Learning for Data Synthesis: Joint Local Salient Projection and Adversarial Network Optimization for Vehicle Re-Identification

. 2022 Dec 6;22(23):9539. doi: 10.3390/s22239539

Algorithm 1: Salient Projection Region Location Procedure

Input: image P, area A, ratio of width and height R, area ratio ranging from

A_{1}

A_{2}

,
aspect ratio ranging from

R_{1}

R_{2}

;

Output: selected region

W \leftarrow P . width

H \leftarrow P . height

A_{t} \leftarrow rand (A_{1}, A_{2}) \times A

R_{t} \leftarrow rand (R_{1}, R_{2})

H_{t} \leftarrow \sqrt{A_{t} \times R_{t}}

W_{t} \leftarrow \sqrt{A_{t} \div R_{t}}

W e i g h t \leftarrow S R A (P)

8: SRA is Spatial Relation-Aware Attention [13]. The weight matrix W with

{Weight}_{P}

rows
and

{Height}_{P}

columns can be obtained from the SRA [13]. Obtain the weights.

9: for i In

W_{t}

10: for j In

H_{t}

11:

12: if

i + W_{t} \leq W \land j + H_{t} \leq H

then

13:

S \leftarrow \sum_{i}^{i + W_{t}} \sum_{j}^{j + H_{t}} Weight

14: end if

15: end for

16: end for

17: Use the loop to find the

max S

and its

P (x, y)

18: return region

(P, W_{t}, H_{t})