Abstract: In order to improve the accuracy of cloth landmark estimation, a method with feature-guided attention module is proposed. Drawing on the idea of depthwise separable convolution, the attention module is constructed, which not only enhances the spatial features in network, but also strengthens the information interaction between different feature channels. Then, it is put into each stage of the HRNet, so the spatial information of the input features could be modelled in a more granular way. Secondly, unbiased data processing is used, which converts the data from discrete space to continuous space, to reduce quantization error introduced by data argumentation. Finally, a coarse-to-fine training strategy is adopted which further reduce heavy computing costs and improve accuracy. The proposed method achieves the state-of-the-art result with 67.4% accuracy in DeepFashion2 dataset cloth landmark estimation task.