[ICML 2026 Oral] Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence

Todo List

• We release the project page of Holi-Spatial, check on here.
• We have released a subset of the Holi-Spatial dataset, featuring 2,000+ high-quality Gaussian models, each accompanied by its corresponding mesh and scene-level 3D bounding box annotations, together with HoliSpatial-QA-2M, a 2-million-scale QA dataset built upon part of these richly annotated 3D assets.
• We release all model checkpoints.
• We will release the pipeline code in July.

Holi-Spatial Pipeline

This repository contains the data curation pipeline used to turn ScanNet, ScanNet++, or DL3DV scenes into 3DGS geometry, mesh-guided masks, object/region 3D annotations, spatial QA, and LLaMA-Factory training data.

The pipeline follows the three stages described in the Holi-Spatial paper:

1. Geometric optimization: DA3 depth/point cloud initialization, then PGSR/3DGS training.
2. Image-level perception: VLM class or region discovery, then SAM3 mask generation.
3. Scene-level lift and refinement: lift masks into 3D, merge/filter boxes, caption instances, and synthesize spatial QA.

Paper: https://arxiv.org/abs/2603.07660

Repository Layout

• inference_da3_scannetppv2.py, run_da3.sh: unified DA3 preprocessing for ScanNet, ScanNet++, and DL3DV.
• PGSR/: PGSR training, rendering, and mesh-to-mask code used by this pipeline.
• 3dgs_train.sh: batch PGSR/3DGS training entrypoint for ScanNet, ScanNet++, and DL3DV.
• mesh.sh: render meshes and generate mesh-guided masks for ScanNet v2, ScanNet++, and DL3DV.
• classic_vllm.py, classic_region.py: VLM-based object and functional-region label discovery.
• sam3.py: SAM3 text-prompted mask generation.
• 3d_bounding_instance_gs_rerun_da3.py: object 2D-to-3D lifting/refinement.
• 3d_bounding_instance_gs_region.py: functional-region 2D-to-3D lifting/refinement.
• postprocess_3d_bbox_aabb.py: floor-aligned AABB-style box post-processing before QA generation.
• scannetppv2_new.sh, scannetppv2_region.sh, scannetv2_all.sh, scannetv2_region.sh, DL3DV_new.sh, dl3dv_new_2k.sh: dataset-specific annotation batch scripts.
• qa_generation/: instance captions, spatial QA generation, QA filtering, and LLaMA-Factory conversion.

Setup

Create a CUDA-enabled Python environment, then install the common dependencies:

pip install -r requirements.txt
pip install -e PGSR/submodules/diff-plane-rasterization
pip install -e PGSR/submodules/simple-knn

Install or expose the external model packages required by the pipeline:

• Depth Anything 3: provides depth_anything_3.api.DepthAnything3.
• SAM3 checkpoint/assets: sam3.py downloads the default checkpoint and BPE asset from facebook/sam3 when local paths are not provided.
• A VLM served through an OpenAI-compatible vLLM endpoint. All scripts default to http://localhost:8000/v1 or http://localhost:8000/v1/chat/completions.

Expected Data Layout

ScanNet v2:

scannetv2/scans/<scene_id>/
  color/
  pose/
  intrinsic/

ScanNet++:

scannetppv2/data/<scene_id>/
  dslr/nerfstudio/transforms_undistorted.json
  dslr/resized_undistorted_images/

DL3DV:

processed_dl3dv_ours/1K/<scene_id>/
  dense/cam/
  dense/rgb/

The QA stage also needs a covisibility root, for example scannetppv2_wai/<scene_id>/covisibility/v0/*.npy or dl3dv_wai/1K_<scene_id>/*.npy.

End-to-End Pipeline

1. DA3 Depth and Point Cloud

Process a directory of scenes:

SCENE_ROOT=scannetppv2/data bash run_da3.sh

Process a single scene:

SCENE_DIR=scannetppv2/data/0a5c013435 bash run_da3.sh

Outputs per scene:

• depth_da3/<image_stem>.npy
• pointcloud_da3.ply

DL3DV example:

SCENE_ROOT=processed_dl3dv_ours/1K bash run_da3.sh

For pre-split scene lists, such as ScanNet v2 lists:

SPLIT_DIR=scannetv2/splits NUM_GPUS=8 bash run_da3.sh

2. PGSR / 3DGS Training

The generic batch script supports ScanNet v2, ScanNet++, and DL3DV. Set roots explicitly for your dataset.

ScanNet v2 example:

SCENE_ROOT=scannetv2/scans \
OUTPUT_ROOT=pgsr_scannetv2_all \
LOG_FILE=train_pgsr_progress_scannetv2.json \
bash 3dgs_train.sh

ScanNet++ example:

SCENE_ROOT=scannetppv2/data \
OUTPUT_ROOT=pgsr_scannetppv2_all \
LOG_FILE=train_pgsr_progress_scannetppv2.json \
bash 3dgs_train.sh

The expected trained model output is:

<OUTPUT_ROOT>/<scene_id>/point_cloud/iteration_30000/point_cloud.ply

3. Mesh and Mesh-Guided Masks

ScanNet v2 example:

DATA_ROOT=scannetv2/scans \
OUTPUT_ROOT=pgsr_scannetv2_all \
bash mesh.sh

ScanNet++ example:

DATA_ROOT=scannetppv2/data \
OUTPUT_ROOT=pgsr_scannetppv2_all \
bash mesh.sh

Outputs:

• <3dgs_output>/<scene_id>/mesh/tsdf_fusion_post.ply
• <data_root>/<scene_id>/mask/

4. Object and Region Annotation

Start an OpenAI-compatible vLLM server, then run object annotation.

