Skip to content

data_preprocess.md

Latest commit

 

History

History
58 lines (43 loc) · 2.33 KB

data_preprocess.md

File metadata and controls

58 lines (43 loc) · 2.33 KB
Raw

ScanNet++

  1. Download the dataset, extract RGB frames and masks from the iPhone data following the official instruction.

  2. Preprocess the data with the following command:

python datasets_preprocess/preprocess_scannetpp.py \
--scannetpp_dir $SCANNETPP_DATA_ROOT\
--output_dir data/scannetpp_processed

the processed data will be saved at ./data/scannetpp_processed

We only use ScanNetpp-V1 (280 scenes in total) to train and validate our SLAM3R models now. ScanNetpp-V2 (906 scenes) is available for potential use, but you may need to modify the scripts for certain scenes in it.

Aria Synthetic Environments

For more details, please refer to the official website

  1. Prepare the codebase and environment
mkdir data/projectaria 
cd data/projectaria
git clone https://github.com/facebookresearch/projectaria_tools.git -b 1.5.7
cd -
conda create -n aria python=3.10
conda activate aria
pip install projectaria-tools'[all]' opencv-python open3d
  1. Get the download-urls file here and place it under ./data/projectaria/projectaria_tools. Then download the ASE dataset:
cd ./data/projectaria/projectaria_tools
python projects/AriaSyntheticEnvironment/aria_synthetic_environments_downloader.py \
--set train \
--scene-ids 0-499 \
--unzip True \
--cdn-file aria_synthetic_environments_dataset_download_urls.json \
--output-dir $SLAM3R_DIR/data/projectaria/ase_raw

We only use the first 500 scenes to train and validate our SLAM3R models now. You can leverage more scenes depending on your resources.

  1. Preprocess the data.
cp ./datasets_preprocess/preprocess_ase.py ./data/projectaria/projectaria_tools/
cd ./data/projectaria
python projectaria_tools/preprocess_ase.py

The processed data will be saved at ./data/projectaria/ase_processed

CO3Dv2

  1. Download the dataset

  2. Preprocess the data with the same script as in DUSt3R, and place the processed data at ./data/co3d_processed. The data consists of 41 categories for training and 10 categories for validation.