JFE output feature channels number

[Eric-chuan]

Hi, well done on your new paper. Can I ask you the channels number of each JFE's output respectively. Or where can I find the supplementary material of the paper to learn about more details of your awesome work.

[feinanshan]

Hi. The output channels for JEFs are 32, 64, 128, and 256 respectively. We are hoping to finish the internal check soon and release the codes.

[Eric-chuan]

As you mentioned in your paper, each JFE module has a two-layer downsampling CNN. The inputs of them are the original feature and the warped feature, then I get the input channels and output cnannels as follows:

• JFE_0: 3+2($I0 concat F{01}$)--------32
• JFE_1: 32+16-----------64
• JFE_2: 64+32----------128
• JFE_3: 128+64----------256 Are these right？I'm having some trouble reproducing it and would be grateful for any help you can provide

[feinanshan]

Hi, the legal check is longer than I expected. The input/output for JFEs is like the following.

• 3+3+2 -> 32
• 32+32+32 -> 64
• 64+64+64 -> 128
• 128+128+128 ->256

[Eric-chuan]

Okay, thank you for your warm answer. Do you have an expected date for the code release?

[feinanshan]

We are trying to work it out before July, but no guarantee for that.

[Eric-chuan]

Get it, thank you so much

[Eric-chuan]

Hi， I have reproduced a version of the M2MModel based on your instructions. currently I am using only laplacian loss for training and have not added the LFM module and can get a PSNR of 34.8 on Vimeo90k. but there is a problem, I find that the estimated multiple optical flows are sparse (looks like a bunch of spots), can you provide any helpful advice? thanks

[feinanshan]

Hi, codes are released.