Rain Rendering on custom dataset

[LeR75019]

Hi, I would like to use your Rain Rendering research for my PhD research in domain adaptation on a custom dataset.

I extracted the depth maps from Midas v3.1, which are 3 channels scaled between 0 and 1 and dtype = float32.

You wrote : # Depth images (16 bits, with depth_in_meter = depth/256.)

So am I supposed to do something like :

depth = depth.mean(0) # [3, H, W] to [H, W]
focal_length, baseline = 2262, 0.22
depth = (baseline * focal_length) / (depth*2048)
depth_enc = np.minimum(depth * 256., 2 ** 16 - 1).astype(np.uint16)

And just use the final depth_enc as depth labels or just feed the model the original depth maps ?

Thanks !

[rdecharette]

hey,

so sorry your message, went below my radar. Yes, you should encode it as you wrote depth_enc. It seems valid at first sight, let me know if you're still stuck.

Best.