Cross-Domain Spatial Co-Attention Network for Sketch-Based Image Retrieval

Lingzhi Yu; Xifan Zhang

doi:10.3788/LOP202259.2215009

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 22 >Page 2215009 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 22, 2215009 (2022)

Cross-Domain Spatial Co-Attention Network for Sketch-Based Image Retrieval

Lingzhi Yu^* and Xifan Zhang

Author Affiliations

School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China

show less

DOI: 10.3788/LOP202259.2215009 Cite this Article Set citation alerts

Lingzhi Yu, Xifan Zhang. Cross-Domain Spatial Co-Attention Network for Sketch-Based Image Retrieval[J]. Laser & Optoelectronics Progress, 2022, 59(22): 2215009 Copy Citation Text

show less

Fig. 1. Network architecture of the proposed method

Download full size

Fig. 2. Some retrieval examples of proposed method on Sketchy dataset

Download full size

Method	Sketchy	TU-Berlin
3D shape^［21］	0.084	0.054
HOG^［22］	0.115	0.091
GF-HOG^［6］	0.157	0.119
SHELO^［23］	0.161	0.123
LKS^［24］	0.190	0.157
SaN^［7］	0.208	0.154
Siamese CNN^［25］	0.481	0.322
Siamese-AlexNet^［18］	0.518	0.367
GN Triplet^［17］	0.529	0.187
Triplet-AlexNet^［18］	0.573	0.448
DSH^［18］	0.783	0.570
GDH^［10］	0.810	0.690
Semi3-Net^［11］	0.916	0.800
Proposed method（discrete）	0.932	0.797
Proposed method（continuous）	0.933	0.799

Table 1. mAP of different methods on Sketchy and TU-Berlin datasets

Method	mAP
w/o edge map branch	0.899
w/o spatial co-attention	0.923
w/o intra-class loss	0.928
w/o quantization loss	0.931
Full model	0.933

Table 2. Ablation study on Sketchy dataset

Input of auxiliary classifier	Precision	mAP
High-level feature	0.996	0.911
Code	0.999	0.933

Table 3. Evaluation of auxiliary classifier on Sketchy dataset

Lingzhi Yu, Xifan Zhang. Cross-Domain Spatial Co-Attention Network for Sketch-Based Image Retrieval[J]. Laser & Optoelectronics Progress, 2022, 59(22): 2215009

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information