search

ChrisChen1023 / HINT

HINT: High-quality INpainting Transformer with Enhanced Attention and Mask-aware Encoding
MIT License
22 stars 3 forks source link

about metrics #16

Open Marm7 opened 2 weeks ago

Marm7 commented 2 weeks ago

您好，我想请问一下，我是用敦煌预训练模型测试敦煌数据集时，得到的评价指标与论文中有些差距，可能是我的某些操作有问题

图一最后一行是我的复现结果 image

同时，我也测试了MISF在敦煌数据集的指标（图二），同样与图一有很大差距 image

上述测试结果使用scikit-image==0.18.2，与MISF的版本相同

所以我想请问一下，是否我有没注意到的设置或操作导致结果产生了偏差 期待回复！Thanks！！

ChrisChen1023 commented 2 weeks ago

Hi,

The issue happens to me before. But the reason could be different. You may wish to check the image size (after dataloder). Also, for dunhuang dataset, the mask threshold need to be 100, which is followed by MISF.

Marm7 commented 2 weeks ago

感谢您的回复，我dataloder后的size均保持在256*256。将mask threshold设置为了100，得到的指标与论文中的更加接近，但也似乎仍在允许的误差之外 image

ChrisChen1023 commented 6 days ago

Hi,

This issue does not happen to me. You could check are the images cropped before the dataloader. Others I guess it might be because of the environment setting.