返回
Virtual try-on based on attention U-Net
DOI:10.1007/s00371-022-02563-6.png)
摘要
En 中文
Image-based virtual try-on, aiming to fit new in-shop clothes into a person image, has gained extensive attention in the fields of computer vision and image process community. Most of the current virtual try-on methods are based on thin plate spline transformation and composition mask. Such methods are limited by the spatial feature retention performance of the network, hard to warp clothes aligning with new body when the body shape and posture change largely, and its difficult for them to handle self-occlusion. We employ a two-stage approach, warping clothes in the clothes warping module (CWM), and generating try-on results in cross-domain fusion module (CFM). To address the problem of hard to warping clothes aligning with new body, we add a combined loss to the Clothes Warping Module, including a perceptual loss for the clothes parsing region and a L1 loss for the whole image. To solve the self-occlusion problem, firstly, we adopt an attention-based U-Net network as the backbone of the cross-domain fusion module. Then, we improved the framework of CFM to generate composition mask, adjusted clothes and rendered person, and composite the final try-on result through mask operations. Experiments on the Zalando dataset demonstrate that this work can warp clothes naturally with details preserved and produce photo-realistic try-on results without self-occlusion.
Keyword:
Virtual try-on
Cascaded attention mechanism
Cross-domain fusion
Thin plate spline
Reduce occlusion
期刊
IF:
2.9
论文数:
4.6K
被引数:
6.5K
机构
引用论文
Caw’s Walking State Recognition Based on Accelerometers and Gyroscopes Installed on Ear-Tags and Collar-Tags基于安装在耳标和项圈上的加速度计和陀螺仪的Caw步行状态识别
The RNA Chaperone Hfq Is Involved in Colony Morphology, Nutrient Utilization and Oxidative and Envelope Stress Response in Vibrio alginolyticus
PLOS ONE
IF0

