Cross-Classification-Based Sketch Person Re-Identification in Inconsistent Cross-Modal Identity Scenes

Bochun Huang; Fan Li; Shujuan Wang

doi:10.3788/LOP212820

Laser & Optoelectronics Progress, Volume. 60, Issue 4, 0410006(2023)

Cross-Classification-Based Sketch Person Re-Identification in Inconsistent Cross-Modal Identity Scenes

Bochun Huang^1,2, Fan Li^1,2, and Shujuan Wang^1,2、*

¹Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650500, Yunnan, China

²Yunnan Key Laboratory of Artificial Intelligence, Kunming 650500, Yunnan, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Cross modal identity inconsistence. (a) Perfect scene; (b) real scene

Download full size

Fig. 2. Triplet loss and modal information alignment based on distance. (a) Original data; (b) result with triplet loss; (c) result with our alignment loss

Download full size

Fig. 3. Sample images of the newly constructed dataset

Download full size

Fig. 4. General structure of the model

Download full size

Fig. 5. Dimensionality reduction and visualization of eigenvectors. (a) Result without triplet loss; (b) result with triplet loss; (c) result with our alignment loss

Download full size

Table 1. Composition of S-Market1501

View table

Table 1. Composition of S-Market1501

Training set number	Identity inconsistent sketch image	Identity inconsistent RGB image	Identity consistent image	All identity inconsistent image	Proportion of inconsistent identity
1	0	0	12936	0	0
2	1618	1616	9702	3234	1/4
3	2157	2155	8624	4312	1/3
4	3235	3233	6468	6468	1/2
5	4313	4311	4312	8624	2/3
6	4852	4850	3234	9702	3/4
7	6468	6468	0	12936	1

Table 2. Training algorithm

View table

Table 2. Training algorithm

Input：Initialize encoder $E$ ，classifier $C_{r}$ ， $C_{s}$ and $C$ ，the maximum number of iterations T，dataset $D$
for t=1，2，3，…，T do：
Step1：
Calculate $f_{i}^{r}$ ， $f_{j}^{s}$ by Eq（1）
Calculate ${\hat{y}}_{i}^{r}$ ， ${\hat{y}}_{i}^{s}$ by Eqs（2）and（3）
Update $E$ ， $C_{r}$ ， $C_{s}$ and $C$ by minimizing Eqs（4），（8）and（9）
Step2：
Extractor features with $E$
Calculate ${\tilde{y}}_{i}^{r}$ ， ${\tilde{y}}_{i}^{s}$ by Eqs（5）and（6）
Update $E$ ， $C$ by minimizing Eqs（7），（8）and（9） while freeze $C_{r}$ ， $C_{s}$
end for
Output：Encoder 方正汇总行 $E$

Table 3. Performance of the model on S-Market1501

View table

Table 3. Performance of the model on S-Market1501

Method		Proportion of inconsistent identity
Method		0	1/4	1/3	1/2	2/3	3/4	1
Baseline	Rank-1 /%	37.4	36.4	35.9	32.6	30.8	29.2	18.0
Baseline	mAP /%	21.1	20.6	19.3	17.1	15.2	15.7	11.8
Baseline+CycleGan	Rank-1 /%	37.7	36.4	37.5	37.3	35.9	30.9	22.7
Baseline+CycleGan	mAP /%	21.0	21.1	20.6	20.1	20.8	17.3	13.6
Ours	Rank-1 /%	41.8	41.6	40.7	40.5	39.7	39.9	29.1
Ours	mAP /%	23.5	23.2	23.1	22.0	21.1	21.4	20.9

Table 4. Performance comparison of Sketch Re-ID
View table
Table 4. Performance comparison of Sketch Re-ID
Method Rank-1 /% mAP /%
Dense-HOG+LBP+rankSVM 5.1 -
Triplet SN 9.0 -
GN Siamese 28.9 -
AFL Net 34.0 -
LMDIF 49.0 -
Ours 60.0 54.4

Table 5. Ablation experiment
View table
Table 5. Ablation experiment
Method Rank-1/% mAP/%
Baseline 32.6 17.1
Baseline+C 36.1 19.7
Baseline+D 34.9 19.9
Baseline+C+D（Ours） 40.5 22.0

Table 6. Performance of Baseline trained by Market-1501 and S-Market1501 (1) and test on Sketch Re-ID
View table
Table 6. Performance of Baseline trained by Market-1501 and S-Market1501 (1) and test on Sketch Re-ID
Datasets Rank-1 /% mAP /%
Market-1501 2.0 1.4
S-Market1501（1） 16.0 10.2

Tools

Get Citation

Copy Citation Text

Bochun Huang, Fan Li, Shujuan Wang. Cross-Classification-Based Sketch Person Re-Identification in Inconsistent Cross-Modal Identity Scenes[J]. Laser & Optoelectronics Progress, 2023, 60(4): 0410006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites