How to generate cubic B-spline coefficients and confidence masks

[GaryGuTC]

Hi Jinhui,

Thanks for sharing such a good paper, I have some problems after reading the paper when I checked the code:

1. The paper has mentioned that 'cubic B-spline coefficients and confidence masks were pre-computed', I want to ask if could you give me some suggestions about how to generate them? or maybe part of them generated by random.
2. I have checked the 'uv_com_100.npy' file, and I have found that it's 240 * 128, while in the paper it has mentioned that 1080p video will be downsampled by 8 to be 240 * 135. I want to ask why this 128 is different from the 135, maybe some extra steps I lost.

Thanks a lot for your help
Best regards,
Tiancheng

[jhxiong]

Thanks for your interests in the work.

1. The coefficients has also been addressed in another thread (Degree of the B-spline surface #3). Currently I do not have the complete version at my place to generate the coefficients, but I prepared a sample code for your reference, which contains majority part for the pre-computing. Send me an email at jinhui.xiong@kaust.edu.sa and I will attach the python file. The confidence map is computed by comparing the difference between the warped frame and reference frame.
2. The flow computation only supports resolution in a multiplier of 32 or 64 which i do not remember correctly due to downsample and upsample operations in the network, so the original video is cropped to the closest size which can fulfill this multiplier.