Questions of the depth data

[moyutianque]

Here is a calculation (depth_gt * 50 + 0.5) before depth loss in mono-prior. Does anyone know what is 50 and 0.5 in the following line. Isn't the depth value stored in meter?

link

[niujinshuchong]

Hi, the monocular depth is up to scale. Please refer to autonomousvision/monosdf#18. The value is chosen based on the replica dataset which might not work well for other datasets.

[moyutianque]

thanks for the information, could I know how this value was designed? I tried depth extracted from omnidata with script link in sdfstudio on my dataset collected from matterport simulator. But it does not perform well, it will grateful if you can give suggestions on how to tune these values.

PS: I verify the collected data on nerfstudio instant-ngp, and it performs well.

[niujinshuchong]

Hi, I adjusted the scale and shift based the omnidata model's output such that the range is in [0, 2]. But it also interleaves with monocular depth loss weight in the loss function. You might need to try different values. Can you share some of your results? Did you also try monocular normal loss?

[moyutianque]

Thanks, I try to change the scale of my poses in nerfstudio transforms.json to match meta_data.json scale, it performs well now. But I found in process_nerfstudio_to_sdfstudio.py the scale for indoor scene is set to 5, is there any insight of this number? I tried args.scene_scale_mult=2 which makes poses more sparse within the scene-box the performance drops 4 psnr in the end.

[niujinshuchong]

Hi, the value is chosen based on a very specific scene so it might not work well for your data. You need to normalize your poses such that the main object is inside the unit cube. Maybe you could also try this one: #90. Btw, if you are using monocular prior, you could try to enable only normal loss or only depth loss to compare the differences. Normal is usually more robust.

[moyutianque]

OK thanks.