Laser & Optoelectronics Progress, Volume. 57, Issue 14, 141011(2020)
Pose-Guided Human Image Synthesis Based on Fusion Feature Feedback Mechanism
Figures & Tables(10)
Fig. 3. VGG19 network structure for extracting character image information. (a) Process of extracting characteristic information; (b) extract results of characterization information
Fig. 4. Image effect generated by DeepFashion[20] and Market-1501[21] dataset. (a) Generate a person image from a given pose on DeepFashion dataset; (b) generate a person image from a given pose on Market-1501 dataset; (c) generate a series of person images from a given series of poses on DeepFashion
Fig. 5. Test results on DeepFashion and Market-1501 datasets. (a) Condition images; (b) target pose; (c) target images; (d) G1-L1-VGG19; (e) G1-L1; (f) G1-G2-D; (g) G1-G2-VGG19-D
Fig. 6. Validation results on DeepFashion dataset training model. (a) Condition images; (b) target images; (c) synthesize images
Fig. 7. Validation results on Market-1501 dataset training model. (a) Condition images; (b) target images; (c) synthesize images
Table 1. Comparison of SSIM and IS of different models on different datasets
View tableTable 1. Comparison of SSIM and IS of different models on different datasets
Model DeepFashion Market-1501 SSIM IS SSIM IS MASK-SSIM MASK-IS G1-L1 0.735 2.427 0.304 3.006 0.809 2.455 G1-L1-VGG19 0.638 2.198 0.256 2.258 0.800 3.249 G1-G2-D 0.762 3.090 0.253 3.460 0.792 3.435 G1-G2-VGG19-D 0.795 2.799 0.262 2.911 0.796 3.520
Table 2. Comparison of SSIM and IS with other image generation models on different datasets
View tableTable 2. Comparison of SSIM and IS with other image generation models on different datasets
Model DeepFashion Market-1501 SSIM IS SSIM IS MASK-SSIM MASK-IS Ref. [10] 0.762 3.090 0.253 3.460 0.792 3.435 Ref. [26] 0.614 3.228 0.099 3.483 0.614 3.491 Ref. [27] 0.786 3.087 0.353 3.214 0.787 3.249 Ours 0.795 2.799 0.262 2.911 0.796 3.520
Table 3. Computational complexity of model
View tableTable 3. Computational complexity of model
Dataset DeepFashion Market-1501 Total parameters 8.2×107 3.6×107 Train time per epoch/min 28.92 3.44 Real-data/(sheets·s-1) 4 30
Tools
Get Citation
Copy Citation Text
Youwen Huang, Peng Zhao, Yadong You. Pose-Guided Human Image Synthesis Based on Fusion Feature Feedback Mechanism[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141011
Paper Information
Category: Image Processing
Received: Oct. 16, 2019
Accepted: Dec. 11, 2019
Published Online: Jul. 28, 2020
The Author Email: Peng Zhao (xrinosina163@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20