casia-iva-lab / isp-reid Goto Github PK

Hi,thanks your nice work.
I reproduce the model based on Occluded_DukeMTMC, and the result is much beterr than the result posted on paper, and I did not adjust any relevant hyperparameters

FileNotFoundError: [Errno 2] No such file or directory: '/data/kzhu/DukeMTMC-reID/train_pseudo_labels-ISP-7'

Thanks for your publication of the code. As you have said in your paper, the clustering is conducted on the images for the same person, respectively. But how to keep the semantic information for the k clusters unchanged during the clustering on different person's images? Intuitively, the clustering on different person's images will generate k clusters with different meanings. At least, the semantic information of k clusters can not keep the same order among the clustering. Waiting for your reply, thx. : )

请问项目中是只有HRnet作为backbone吗？因为我没有发现别的backbone，但是在文章中是有resnet50、seresnet、hrnet的对比，如果项目中有别的backbone希望作者能够指出，不胜感激！

Hello, I think your idea is interesting and effective.But I have one problem. During each clustering process, there will have K clusters.But how can you recognize which one correspond to head and which one correspond to the body?
Looking forward to your reply and code.Thank u.

Your idea is very interesting, but when I try to get the predicted ground-truth for person re-ID datasets on PGT, the code is incorrect, maybe you forget to update your code?

您好，感谢分享论文的代码。您能分享下利用ISP分割后的数据集吗？

Thanks for the amazing implementations.
Some questions regarding reproducing the reported performances. I've been able to obtain really close performances on Market and Duke, however on Occluded Duke, using defalut bash file, I only got around 50.4/58.8(mAP/Rank-1) which is lower than reported 52.3/62.8. I also tried to increase the number of clusters, the performance was droped.
Can you shed some insights on how to improve the performance? Many thanks!

Also why there are two evaluations (with and without arm)?

Hi，My hardware is not very good, Can you provide the model you trained.

PGT文件的生成是用的SCHP代码的哪一个模型？作者将SCHP代码做修改了吗？想了解一下

如题

At line 149 of the engine/trainer.py which is the compute_feature function, I found the detach operation causes memory problem of CPU. I notice you have tried Cuda, but after I use Cuda the memory problem(Cuda memory) is still there. I wanna know how I can fix it.

您好，您的文章中方法部分3.2节中，生成前景图和各个部位的图，是将相同人物身份的所有图像放在一起聚类还是指单个图像进行操作呢，3.2中的公式2和公式3计算得到的结果是3.1节中的Pk做的解释吗？谢谢！

Hello,when I run visualize.sh , it report errors as shown above,so I want to ask why it report errors,thank you!

when i run this code ,the cpu memory will be exploded.my cpu memory is 32G size,

Hi, I cannot download your SCHP predicted ground-truth as I don't have a Baidu account (only for chinese phone number holders if I'm not wrong?), can you share them with me using another medium? (email, Google Drive, ...)
Thank you :-)

Nice work ! It's a very interesting idea , but I can't train with this code . what is the version of ignite？or install it by source?

Thanks for releasing your code!

I have a quick question. Is the provided "pre-trained HRNet32" trained on ImageNet only? i.e. without using any kinds of human pose datasets?

Above image is captured from an official github of HRNet. The pre-trained model you use might be "hrnet_w32-XXX" or "hrnet_w48-XXX" right?

Why are the predicted ground-truth for person re-ID datasets on PGT all black pictures？

Hi, can you provide the performances of ISP-ResNet50 on Market1501, DukeMTMC, Occluded-Duke, and CUHK03-NP?
Thank you very much. I plan to compare our method with ISP-ResNet50.