Scripts to download ARKitScenes data based on scene IDs from ARKitSceneRefer dataset. Supports chunked downloading to manage large datasets incrementally.
ARKitSceneRefer contains 1,605 unique scenes with 3D object referring expressions. These scripts extract the scene IDs and download the corresponding data from ARKitScenes (either 3DOD or RAW format) in configurable chunks.
- Python 3.12
- Pytorch 2.8
- curl (for downloading)
- OpenRouter API Key
#create workspace
mkdir workspace
cd workspace
#copy project.zip in workspace
cp <path>/project.zip .
#unzip project
unzip projects.zip
# Install dependencies
pip install pandas
pip install pycolmap-cuda12
pip install torchvision
#install segment-anythging
pip install git+https://github.com/facebookresearch/segment-anything.git
#download sam model
curl https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth --output sam_vit_b_01ec64.pth
#install opencv
pip install opencv-python
#install numpy
pip isntall numpy
#install scipy
pip install scipy
#install open3d
pip install open3d
#install gdown
pip install gdown
gdown 1EkrdJhtHZuG1ch9UBjnEkzzuTgMjJI_o
unzip data.zip python run_colmap.py --scene_path data/3dod/Training/40777060When it's finish you will see something like
2025-12-16 03:14:38,071 - INFO - Reconstruction done: 209 images, 2911 points
2025-12-16 03:14:38,073 - INFO - Exporting reconstruction
2025-12-16 03:14:38,087 - INFO - exported point cloud: /workspace/project/data/3dod/Training/40777060/colmap_output/points3D.ply
2025-12-16 03:14:38,099 - INFO - Exported json: /workspace/project/data/3dod/Training/40777060/colmap_output/reconstruction.json
Reconstruction summary
Registered images: 209
3D points: 2911
Observations: 46429
Mean reprojection error: 0.3967 px
2025-12-16 03:14:38,100 - INFO - Finished successfully. output: /workspace/project/data/3dod/Training/40777060/colmap_output python sam_segmenter.py --scene_path data/3dod/Training/40777060 --keyframe_interval 1When it's finish you will see something like
2025-12-16 03:34:24,373 - INFO - Segmenting 40777060_148.061...
2025-12-16 03:34:41,759 - INFO - Segmenting 40777060_151.060...
2025-12-16 03:34:59,257 - INFO - Segmenting 40777060_154.059...
2025-12-16 03:35:28,878 - INFO - Segmenting 40777060_157.057...
2025-12-16 03:35:46,948 - INFO - Segmenting 40777060_160.056...
2025-12-16 03:36:01,151 - INFO - Segmenting 40777060_163.055...
2025-12-16 03:36:17,655 - INFO - Segmented 22 frames
2025-12-16 03:36:17,655 - INFO - Frame 40777060_100.047: 102 objects detected python lift_masks_to_3d.py --scene_path data/3dod/Training/40777060 --min_mask_area 200 --min_points 75 --fuse_radius 0.2 --min_views 2 --max_depth 5.0When it's finish you will see something like
2025-12-16 03:44:15,340 - INFO - Processing 22 SAM mask files
2025-12-16 03:44:18,142 - INFO - Saved 258 fragments
2025-12-16 03:44:20,197 - INFO - Saved 37 fused objects
2025-12-16 03:44:20,218 - INFO - Summary written to /workspace/project/data/3dod/Training/40777060/objects3d/summary.jsonexport OPENROUTER_API_KEY=<YOUR_OPENROUTER_API_KEY>
python run_evaluation.py --data_dir data/3dod/ --max_scenes 1When finished you will see the report
CAPTIONING METHOD COMPARISON
Metric baseline multiview context_aware
Objects 37 37 37
Avg Word Count 86.4 179.7 107.2
Avg Unique Words 63.3 114.6 79.4
Avg Attributes 7.7 12.1 7.6
Detail Score 1.000 1.000 1.000
Hallucination Rate 56.8% 48.6% 54.1%
Label Accuracy N/A N/A N/A
CLIP Score N/A N/A N/A
Consistency Score N/A N/A N/A
SUMMARY:
Multi-view vs Baseline detail score: +0.000
Context-aware vs Baseline detail score: +0.000
#from project go up one level in workspace
cd ..
#clone both repositories ARKitScenes and ARKitSceneRefer
git clone https://github.com/apple/ARKitScenes.git
git clone https://github.com/ku-nlp/ARKitSceneRefer.git
#download more scenes
cd project
python download_3dod_from_refer.py --chunk_index 0
#this will download a chunk of 100 scenes