Got higher PSNR with your uploaded weight file than that in the paper on DAVIS testset. about fastdvdnet HOT 27 CLOSED

And the results on "Set8" is almost the same as your represented results. Should the DAVIS testset be in "png" format?

from fastdvdnet.

I got sigam50: 31.90db on the DAVIS dataset which is downloaded from the link you provided.

from fastdvdnet.

I got sigam50: 31.90db on the DAVIS dataset which is downloaded from the link you provided.

What about sigma 10?

from fastdvdnet.

38.93db in the sigma10

from fastdvdnet.

38.93db in the sigma10

Still 0.22dB higher than the results in the paper. Pretty strange, ha

from fastdvdnet.

老哥也在撸KPN 去躁那些吗哈哈

from fastdvdnet.

I have the same question with the provided model. I use DAVIS-2017-test-dev-480p as the test sequences, and set the maximum number of frames per sequence as 85.
The results of PSNR (average of 30 sequences, and for a single sequence, average of all the frames) are 39.2444dB (sigma 10), 36.1023dB (sigma 20), 34.2924dB (sigma 30), 33.0229dB (sigma 40), 32.0329dB (sigma 50). All of the results are obviousely higher than in the paper.
Can I have any idea of the specific DAVIS test sequences you (I mean the author) use?
BTW, I use the 4 gopro sequences downloaded from your dropbox and also get higher results for each sequence and different sigma values.

from fastdvdnet.

I spent some time trying to figure out why there a difference, but I'm still not sure.
In my davis_480p testset folder, I have 30 sequences, composed of jpgs.
As metioned, I compute the sequence PSNR as an average of the frame PSNRs.
I consider a maximum of 85 frames.
For each frame, I use the function compare_psnr from skimage.measure.simple_metrics, with data_range = 1.0.

You can find my davis_480p testset here (password: fastdvdnet, the link expires on 15 days)

In any case, it is not that bad that you got higher PSNRs than in the paper---the opposite would be very suspicious :)

from fastdvdnet.

Hi, I am reproducing the results of VNLB you present in your paper. As VNLB outputs .tif files, how did you process them by python? I found that the data range in .tif is larger than .png. I turned .tif to .png and the PSNRs are not good.

from fastdvdnet.

Hi Aithusa1113,

Please bear in mind that I computed the results in the paper about three years ago, so the current VNLB version might be different with respect to the version I used. For example, the version I used output .png files, not .tif.
As I side note, maybe clipping the .tif images before saving them as .png helps?

from fastdvdnet.

Hi Aithusa1113,

Please bear in mind that I computed the results in the paper about three years ago, so the current VNLB version might be different with respect to the version I used. For example, the version I used output .png files, not .tif.
As I side note, maybe clipping the .tif images before saving them as .png helps?

Thank you so much for your reply! I'll try clipping and test the results.

from fastdvdnet.

I spent some time trying to figure out why there a difference, but I'm still not sure.
In my davis_480p testset folder, I have 30 sequences, composed of jpgs.
As metioned, I compute the sequence PSNR as an average of the frame PSNRs.
I consider a maximum of 85 frames.
For each frame, I use the function compare_psnr from skimage.measure.simple_metrics, with data_range = 1.0.

You can find my davis_480p testset here (password: fastdvdnet, the link expires on 15 days)

In any case, it is not that bad that you got higher PSNRs than in the paper---the opposite would be very suspicious :)

Can you please reshare the link for the davis_480p testset as it has expired. I am trying to replicate the results, and this dataset('https://davischallenge.org/davis2017/code.html') seems to be different from the one you(author of the paper) used.

from fastdvdnet.

Hi AdithyaK243, I just replied to your email with the link.

from fastdvdnet.

Can you also share the link for davis_480p with me??

from fastdvdnet.

Hey there, here you have the davis_480p saved as jpgs

from fastdvdnet.

Can you also share the link for davis_480p with me?? Email: [email protected]

from fastdvdnet.

@huanzheng551803 just sent you an email

from fastdvdnet.

@m-tassano Can I get DAVIS dataset link
email: [email protected]

from fastdvdnet.

@m-tassano Can you share the DAVIS dataset link with me? My email is [email protected] . Thank you!

from fastdvdnet.

Can you share the DAVIS testset link with me? My email is [email protected]. Thank you!

I spent some time trying to figure out why there a difference, but I'm still not sure. In my davis_480p testset folder, I have 30 sequences, composed of jpgs. As metioned, I compute the sequence PSNR as an average of the frame PSNRs. I consider a maximum of 85 frames. For each frame, I use the function compare_psnr from skimage.measure.simple_metrics, with data_range = 1.0.

You can find my davis_480p testset here (password: fastdvdnet, the link expires on 15 days)

In any case, it is not that bad that you got higher PSNRs than in the paper---the opposite would be very suspicious :)

from fastdvdnet.

@Estherisok666 I added the link in the README

from fastdvdnet.

@m-tassano
I see it. Thank you very much. Could you please share the DAVIS validation set used in your paper?My email is [email protected]

from fastdvdnet.

Please note that the performance of the model trained with the old DALI version (the one originally used to train the weights shared in the Github) has been reported to be superior than the one obtained with the new DALI version.
Or inversely, the new DALI version is linked to a drop in the performance of the model.
See #51 for more details.

from fastdvdnet.

@m-tassano Excuse me，can you also share the link for davis_480p with me （The 2017 DAVIS dataset, trainset）? Email: [[email protected]]

from fastdvdnet.

@wsqat please read the README, you'll find the link there

from fastdvdnet.

@m-tassano ，Sorry to bother again, if I get a model A after training based on the noise model in the paper. Then use this model A as a pre-training model to train and remove other noise models, and then obtain model B. Can this model B remove both kinds of noise.

from fastdvdnet.

@wsqat It'd probably do a decent job at removing both types of noise, but I expect this model to be a bit suboptimal compared to a model trained specifically to remove each type of noise. In any case, what would give the definite answer is doing the experiment.

from fastdvdnet.