search

JuewenPeng / BokehMe

BokehMe: When Neural Rendering Meets Classical Rendering (CVPR 2022 Oral)
Apache License 2.0
181 stars 9 forks source link

blur degree #9

Closed C-water closed 2 months ago

C-water commented 5 months ago

Hi, juewen! Thank you for the BLB and training datasets.

I tried to train a CNN network to generate bokeh images in an end-to-end manner. When K=10 or 20 (the training set and the BLB are equally blur), the output is normal. However, when K increases from 30 to 50, the model cannot learn the corresponding degree of blur from the input defocus map. How did you solve this problem when training neural render (in Bokehme)?

Thank you very much!!!

JuewenPeng commented 5 months ago

With the increase of K, it will be much harder for the neural renderer to learn, since it needs to deal with a wider range of blur amounts. You can try to increase the capacity of the network or modify the architecture. However, from my experience, constraining the blur range of ARNet to a low level (eg. 0-10 or 0-12), and iteratively increase the blur range by IUNet is more efficient and effective.

C-water commented 5 months ago

Thank you for your reply.

I tried to increase the model capacity and the blur effect seemed to increase, along with artifacts (difficult to learn the defocus blur of unknown degree). As you suggested, I will adopt the iteratively up-sampling structure, such as IUNet, to increase the blur range.

Thank you again!!!