Request for code of datasets preprocess

[HarshWinterBytes]

Thank you for your excellent work!

I noticed that you used four larger datasets: hypersim, Replica, 3D Ken Burns, and Objaverse, and filtered these datasets before training. I think this filtering operation is crucial for the final training result, so I would like to ask if you can release this part of the code or filtered filename list. This will help a lot.

Waiting for your early reply.
Best wishes!

[Guirassy43]

Hi,

Would it be also possible that you share the processing code that you used for the surface normal estimation test datasets (or even release the processed data)? I saw that you mentioned some filtering/processing to reduce the noise in the GT.

Thank you very much!

[fuxiao0719]

Hi, thanks for your interest!

(1) To HarshWinterBytes: Many of our training datasets are sourced from the Metric3D V2 server, including Hypersim, Replica, and 3D Ken Burns. While it might be challenging to prepare specific items for all training data, we can list the main training components and potential alternatives:

1.Hypersim: Filter out 191 scenes without tilt-shift photography (see here)
2.Replica: Exclude samples with fewer than 50 invalid pixels.
3.3D Ken Burns: All samples
4.Synthetic Urban Dataset (Supp. B.2): These data are owned by DJI Auto and may not be publicly available due to company licensing. However, StableNormal (another great work) uses MatrixCity as an alternative, which you may also consider.
5.Objaverse: Similar to this filter

(2) To Guirassy43: The evaluation surface normal test sets are also sourced from the Metric3D V2 server and have been further manually processed to address "over-smooth" (potentially erroneous) regions with the aid of the company. We will catch up later regarding this part. (similar to (1)-4, restricted by DJI company licensing, the release of data is not as easy as the release of code/weight)

[HarshWinterBytes]

Hi, thanks for your interest!

(1) To HarshWinterBytes: Many of our training datasets are sourced from the Metric3D V2 server, including Hypersim, Replica, and 3D Ken Burns. While it might be challenging to prepare specific items for all training data, we can list the main training components and potential alternatives:

1.Hypersim: Filter out 191 scenes without tilt-shift photography (see here) 2.Replica: Exclude samples with fewer than 50 invalid pixels. 3.3D Ken Burns: All samples 4.Synthetic Urban Dataset (Supp. B.2): These data are owned by DJI Auto and may not be publicly available due to company licensing. However, StableNormal (another great work) uses MatrixCity as an alternative, which you may also consider. 5.Objaverse: Similar to this filter

(2) To Guirassy43: The evaluation surface normal test sets are also sourced from the Metric3D V2 server and have been further manually processed to address "over-smooth" (potentially erroneous) regions with the aid of the company. We will catch up later regarding this part. (similar to (1)-4, restricted by DJI company licensing, the release of data is not as easy as the release of code/weight)

Thank you for your early reply! I'm sorry to bother you again.

I have not worked with simulators before. Right now I'm having trouble with replica. May I ask you how to use replica to generate the depth and normal? If it possible for you to release the code of generation or even the processed data?

And I also noticed that the issue(apple/ml-hypersim#24)) has been solved. The rgb, depth and normal of the example scene in this issue now is aligned. I wonder if it is necessary now to filter out Hypersim and align the normal and depth in the process of dataloader?

Thank you a lot!

[Baijiong-Lin]

the same issue. could you share those .json files or the generated code? thanks.

GeoWizard/geowizard/training/dataloader/mix_loader.py

Line 39 in ee814b7

data_dir = os.path.join(self.data_dir, 'Hypersim', 'annotations', 'annos_all.json')

[fuxiao0719]

It is necessary to align depth and normal during data preprocessing in Hypersim.

For code generation of .json files and replica data, I will reach out to Metric3D v2 team for their initial preparation (as our data are sourced from their server) and see if it is convenient to release.

[HarshWinterBytes]

It is necessary to align depth and normal during data preprocessing in Hypersim.

For code generation of .json files and replica data, I will reach out to Metric3D v2 team for their initial preparation (as our data are sourced from their server) and see if it is convenient to release.

Thank you for your early reply!
Looking forward to your good news!