prunetruong / glu-net Goto Github PK

Hi @PruneTruong ,
Thanks for the great work! I found that DGC-Net can only accept 240x240 image but there are evaluation results of DGC-Net based on the original size of hpatches. Which strategy did you use?

Hi, Thanks for the great work! I have a question about the Hpatches evaluation on the aepe and pck metric. I use the script eval.py in this repo :
python eval.py --pre_trained_models DPED_CityScape_ADE --data_dir path_to_hpatches --hpatches_original_size True
Then it generated a result file named ./evaluation/metrics_GLUNet_HpatchesDataset.txt. I can read the mean value of the aepe( I added it in your code) , pck-1 and pck-5 metric in every scene for 1 to 5, so I dircetly compute the total aepe and pck-1 pck-5, but there is a big difference between the result I got and that in your paper.

from col 1 to col 3 is HP 240x240, from col 4 to col 6 is HP. There is a clear gap in pck-1 but the rest metric including aepe and pck-5 is close. So I was doubt whether I run the code in wrong way.

Hi Prune, thanks for the very nice work!

I think you missed the description of ETH3D testing data in README. Will you update it? And I'm wondering the criterion of identifying a potential pair for testing? As you mentioned in the paper:

This results in about 500 images pairs in total for each selected interval.

Could you also upload the selected pair ids?

Thanks!

Hi, thanks for the great work!
Did you observe any test-time performance difference by training with 1) creating synthetic training and evaluation pairs on the fly (data might be different in each epoch) and 2) save them to disk and load them during training (data are mostly same in each epoch) ?

Thanks for sharing your great work. May you please let me know how long it takes to train on three databases?

Best,

As described in Line 369, GLU-Net.py, the 32*32 flow3=res_flow3 + up_flow4, where up_flow4 is not scaled by the ratio defined in Line 358. However, flow2 and flow1 are all based on the scaled up_flow, which is used in warping. Why are the ratios are used consistent?

Hi @martin-danelljan @PruneTruong

Thanks for the great repository. I am trying to train the GLUNet model for my custom data (training on KITTI) and so wanted to know how to generate the CSV file for any data other than DPED+CS+ADE. Would really appreciate if you could share some information on different columns and picking/finding values for them.

Thanks in advance.

I can't download the model at: https://drive.google.com/open?id=15yXIi8kJbCyXCAHzg-UbMQ6RD2lJ1nOi
Can you give me at email:[email protected]
Thanks!

Hi,

I don't know what can be done wrt this, but you probably should know.
I have requested access to ADE20K and they provided me with a link. However, after unzipping, directory named ADE20K_2021_17_01, not ADE20K_2016_07_26, expected by GLU-Net training appears.
OK, I made a softlink. Then it turns out, that the structure of the dataset has completely changed.
For example, GLU-Net expects following file:

ADE20K_2016_07_26/images/training/l/living_room/ADE_train_00000699.jpg

However, in the current ADE, directory training contains no l, instead the following dirs:
cultural home_or_hotel industrial nature_landscape shopping_and_dining sports_and_leisure transportation unclassified urban work_place.

I believe that homo_aff_tps_test_DPED_CityScape_ADE.csv may need regeneration for the current version of the dataset. Moreover, ADE20K website does not allow to download some kind of older version download.

你他妈的狗比，训练好的模型不加进去？