panpanfei / dual-pixel-exploration-simultaneous-depth-estimation-and-image-restoration Goto Github PK

I want to use the simulator for making synthetic DP datasets.
But I have a problem on setting parameters.
What are meanings and units of parameters in the code?
I wonder that the Fd means the F value referred in the paper and meanings of the f_pix and f in the code.

Thank you

I noticed that the reported results (25.4136 / 0.8394 in PSNR/SSIM) are evaluated with the 8-bit mode in the DPDBlur. Since the image mode of DPDBlur is 16-bit, I think the reported results with 8-bit mode should be highlighted in ReadMe.md.

In the simulator_dp_nyu_testingdata.m and simulator_dp_nyu_trainingdata.m, there's a line:

[img_left,img_right,disp_left] = generatedpimage(RGB_img,disp);

which gives an error "Too many output arguments.". In simulator_dp_main.m, the line is like this

[img_left,img_right] = generatedpimage(RGB_img,disp);

Could you clarify if generatedpimage.p was updated to return 3 values, but not updated in github? or what is the problem?

Thank you!

May I ask what image resolution is used for DPD-blur testset? Thank you very much!

Originally posted by @liyucs in #1 (comment)

Hi,
When the code will be released?

Hello,

would you please share a readable code for this part? generatedpimage?

thank you!

Hi, thanks for your inspired work, I have some issues about the experiments in the paper.
In Table 1, the result of the DPDNet is reported as 25.53 in DPD-blur dataset. However, the original paper given 25.13. So is the SSIM(0.826/0.786). Can you explain it? Thank you!

Hi and thanks for the interesting work and code,

Could I ask for the fine-tuned model using reblur-loss for DP-disp results (figure 6 + table 2)

Thanks,
Sagi Monin

Hi, Thanks for your interesting work.
When the code will be released?

Hi,

In your implementation of simulator, the predicted depth map is used to calculate the positions of four boundary points. Then, the positions are used in index operation to search for certain regions of the integral image. However, the index operation does not support backpropagation in general, so how could you supervise your predicted depth map in simulator?

Best Regards

Hi, thanks for your inspired work, I have some issues about the experiments in the paper.
In Table 1, the result of the DPDNet is reported as 25.53 in the DPD-blur dataset. However, the original paper given 25.13. So is the SSIM(0.826/0.786). Can you explain it? Thank you!