Ge Teng

icon company Zhejiang University icon location China, Zhejiang, Hangzhou icon location B.Sc. degree, Zhejiang University, Hangzhou, China, 2015-2019. Ph.D. degree, Electronic Information Technology and Instrument Lab, Zhejiang University, 2019-.

Results 4 comments of Ge Teng

您好，我在您给出的数据集链接中下载的USID-D数据集感觉明显是缺失了一部分，导致训练过程中会存在空数据。

你好，我们也遇到过这样的问题，但这个数据集链接是由原作者提供的，我们不方便更改。你可以尝试 - 原作者提供的百度云链接下载，可能会是完整的，https://github.com/nie-lang/UnsupervisedDeepImageStitching - 直接删除不匹配的多余图像

请问您的这个pytorch版本在第一步做mask和warp用到的模型和论文里一样吗？

关于问题：Origin版本用的模型是和原论文一致的，YOLO和Variant版本做了修改，主要包括用Concat替代Cost Volume以及更换Backbone。 关于流程图：原作者在论文中提及，Align阶段的网络沿用自这篇论文 https://arxiv.org/abs/2012.06194 ，你可以在里面找到流程图，Fuse阶段的网络在原论文中就有 https://ieeexplore.ieee.org/abstract/document/9472883

ERROR: Warp Cashed. Results saved in ./tmp/

![图片](https://user-images.githubusercontent.com/35397867/200509512-e546b8a5-7d44-41a2-978e-88fa23d2415a.png) This can happen when the warped mask is all zeros. Please check the "./tmp" directory to find out what went wrong.

ERROR: Warp Cashed. Results saved in ./tmp/

There are multiple causes of this problem, including datasets and training settings. You can check the output of the network during training and testing to find more hints about this...