Can't get correct evaluation results by following instructions

[Wei-TianHao]

Hi, Thanks for your work! I followed the instructions to evaluate on segmentation propagation but failed. I tried to evaluate all folds by executing following commands:

python run_segmentation.py 1 0 python run_segmentation.py 1 1 python run_segmentation.py 1 2 python run_segmentation.py 1 3 python run_segmentation.py 1 4 python eval.py ../data/seg_results/STAGE1_RESULT/ ../data/seg_results/ python eval_view.py ../data/seg_results/STAGE1_RESULT.h5

this is what I have got:

I got the same result for STAGE0.

Thanks for your help!

[varunjampani]

Hi Wei,

Thanks for your interest. Good that you got the same result for STAGE0. So, it seems there is something wrong with the 'caffemodels' that you got by downloading or re-training. Can you please check them? With out further information, it is difficult to find out what is the issue.

Thanks, Varun

[Wei-TianHao]

@varunjampani Sorry, I mean the same as above.

[varunjampani]

Can you check the output of the network? Are they all zeros? And, what about the loaded network weights? Are they also all zeros?

[Wei-TianHao]

I thought there might be problems with gSLICr superpixels. Here is what I got for bear/00000.pgm I rescaled the pixel value with a factor of 5 and it looks like this. Is this normal? Could you please post some superpixels result pictures for reference？ Thank you very much.

[Wei-TianHao]

For the first picture, the pixel value increases from 1 to 44 continuously.

[varunjampani]

I doubt that this is a problem of superpixels as 'STAGE0' (BNN-Identity) worked for you. I think, the above super pixel problem you mentioned is an issue of visualization. Can you please confirm if the STAGE1 caffemodels that you downloaded have around 79MB size?

[Wei-TianHao]

Actually, STAGE0 didn't work for me. I actually meant I got the same result as STAGE1 for STAGE0 when I said I got the same result. I'm sorry if my words made you confused.

[varunjampani]

Sorry that I misunderstood your message. Then, this might be a problem with superpixels. It is difficult to compare with the visualizations. Can you email me a couple of superpixel files (ppm files) so that I can compare them with the superpixel files that I have?

[Wei-TianHao]

Thanks very much! I have sent the files to your email address at tue.mpg.de.

[varunjampani]

There is nothing wrong with the superpixels. I just compared yours with the superpixels I extracted. Problem seem to be somewhere else.

[varunjampani]

Did you find where the problem is?

[Wei-TianHao]

Hi Varun, Thanks for your concern! I'm sorry for forgetting to reply. Actually I didn't find out where the problem is. For other reasons, I have changed my research topic and didn't go deeper into it. Thanks for your help and concern! Really appreciate it! I will close the issue.

Best regard, Tianhao Wei

2017-09-08 21:13 GMT+08:00 Varun Jampani notifications@github.com:

Did you find where the problem is?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/varunjampani/video_prop_networks/issues/2#issuecomment-328099782, or mute the thread https://github.com/notifications/unsubscribe-auth/ARHQNjQA_o2V1KikqUCJFqHrVsqY02lzks5sgT2TgaJpZM4OvKI3 .

-- Best regards, 魏天昊 | Tianhao Wei Phone:(+86)155-5803-6613

[varunjampani]

Hi Wei, Thanks for the information.

[Wei-TianHao]

Hi @varunjampani , I figured out where the problem is! It's related to the cuda version. I degraded the cuda version to 7.5 and got the right results!

[varunjampani]

Thats good to know. Thanks for the information.