EIT-NLP/GeoRG-67K-CAT

GeoGR-67K-CAT

This repository contains the data used in GeoGR: Spatial–Semantic Token Compression and Group Routing for Efficient 3D VLMs.

The release includes ScanNet scene data and derived 3D representations, EmbodiedScan scene metadata, geometry and frame-selection metadata, benchmark annotations for five downstream 3D-language tasks, and the processed training data used in our experiments.

The scene representation and data preprocessing follow the general pipeline of Video-3D-LLM.

Dataset Organization

After extracting the ScanNet pose data, the expected directory structure is:

GeoGR-67K-CAT/
├── scannet/
│   ├── scans/
│   ├── pose/
│   ├── pcd_with_object_aabbs/
│   └── mask/
│
├── embodiedscan/
│   ├── embodiedscan_infos_train.pkl
│   ├── embodiedscan_infos_val.pkl
│   └── embodiedscan_infos_test.pkl
│
├── metadata/
│   ├── scannet_select_frames.json
│   ├── pcd_discrete_0.1.pkl
│   ├── scannet_train_gt_box.json
│   ├── scannet_val_gt_box.json
│   ├── scannet_val_pred_box.json
│   └── scannet_test_pred_box.json
│
├── benchmark/
│   ├── Sqa3D/
│   ├── ScanQA/
│   ├── ScanRefer/
│   ├── Scan2Cap/
│   └── Multi3DRefer/
│
└── processed/
    ├── full/
    └── subsets/

ScanNet Data

The scannet/ directory contains the ScanNet scene data and derived scene representations used in GeoGR.

Scans

scannet/scans/ contains the original ScanNet v2 scene files. Each scene follows the standard ScanNet organization:

scannet/
└── scans/
    └── sceneXXXX_XX/
        ├── sceneXXXX_XX_vh_clean_2.ply
        ├── sceneXXXX_XX_vh_clean_2.0.010000.segs.json
        ├── sceneXXXX_XX.aggregation.json
        ├── sceneXXXX_XX.txt
        └── sceneXXXX_XX.sens

The original ScanNet v2 data can be obtained from the official ScanNet website.

Pose

scannet/pose/ contains the preprocessed multi-view observations for each scene, including RGB images, depth information, and camera poses.

The pose data are distributed as five split archive files (posed_images_part_aa to posed_images_part_ae). After downloading, merge and extract them into scannet/pose/ with:

cd scannet && cat posed_images_part_a{a,b,c,d,e} > posed_images.tar.gz && mkdir -p pose && tar -xzf posed_images.tar.gz -C pose && rm posed_images.tar.gz

Point Clouds and Object Boxes

scannet/pcd_with_object_aabbs/ contains processed scene point clouds together with object-level axis-aligned 3D bounding boxes.

Object Masks

scannet/mask/ contains the object-mask and proposal information used by the scene processing and evaluation pipeline.

The corresponding ScanNet preprocessing scripts are available in the Video-3D-LLM 3D preprocessing directory.

EmbodiedScan Metadata

The embodiedscan/ directory contains the scene-information files used to associate multi-view observations with their corresponding 3D scenes.

embodiedscan/
├── embodiedscan_infos_train.pkl
├── embodiedscan_infos_val.pkl
└── embodiedscan_infos_test.pkl

The original EmbodiedScan data can be obtained from the official EmbodiedScan repository. Access may require completing the corresponding data request procedure.

Metadata

The metadata/ directory contains auxiliary geometry, frame-selection, and object-box information used by the GeoGR data pipeline.

metadata/
├── scannet_select_frames.json
├── pcd_discrete_0.1.pkl
├── scannet_train_gt_box.json
├── scannet_val_gt_box.json
├── scannet_val_pred_box.json
└── scannet_test_pred_box.json

scannet_select_frames.json stores the selected frames for each ScanNet scene.

pcd_discrete_0.1.pkl contains discretized point-cloud information used during geometry-aware scene processing and frame selection.

The remaining JSON files contain ground-truth or predicted 3D object-box information for the corresponding ScanNet splits.

Benchmark Annotations

The benchmark/ directory documents or stores the source annotations used to construct the processed files. The processed annotations use the Video-3D-LLM LLaVA-style record format, for example: json { "id": "...", "video": "scannet/scene0249_00", "conversations": [ {"from": "human", "value": "<image> ..."}, {"from": "gpt", "value": "..."} ], "metadata": {"dataset": "scanqa"} } The video value is a scene key. It is resolved through the EmbodiedScan scene metadata and does not point to a standalone video file.

benchmark/
├── Sqa3D/
├── ScanQA/
├── ScanRefer/
├── Scan2Cap/
└── Multi3DRefer/

Benchmark	Task	Original Project
SQA3D	Situated 3D question answering	SQA3D
ScanQA	3D question answering	ScanQA
ScanRefer	Single-object referring and grounding	ScanRefer
Scan2Cap	3D object captioning	Scan2Cap
Multi3DRefer	Multi-object referring and grounding	Multi3DRefer

Processed Annotations

The processed/ directory contains annotations converted into the unified LLaVA-style format used by the GeoGR training and evaluation pipeline.

processed/
├── full/
└── subsets/

A processed record follows the general structure:

{
  "id": "...",
  "video": "scannet/scene0249_00",
  "conversations": [
    {
      "from": "human",
      "value": "<image> ..."
    },
    {
      "from": "gpt",
      "value": "..."
    }
  ],
  "metadata": {
    "dataset": "scanqa"
  }
}

The video field represents the corresponding 3D scene key rather than a standalone video file.

Full Annotations

processed/full/ contains the complete processed annotations for the five downstream tasks, including the available training and evaluation splits.

processed/
└── full/
    ├── sqa3d_*_llava_style.json
    ├── scanqa_*_llava_style.json
    ├── scan2cap_*_llava_style.json
    ├── scanrefer_vg_*_llava_style.json
    └── multi3drefer_*_llava_style.json

GeoGR Training Subsets

processed/subsets/ contains the PRISM-selected training subsets used in the GeoGR post-training experiments.

The subset selection only changes which annotation records are used for training. It does not change the RGB images, depth maps, camera poses, point clouds, scene geometry, or scene IDs.

processed/
└── subsets/
    ├── sqa3d_train_llava_style.json
    ├── scanqa_train_llava_style.json
    ├── scan2cap_train_llava_style.json
    ├── scanrefer_vg_train_llava_style.json
    └── multi3drefer_train_llava_style.json

The released subsets contain approximately 30% of the corresponding full training annotations.

Task	Training Records
ScanRefer	11,000
Multi3DRefer	13,151
Scan2Cap	11,000
ScanQA	7,955
SQA3D	23,834
Total	66,940

License and Attribution

This repository is a derived data release and does not supersede or modify the licenses and access conditions of the original datasets.

Users should follow the original terms of ScanNet, EmbodiedScan, SQA3D, ScanQA, ScanRefer, Scan2Cap, and Multi3DRefer when using or redistributing the corresponding data.

Please retain the original dataset attribution and license notices when creating mirrors or derivative releases.