cydiachen

> @cydiachen I have the idea to reproduce the network on MMDet. I hope we can reproduce it together hello, I have contacted you on Zhihu. Feel free to contact...

这篇论文稍微存在一些复现性问题，我这里复现出来的效果只能达到大约PSNR：22-23的样子，我发出来给大家一起跑跑看 ，看看这个分数的gap在哪里。我最近也有把这个程序dataloader重写一下，有空的话我整理一下发出来。随时欢迎提issue

Dear friend, This code is originally altered from Wavelet-SR base code, you can refer to that code and copy the dataloader from the base code and simply replace train batches...

THX. I am uploading the pretrained weight and I rewrite all the dataloader for you. You can download the NEW_VERSION.

> Thanks @cydiachen for the update. I'll check all and come back to you. > Just a question : in your approach, the HourGlass is generating parsing maps and Landmarks...

> 帅哥，有预训练模型吗？可以提供一个预训练模型嘛。 我现在使用的pretrain的model的gt 图像数目比较少，请问你有没有相关含有比较清晰parsing map的数据集可以参考？我现在采用的是CelebAHD+ 使用GAN生成的对应Parsing map的数据集，数据量级在15000左右，和原作者的差一些

> 我现在就有Hellen的人脸解析图>_

THX. I am uploading the pretrained weight and I rewrite all the dataloader for you. You can download the NEW_VERSION.

THX. I am uploading the pretrained weight and I rewrite all the dataloader for you. You can download the NEW_VERSION.

TODO: 1. Not yet tested on Visdrone, need to debug; 2. Visdrone has inherent label for crowd object and occluded objects, we should consider imtroduces these properties; 3. Dataset definition...