search

yuhsuanyeh / BiFuse

[CVPR2020] BiFuse: Monocular 360 Depth Estimation via Bi-Projection Fusion
MIT License
177 stars 29 forks source link

Image size to match original? #3

Open gateway opened 4 years ago

gateway commented 4 years ago

Is their a way that we can have an option or maybe I'm missing it where it doesn't crop/scale down the size? I have images that are 7296 x 3648 which comes out of my Theta Z1 360 camera.

gateway commented 4 years ago

tre1

Test Image I'm using.. equirectangular image.

Thank you for doing this!!

gateway commented 4 years ago

Is this possible, I see in the code this is locked to 1024 x 512 ?

fuenwang commented 4 years ago

@gateway Because our training images are all fixed as 1024x512, so if you apply to other resolution I cannot promise the performance will be good. In you case, I guess you can just upsample the predicted depth map to 7296 x 3648 if you just need reasonable result.

gateway commented 4 years ago

So here is what I have done which I get used to after doing 1000's of 360 panos for my work. I use Topaz Gigapixel AI which isnt free but there are a few neural network upscales available. I have bought the depth map back up to the same resolution of my original pano. Then in photoshop or affinity photo I edit the pano in 3d mode and use photoshops healing/clone tools to fix the zenith/nadir areas that where originally just noise. Its a first step but thats at least a workaround. Then the last step is to run it though Topaz DeNoise (also not free) but a part of my toolbox.