_{t = 1}^{V}

1:
Randomly initialize parameters $w_{6}$ - $w_{9}$
2:
Draw training samples $X_{0}$ around $b_{0}^{*}$ , $M_{S} .$ push $(X_{0})$ , $M_{L} .$ push $(X_{0})$
3:
Update parameters $w_{4}$ - $w_{9}$
4:
for $t = 1, \dots, V$ do
5:
$b_{t}^{0} \leftarrow b_{t - 1}^{*}$
6:
for $k = 1, \dots, m a x_i t e r$ do
7:
$p_{t}^{k} = f_{b} (b_{t}^{k - 1}, F_{t})$
8:
Select N refinements $φ_{i}$ based on Transformation-Net $(p_{t}^{k})$
9:
Apply $φ_{i}$ with $i \in [1, N]$ to the bounding-box to estimate $b_{t}^{k}$
10:
end for
11:
Evaluate confidence score $q^{*} =$ Confidence-Net $(f_{b} (b_{t}^{k}, F_{t})))$
12:
if $q^{*} > 0.5$ then
13:
$b_{t}^{*} \leftarrow b_{t}^{k}$
14:
Draw sample $X_{t}$ around $b_{t}^{*}$
15:
Update $M_{S}$ and $M_{L}$ by adding $X_{t}$ and limiting their size
16:
end if
17:
if $q^{*} < = 0.5$ then (failure!) apply Re-detection to find $b_{t}^{*}$
18:
Evaluate confidence score $q^{*} =$ Confidence-Net $(f_{b} (b_{t}^{*}, F_{t}))$
19:
end if
20:
if $q^{*} < = 0.5$ then Update $w_{4} - w_{9}$ by using $M_{S}$
21:
else if t $m o d$ $10 = 0$ then Update $w_{4} - w_{9}$ by using $M_{L}$
22:
end if
23:
end for