Data Preprocessing

October 17, 2025 · View on GitHub

We preprocess ScanNet, ScanNet++ and ARKitScenes following the data processing pipeline provided by VLM-3R. Specifically, for each scene dataset, we extract and store the following components: RGB frames, depth frames, camera parameters (both intrinsics and extrinsics), and axis alignment matrix. Since we only adopt part of the original pipeline and make several modifications to the code, we provide a simplified preprocessing version here for convenience and reproducibility.

After preprocessing, the directory structure should look like this:

GS-Reasoner
|—— data
    |—— processed_data
        |—— scannet
            |—— axis_align_matrix
            |—— color
            |—— depth
            |—— intrinsic
            |—— pose
        |—— scannetpp
        |—— arkitscenes

1. Download Raw Dataset

ScanNet: Download data to data/raw_data/scannet . Follow instructions at https://github.com/ScanNet/ScanNet.
ScanNet++: Download data to data/raw_data/scannetpp and clone repository to data/datasets/scannetpp. Follow instructions at https://github.com/scannetpp/scannetpp.
ARKitScenes: Download data to data/raw_data/arkitscenes and clone repository to data/datasets/ARKitScenes. Follow instructions at https://github.com/apple/ARKitScenes.

2. Extract Data

Extract RGB frames, depth frames, camera parameters (both intrinsics and extrinsics), and axis alignment matrix using following script:

ScanNet

# extract frame info
python scripts/3d/preprocessing/scannet/export_sampled_frames.py \
    --scans_dir data/raw_data/scannet/scans \
    --output_dir data/processed_data/scannet \
    --train_val_splits_path scripts/3d/preprocessing/scannet \
    --num_frames 10000 \
    --max_workers 32 \
    --image_size 480 640

ScanNet++

# install dependency
pip install munch

# extract frame info
# train split
export PYTHONPATH="$PYTHONPATH:./data/datasets/scannetpp"
python scripts/3d/preprocessing/scannetpp/export_iphone_frames.py \
    --data_root data/raw_data/scannetpp/data \
    --scene_list_file data/raw_data/scannetpp/splits/nvs_sem_train.txt \
    --output_dir data/processed_data/scannetpp \
    --num_frames 10000 \
    --num_processes 32 \
    --image_size 480 640 \
    --split_name train

# val split
export PYTHONPATH="$PYTHONPATH:./data/datasets/scannetpp"
python scripts/3d/preprocessing/scannetpp/export_iphone_frames.py \
    --data_root data/raw_data/scannetpp/data \
    --scene_list_file data/raw_data/scannetpp/splits/nvs_sem_val.txt \
    --output_dir data/processed_data/scannetpp \
    --num_frames 10000 \
    --num_processes 32 \
    --image_size 480 640 \
    --split_name val

# generate axis align matrix
# train spilt
python scripts/3d/preprocessing/axis_align_pcd.py \
    --scene_list data/raw_data/scannetpp/splits/nvs_sem_train.txt \
    --split train \
    --output_dir data/processed_data/scannetpp/ \
    --data_type scannetpp

# val split
python scripts/3d/preprocessing/axis_align_pcd.py \
    --scene_list data/raw_data/scannetpp/splits/nvs_sem_val.txt \
    --split val \
    --output_dir data/processed_data/scannetpp/ \
    --data_type scannetpp

ARKitScenes

# extract frame info
# train split
export PYTHONPATH="$PYTHONPATH:./data/datasets/ARKitScenes/threedod/benchmark_scripts"
python scripts/3d/preprocessing/arkitscenes/export_sampled_frames.py \
    --arkit_data_root data/raw_data/arkitscenes/3dod \
    --scene_list_file scripts/3d/preprocessing/arkitscenes/scenes_train.txt \
    --split Training \
    --output_dir data/processed_data/arkitscenes \
    --num_frames 10000 \
    --image_size 480 640 \
    --max_workers 32

# val split
export PYTHONPATH="$PYTHONPATH:./data/datasets/ARKitScenes/threedod/benchmark_scripts"
python scripts/3d/preprocessing/arkitscenes/export_sampled_frames.py \
    --arkit_data_root data/raw_data/arkitscenes/3dod \
    --scene_list_file scripts/3d/preprocessing/arkitscenes/scenes_val.txt \
    --split Validation \
    --output_dir data/processed_data/arkitscenes \
    --num_frames 10000 \
    --image_size 480 640 \
    --max_workers 32

# generate axis align matrix
# train split
python scripts/3d/preprocessing/axis_align_pcd.py \
    --scene_list scripts/3d/preprocessing/arkitscenes/scenes_train.txt \
    --split train \
    --output_dir data/processed_data/arkitscenes/ \
    --data_type arkitscenes

# val split
python scripts/3d/preprocessing/axis_align_pcd.py \
    --scene_list scripts/3d/preprocessing/arkitscenes/scenes_val.txt \
    --split val \
    --output_dir data/processed_data/arkitscenes/ \
    --data_type arkitscenes