Questions about dataset preparation

[Rickyeeeeee]

Dear GaussianRoom Authors,

Thank you for your fantastic work! Our team is currently working on a project that aligns with your work, and we are keen on making a quantitative comparison. Could you please provide us with some imformation about your dataset preparation process?

1. What are the list of scenes that were selected for your experiments from both ScanNet(V2) and ScanNet++?
2. How was the train-test splitting done?
3. Which camera type is used, DLSR or iPhone?
4. Were the model trained and tested using the original image resolution?

Thank you for your time and consideration.

Best regards, Ricky

[xhd0612]

Greetings, thank you for your attention to our work.

1. The selected scenes are scene0050_00, scene0085_00, scene0114_02, scene0580_00, scene0616_00, scene00617_00, scene0721_00, 8b5caf3398, 8d563fc2cc.
2. We take one image as a test for every 10 images.
3. iPhone.
4. We crop the images to 640x480.

All the best, xhd

[Rickyeeeeee]

Thank you for your respond!

I have a couple more questions about the iPhone datasets:

1. Does the training and testing sets include every frames of the video?
2. How was the cropping done? Was it just a simple rescaling?

Best regards, Ricky

[xhd0612]

1. For the ScanNet dataset, we adopted a processing approach similar to Neus-based methods (Monosdf, NeuRIS). The authors of NeuRIS have open-sourced their dataset, which can serve as a reference. For ScanNet++, we set the interval for capturing images to 20 or 15 to keep the number of images within an appropriate range.
2. The processing of the ScanNet dataset also references methods such as NeuRIS. For ScanNet++, we perform center cropping to resize the images to 720x960, as direct rescaling can degrade the performance of monocular normal prediction.
3. By the way, if you want to use the dataset released by NeuRIS, note that the resolution of its normal priors is relatively low, so you might want to re-predict a new one.

Best, xhd

[Rickyeeeeee]

Thank you for providing the useful information! I apologize for reaching out again, but our goal is to replicate the exact dataset used in your paper's experiments. We aim to compare Gaussian Room to our method using quantitative metrics (our method only uses RGB images and camera poses as inputs).

I have a few more questions about the dataset:

1. Which scenes in ScanNet++ use 20 or 15 interval for capturing images?

1. For the ScanNet dataset, we adopted a processing approach similar to Neus-based methods (Monosdf, NeuRIS). The authors of NeuRIS have open-sourced their dataset, which can serve as a reference. For ScanNet++, we set the interval for capturing images to 20 or 15 to keep the number of images within an appropriate range.

2. How is the preprocessing actually done for the ScanNet dataset?

1. For the ScanNet dataset, we adopted a processing approach similar to Neus-based methods (Monosdf, NeuRIS). The authors of NeuRIS have open-sourced their dataset, which can serve as a reference. For ScanNet++, we set the interval for capturing images to 20 or 15 to keep the number of images within an appropriate range.

3. Would it be possible for us to obtain your preprocessing script?

All the best, Ricky

[xhd0612]

Sorry for the late reply. I've been quite busy recently.

1. I checked again, and the interval we used was 10.
2. Our data processing script comes from NeuRIS.

Best wishes xhd

[Rickyeeeeee]

Thank you! I'm curious about how the initial point cloud for the ScanNet dataset is generated?

Best, Ricky