Depth for custom data

[harryzhangzhx]

Hello there, I am just curious if we need to infer the depth map on our own if we are using custom data because when I run train_single.sh error said no depth.npy found. Thanks!

[QianyiWu]

Hi,

Yes. As GSrec needs monocular geometry supervision to regulate the Gaussian's orientation and location, we need to preprocess the dataset for monocular cues similar to the operation in here.

[jiekexiongda]

Hello, I’d like to ask how to obtain _mask.npy in a custom dataset. Here’s the code I’m using: mask_image_path = os.path.join(images_folder, image_name + '_mask.npy').

[QianyiWu]

Hi @jiekexiongda

The mask information can be provided by you own. For example, in replica dataset, we adopt mask to ignore the wrong depth supervision similar to here. If you don't want to apply mask, just generate a all-1 tensor for your task.