Results on CUHK-SYSU and MSMT17

[chenhao2345]

Thanks for your code. I managed to reproduce the results for Market, Duke and CUHK03. However, I tried to add CUHK-SYSU and MSMT17 into argument --test_dataset, which could not give results mentioned in the paper. The test_mode argument seems to be not working in test_p_s.py, while test_continual_operation_neck.py only support market/duke/mix. Could you please share more details on how to test CUHK-SYSU and MSMT17 datasets?

[hh23333]

@chenhao2345 hi chen, I also meet this issue, have you reproduce the results on CUHK-SYSU and MSMT17?

[chenhao2345]

@hh23333 My re-implementation results on CUHK-SYSU are higher than those in the paper. Results on MSMT17 are lower than those in the paper.

[ddup-gh]

@chenhao2345 Hello chen, I got an error "all query identities do not appear in gallery" when I trained on the dataset CUHK-SYSU. Do you have any idea why that might be? Do I need to make any changes to the file chuksysu.py besides the file structure？Thanks.

[chenhao2345]

@ddup-gh If I understand correctly, it's the problem in the testing code. CUHK-SYSU has only one camera label. But you have to set different camera labels for query and gallery sets. You should modify some code in operation/test_p_s.py.

[ddup-gh]

@chenhao2345 Thank you very much! You're absolutely right, I have solved this problem.

[whisperH]

@ddup-gh If I understand correctly, it's the problem in the testing code. CUHK-SYSU has only one camera label. But you have to set different camera labels for query and gallery sets. You should modify some code in operation/test_p_s.py.

@chenhao2345 hi, Chen. Did you solve the problem later? here is my results of two experiments (mAP/Rank1)

[chenhao2345]

@whisperH No. I got results that corresponded to those in the paper on Market, Duke and CUHK03. But my results on CUHK-SYSU were higher than those in the paper. Results on MSMT17 were lower than those in the paper.

[whisperH]

@chenhao2345 陈博士您好，我又仔细检查了一下代码和超参数，发现我复现的结果和您的结果刚好相反，我的运行的结果显示Market1501、CUHK03比论文精度高、Duke基本持平，CUHK-SYSU和MSMT17 结果反而没有论文中提到的高，大约低了50%，我们有没有可能交换检查一下代码，或者看一下您当时的配置文件或者工程目录呢？

[llzhaoshuo]

@chenhao2345 陈博士您好，我又仔细检查了一下代码和超参数，发现我复现的结果和您的结果刚好相反，我的运行的结果显示Market1501、CUHK03比论文精度高、Duke基本持平，CUHK-SYSU和MSMT17 结果反而没有论文中提到的高，大约低了50%，我们有没有可能交换检查一下代码，或者看一下您当时的配置文件或者工程目录呢？

您好，我想向您请教一下，cuhksysu数据集相关的.mat（Train.mat, TestG50.mat, Person.mat）文件应该从哪里获取呢？

[whisperH]

https://api.bbs.cvmart.net/articles/4753

------------------ 原始邮件 ------------------ 发件人: "TPCD/LifelongReID" @.>; 发送时间: 2022年5月28日(星期六) 晚上8:27 @.>; @.**@.>; 主题: Re: [TPCD/LifelongReID] Results on CUHK-SYSU and MSMT17 (#4)

@chenhao2345 陈博士您好，我又仔细检查了一下代码和超参数，发现我复现的结果和您的结果刚好相反，我的运行的结果显示Market1501、CUHK03比论文精度高、Duke基本持平，CUHK-SYSU和MSMT17 结果反而没有论文中提到的高，大约低了50%，我们有没有可能交换检查一下代码，或者看一下您当时的配置文件或者工程目录呢？

您好，我想向您请教一下，cuhksysu数据集相关的.mat（Train.mat, TestG50.mat, Person.mat）文件应该从哪里获取呢？

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>

[chenhao2345]

@whisperH 你好，这是我得到的结果。除了上面cuhk-sysu的camera问题，我好像没有改什么。如果有需要，可以发邮件交流hao.chen@inria.fr。

[whisperH]

@whisperH 你好，这是我得到的结果。除了上面cuhk-sysu的camera问题，我好像没有改什么。如果有需要，可以发邮件交流hao.chen@inria.fr。

谢谢陈博，我检查过作者代码，感觉还是有些问题的，

初始化过程中没有放到cuda上，forward过程才放到cuda上，这样会导致训练过程grad一直为none，具体可以参考这里，改完之后有一到两个点的提升，但是还是跟作者论文中的结果对不上，实在是不知道啥情况了，感觉很奇怪

[chenhao2345]

@whisperH 好的，谢谢告知。