Towards Hard-pose Virtual Try-on via 3D-aware Global Correspondence Learning

Authors: Zaiyu Huang, Hanhui Li, Zhenyu Xie, Michael Kampffmeyer, qingling Cai, Xiaodan Liang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on public benchmarks and our Hard Pose test set demonstrate the superiority of our method against the SOTA try-on approaches. 4 Experiments In this section, we conduct extensive experiments to validate the effectiveness of the proposed 3D-GCL network.
Researcher Affiliation Collaboration 1Shenzhen Campus of Sun Yat-Sen University 2Byte Dance, 3 Ui T The Arctic University of Norway 4Peng Cheng Laboratory {huangzy225, xiezhy6} @mail2.sysu.edu.cn {lihh77, caiqingl} @mail.sysu.edu.cn michael.c.kampffmeyer@uit.no, xdliang328@gmail.com
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code Yes Our code and data are provided in the supplementary material.
Open Datasets Yes Our experiments are conducted on two open-source datasets, Deep Fashion [31] and MPV [32], which contain 52,712 and 37,723 fashion images, respectively.
Dataset Splits No To ensure a fair comparison, we follow the train/test split of [20, 22, 33] on the Deep Fashion dataset and use the original split of MPV. In this way, we get 101,622/8,564 train/test pairs for Deep Fashion and 52,236/10,544 train/test pairs for MPV. The paper specifies train/test splits but does not explicitly mention a separate validation split with quantities or percentages.
Hardware Specification Yes The proposed 3D-GCL network is implemented in Py Torch and trained with 4 Tesla V100 GPUs.
Software Dependencies No The proposed 3D-GCL network is implemented in Py Torch and trained with 4 Tesla V100 GPUs. We thank Mind Spore for the partial support of this work, which is a new deep learning computing framwork5. The paper mentions PyTorch and MindSpore but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes We first train the correspondence estimation subnetwork for 20 epochs with a batch-size of 8, and then follow the settings of [20, 25] to train the try-on generator.