prepare_data.md

Prepare data

Create data folder

mkdir data

Colmap

For Colmap, we use the same results as CityGaussian. The following results are sourced from this link.

• Download from Tsinghua Cloud
• Download from Google Drive

Please download and unzip the Colmap results into data/.

Scene partition

We adopt the similar divide-and-conquer strategy as CityGaussian. The following file divide each scene into 8 blocks. Please download and unzip into data/.

• Download from Tsinghua Cloud
• Download from Google Drive

[Optional] If you wish to divide the scene into a different number of blocks, you can follow these instructions (take scene Building as an example):

1. Complete all the subsequent data preparation first before starting scene partition.

2. Train a coarse model for SSIM-based view filtering.

bash script/train/train-building-coarse.sh

3. Divide the scene

bash data_partition.sh <COARSE_MODEL_FOLDER>

4. Modify block_num, partition_name, and block_dim in corresponding training script.

Mill 19 (Building, Rubble)

Please download these two scenes from MegaNeRF and extract them into the data/mill19/ directory. You may refer to the folder structure at the bottom of this page for guidance.

• Dowload Building
• Dowload Rubble

Next, run following script to preprocess images. Please make sure you are in the Momentum-GS/ directory.

bash script/data_preparation/preprocess_mill19.sh

UrbanScene 3D (Residence, Sci-Art)

Please download these two scenes from UrbanScene3D and extract them into the data/urbanscene3d directory. Besides, please download the refined camera poses from MegaNeRF.

• Download Residence Camera Pose
• Download Sci-Art Camera Pose

Next, run the following script to preprocess images.

bash script/data_preparation/preprocess_urbanscene3d.sh

MatrixCity (SmallCity-Aerial)

Please download the small_city-aerial scene from MatrixCity into the data/matrix_city/aerial directory.

Next, run the following script to preprocess images.

bash script/data_preparation/preprocess_matrixcity.sh

Cache images

Following previous methods (e.g. MegaNeRF, VastGaussian, CityGaussian), we downsample all images by a factor of 4 (except for MatrixCity, which will be dawnsampled to 1600*900). This downsampling will be performed during each camera loading process. To enhance efficiency, we will downsample all images in advance.

bash script/data_preparation/downsample.sh

Folder structure

├── data
│   ├── matrix_city
|   |   ├── aerial
|   │   │   ├── train
|   |   |   |   ├── block_all
|   |   │   │   │   ├── data_partitions
|   |   |   |   |   ├── input
|   |   |   |   |   ├── input_cached
|   |   │   │   │   ├── sparse
|   |   │   │   │   │   ├── 0
|   |   │   │   │   │   │   ├── cameras.bin
|   |   │   │   │   │   │   ├── points3D.bin
|   |   │   │   │   │   │   ├── images.bin
|   │   │   ├── test
|   |   |   |   ├── block_all
|   |   │   │   │   ├── input
|   |   │   │   │   ├── input_cached
|   |   │   │   │   ├── sparse
|   |   |   |   │   │   ├── ...
│   ├── mill19
|   │   ├── building-pixsfm
|   │   │   ├── train
|   |   │   │   ├── data_partitions
|   |   │   │   ├── images
|   |   │   │   ├── images_4
|   |   │   │   ├── sparse
|   |   |   │   │   ├── ...
|   |   |   ├── val
|   |   |   │   ├── images
|   |   |   │   ├── images_4
|   |   |   │   ├── sparse
|   |   |   │   │   ├── ...
|   │   ├── rubble-pixsfm
|   │   │   ├── train
|   |   |   |   ... (Similar to building-pixsfm)
|   |   |   ├── val
|   |   |   |   ... 
|   ├── urbanscene3d
|   │   ├── residence
|   │   │   ├── train
|   |   │   │   ├── ...
|   |   │   ├── val
|   |   │   │   ├── ...
|   |   ├── sci-art
|   |   │   ├── train
|   |   │   │   ├── ...
|   |   │   ├── val
|   |   │   │   ├── ...