ScanNet++ object instances:

DATA_ROOT=scannetppv2/data \
SCENES_DIR=pgsr_scannetppv2_all \
CLASS_JSON_DIR=Qwen3VL-32B-Scannetppv2 \
MASK_ROOT=sam3_masks_scannetppv2_new \
OUTPUT_JSON_DIR=output_scannetppv2_new \
VLLM_API_BASE=http://localhost:8000/v1 \
VLLM_CHAT_URL=http://localhost:8000/v1/chat/completions \
bash scannetppv2_new.sh

ScanNet++ functional regions:

DATA_ROOT=scannetppv2/data \
SCENES_DIR=pgsr_scannetppv2_all \
CLASS_JSON_DIR=Qwen3VL-32B-Scannetppv2_region \
MASK_ROOT=sam3_masks_scannetppv2_region \
OUTPUT_JSON_DIR=output_scannetppv2_region_new \
VLLM_API_BASE=http://localhost:8000/v1 \
VLLM_CHAT_URL=http://localhost:8000/v1/chat/completions \
bash scannetppv2_region.sh

The same pattern applies to scannetv2_all.sh, scannetv2_region.sh, DL3DV_new.sh, and dl3dv_new_2k.sh.

Annotation outputs are one JSON per scene, containing instance labels, source images, masks, 3D boxes, and metadata needed by QA generation.

5. AABB Box Post-Processing

Convert the raw 3D object boxes into floor-aligned AABB-style boxes before QA generation:

python postprocess_3d_bbox_aabb.py \
  --input_dir output_scannetppv2_new \
  --output_dir output_scannetppv2_new_aabb \
  --floor_label floor \
  --axis_method largest_face \
  --extent_mode keep

For ScanNet v2, use the corresponding object annotation directory:

python postprocess_3d_bbox_aabb.py \
  --input_dir output_scannetv2_new \
  --output_dir output_scannetv2_new_aabb \
  --floor_label floor

The region JSONs can be kept as generated by the region annotation step.

6. Instance Descriptions and QA

Generate object descriptions:

python qa_generation/generate_3d_instance_description.py \
  --input_dir output_scannetppv2_new_aabb \
  --output_dir output_scannetppv2_new_aabb_with_descriptions \
  --dataset_root scannetppv2/data \
  --vllm-api-url http://localhost:8000/v1/chat/completions

Generate object + region spatial QA:

python qa_generation/generate_two_view_qa_region.py \
  --data-root scannetppv2/data \
  --wai-root scannetppv2_wai \
  --marker-types language_description \
  --bbox-json-folder output_scannetppv2_new_aabb_with_descriptions \
  --region-bbox-json-folder output_scannetppv2_region_new \
  --output output_QA_new_lang_add_region

Filter repeated/ambiguous descriptions:

python qa_generation/filter_qa_repeat_descriptions.py \
  --data-root scannetppv2/data \
  --bbox-json-folder output_scannetppv2_new_aabb_with_descriptions \
  --output output_QA_new_lang_add_region \
  --output-filter-repeat output_QA_new_add_region_filter_repeat \
  --vllm-api-url http://localhost:8000/v1/chat/completions

7. Convert to LLaMA-Factory

python qa_generation/convert_qa_to_llamafactory.py \
  --image-root scannetppv2/data \
  --qa-json-dir output_QA_new_add_region_filter_repeat \
  --output-dir output_llamafactory_QA_add_region_lang

The final output contains dataset.jsonl and copied/rendered images in the format expected by LLaMA-Factory multimodal SFT.

Main Outputs

• DA3: depth_da3/, pointcloud_da3.ply
• PGSR: pgsr_*/<scene>/point_cloud/iteration_30000/point_cloud.ply
• Mesh/masks: mesh/tsdf_fusion_post.ply, <data_root>/<scene>/mask/
• Object boxes: output_scannetppv2_new/<scene>.json
• AABB object boxes: output_scannetppv2_new_aabb/<scene>.json
• Region boxes: output_scannetppv2_region_new/<scene>.json
• Descriptions: output_scannetppv2_new_aabb_with_descriptions/<scene>.json
• QA: output_QA_new_add_region_filter_repeat/<scene>.json
• LLaMA-Factory: output_llamafactory_QA_add_region_lang/dataset.jsonl

Notes

• All large datasets, model checkpoints, masks, QA outputs, and 3DGS outputs are ignored by .gitignore.
• The batch scripts skip scenes whose final output JSON or trained point cloud already exists.
• For single-scene debugging, call the Python scripts directly with --scene, --data-root, --scene-json, --mask-root, and --model-path.

Acknowledgements

This pipeline builds on and benefits from several open research works and foundation models:

• SAM3
• Depth Anything 3
• PGSR
• Qwen3-VL
• Gemini 3 Flash

We also thank the dataset providers:

• ScanNet++
• ScanNet
• DL3DV-10K

We thank the authors and maintainers of these projects for making their work available to the community.

Citation

If you find this work useful, please consider citing:

@article{gao2026holi,
  title={Holi-spatial: Evolving video streams into holistic 3d spatial intelligence},
  author={Gao, Yuanyuan and Li, Hao and Liu, Yifei and Ji, Xinhao and Gong, Yuning and Liao, Yuanjun and Liu, Fangfu and Zhang, Manyuan and Yang, Yuchen and Xu, Dan and others},
  journal={arXiv preprint arXiv:2603.07660},
  year={2026}
}