search

NUST-Machine-Intelligence-Laboratory / SSC

7 stars 1 forks source link

I'd like to ask you some questions about the project #1

Open ajie220209 opened 7 months ago

ajie220209 commented 7 months ago

When I run the run_sample file, it tells me I don't have enough GPU memory, so I try to delete some of the VOC2012 images, and it shows me FileNotFoundError: No such file: 'D:\PycharmProjects\SSC-main\data\VOC2012\JPEGImages\2011_001789.jpg'. I just want to ask if I can continue to complete the code without deleting some data sets. I can't find it either. Why do I have to look for 2011_001789? Shouldn't it be under the existing data set?

xixiaos commented 6 months ago

Hello author, thank you for your contribution to this article. I would like to ask you a question. When I reproduced your code on A6000, the quality of the pseudomask was only 69%, and the result of v3+segmentation was only 69%. However, the quality of the pseudomask in the article was 71.9%, and the result of v2 segmentation was 72.7%. Can you explain this? Thank you

chentao2016 commented 5 months ago

When I run the run_sample file, it tells me I don't have enough GPU memory, so I try to delete some of the VOC2012 images, and it shows me FileNotFoundError: No such file: 'D:\PycharmProjects\SSC-main\data\VOC2012\JPEGImages\2011_001789.jpg'. I just want to ask if I can continue to complete the code without deleting some data sets. I can't find it either. Why do I have to look for 2011_001789? Shouldn't it be under the existing data set?

If you don't have enough GPU memory, you should change the batch size to a smaller value rather than delete some of the VOC2012 images. Sorry for the late reply.

chentao2016 commented 5 months ago

Hello author, thank you for your contribution to this article. I would like to ask you a question. When I reproduced your code on A6000, the quality of the pseudomask was only 69%, and the result of v3+segmentation was only 69%. However, the quality of the pseudomask in the article was 71.9%, and the result of v2 segmentation was 72.7%. Can you explain this? Thank you

Please first check the quality of seeds before irn. Besides, please check whether you can reproduce the result of BECO, your result of v3+segmentation 69 is far from the baseline result of BECO 72.1.

xixiaos commented 4 months ago

Hello author, I would like to ask if your model is not very stable. The seed quality I reproduced is only 56.18, while the result in your article is 58.3

chentao2016 commented 4 months ago

@xixiaos Thanks for your interest. During our experiments, our model with 3090GPU is quite stable. Maybe you can run it multiple times to check whether you can obtain better seeds. Sorry for the inconvenience.

xixiaos commented 4 months ago

Hello author, thank you for your reply. I have reproduced the quality of the best pseudo mask many times on RTX6000 (i.e. after IRN) to be 70.18, and then used this pseudo mask to segment v2, resulting in a segmentation result of 66.25. I am not sure if the code you uploaded is from one of your ablation experiments, as I have never been able to reach the indicator in the paper

chentao2016 commented 4 months ago

Hello author, thank you for your reply. I have reproduced the quality of the best pseudo mask many times on RTX6000 (i.e. after IRN) to be 70.18, and then used this pseudo mask to segment v2, resulting in a segmentation result of 66.25. I am not sure if the code you uploaded is from one of your ablation experiments, as I have never been able to reach the indicator in the paper

For the segmentation result, the baseline is 72.1 (BECO). Maybe you can directly refer to the original segmentation code from here: https://github.com/ShenghaiRong/BECO.

chentao2016 commented 4 months ago

Hello author, thank you for your reply. I have reproduced the quality of the best pseudo mask many times on RTX6000 (i.e. after IRN) to be 70.18, and then used this pseudo mask to segment v2, resulting in a segmentation result of 66.25. I am not sure if the code you uploaded is from one of your ablation experiments, as I have never been able to reach the indicator in the paper

You can aslo leave your wechat here, I will contact you for more convenient disscussion.

xixiaos commented 4 months ago

I'm not sure if my BS settings are incorrect. How many cards are you using

chentao2016 commented 4 months ago

I'm not sure if my BS settings are incorrect. How many cards are you using

I use 2 3090GPU. Honestly, with BECO, even if very poor seed from original irn can reach quite decent performance if you check the paper of BECO.

qingjob commented 4 months ago

您好作者，感谢您的回复。我在 RTX6000 上多次（即 IRN 之后）将最佳伪掩码的质量重现为 70.18，然后使用此伪掩码对 v2 进行分割，得到 66.25 的分割结果。我不确定您上传的代码是否来自您的一个消融实验，因为我从未能够达到论文中的指标

您好作者，感谢您的回复。我在 RTX6000 上多次（即 IRN 之后）将最佳伪掩码的质量重现为 70.18，然后使用此伪掩码对 v2 进行分割，得到 66.25 的分割结果。我不确定您上传的代码是否来自您的一个消融实验，因为我从未能够达到论文中的指标

Hello, the result I reproduce is similar to yours, it is 65.34, have you solved this problem? I'd imagine you asking for a solution. Thank you!

xixiaos commented 3 months ago

您好作者，感谢您的回复。我在 RTX6000 上多次（即 IRN 之后）将最佳伪掩码的质量重现为 70.18，然后使用此伪掩码对 v2 进行分割，得到 66.25 的分割结果。我不确定您上传的代码是否来自您的一个消融实验，因为我从未能够达到论文中的指标

您好作者，感谢您的回复。我在 RTX6000 上多次（即 IRN 之后）将最佳伪掩码的质量重现为 70.18，然后使用此伪掩码对 v2 进行分割，得到 66.25 的分割结果。我不确定您上传的代码是否来自您的一个消融实验，因为我从未能够达到论文中的指标

Hello, the result I reproduce is similar to yours, it is 65.34, have you solved this problem? I'd imagine you asking for a solution. Thank you!

The author of this article used the Beco method when running segmentation. You can refer to what the author mentioned in README.md and then run the segmentation to see the effect.