Towards End-to-End Embroidery Style Generation: A Paired Dataset and Benchmark
Lists
Only you see this
Abstract
Despite the numerous developments in image to image translation, further improvement and applications are limited by monotonous datasets, especially the lack of the datasets with aligned image pairs. In this paper, we present the first dedicated paired embroidery dataset, which contains 9k sets of the input sketches and the corresponding output embroidery images. Specialized on this dataset, we propose a new image-based generative approach as the benchmark for automatically generating a preview of the image produced by an expert on an embroidery machine. The key idea is to segment the input image and use the two-stage generation. The first stage trains a sub-generator function within each segmentation region, while the second stage fine-tunes the output of the first stage with a global generator function. Experimental results demonstrate that our proposed benchmark handles this task best, in terms of both qualitative and quantitative measures.
BibTeX
@inproceedings{ye2021towards,
author = {Jingwen Ye and Yixin Ji and Jie Song and Zunlei Feng and Mingli Song},
title = {{Towards End-to-End Embroidery Style Generation: A Paired Dataset and Benchmark}},
booktitle = {PRCV},
year = {2021},
doi = {10.1007/978-3-030-88013-2\_17},
}Loading related content...