Camera conventions & Num objects in the processed GObjaverse

[mengxuyiGit]

Hi Anpei, thanks for the great work and your effort in processing the GOBjaverse!

I have following questions on the processed data:

1. Camera convention: the original GObjaverse seems to use a complex left hand system under camera view / world view, do you transform it to another system so that can be directly used for 2DGS rendering?
2. Normal type: GOBjaverse provides 2 types of normals: ng.exr is the ground truth normal map. on the other hand, nd.exr is the normal map adding bump to enhance the detail of geometry surface. Which one is in your processed data?
3. Dataset size: the original GObjaverse seems to have 3T, while your processed data is about 1.3T. Do you sample a subset from it?

Thanks!

[apchenstu]

Thank you! Regarding the date preprocessing, we use this script to download and convert the format: prepare_dataset_objaverse.

For the size, we only store RGB and normal maps, so it's a little bit smaller than the Gobjaverse.

[mengxuyiGit]

Thanks for prompt reply!

I have downloaded the processed h5 file part 1 from the hub, but the loaded c2w are all 0s:

LaRa/dataLoader/gobjverse.py

Lines 120 to 121 in c32ae77

	def read_cam(self, scene, view_idx):
	c2w = np.array(scene[f'c2w_{view_idx}'], dtype=np.float32)

I have confirmed that the h5 file is not empty and containing 264755 keys, and the idx are valid integers within a reasonable range.

Is there any hint on how to solve this?

[apchenstu]

Oh, if you want to download part of the dataset, you have to download the root file "gobjaverse.h5" and the "gobjaverse_part_xx.h5". Please also refer to gobjaverse.h5 for the dataloader. The gobjaverse.h5 contains camera poses and part_xx contains rendering information.

[mengxuyiGit]

Yes, both are present under the ./dataset. I manually changed the range(30) inside the downloading script.

The eval_all.py works fine on my own processed h5 data. But the downloaded part doesn't.

[mengxuyiGit]

Is it because the dataloader assumes all the meta information contained in "gobjaverse.h5" are available, which is not true when I have downloaded only partial data?

[apchenstu]

Is it because the dataloader assumes all the meta information contained in "gobjaverse.h5" are available, which is not true when I have downloaded only partial data?

Yes, could be, could you please try the first 10 keys of your download data? they are stored in the first part.

[apchenstu]

I just checked there are 8275 samples in part one, please try this scenes_name = np.array(sorted(self.metas.keys()))[:8275] at here .

[mengxuyiGit]

I tried the first 1000 and it works. Appreciated!

[muses0229]

Hi, I preprocess gobj dataset by my own, and I delete all part file and only retain the merged .h5 file, is it ok for training?

[apchenstu]

Yes, all information should already be stored in .h5 after the packing.