Datasets:

devappendcbangj
/

DL3DV-OVS

pretty_name: DL3DV-OVS
license: cc-by-nc-4.0
task_categories:
  - image-segmentation
tags:
  - 3d
  - open-vocabulary
  - gaussian-splatting
  - object-segmentation
  - evaluation
gated: true
extra_gated_heading: Access DL3DV-OVS
extra_gated_description: >-
  Confirm the conditions below to access the files. You must separately accept
  the official DL3DV-10K terms for both source repositories.
extra_gated_button_content: Agree and access
extra_gated_fields:
  I have accepted the official DL3DV-10K Terms of Use: checkbox
  I will use these files only for non-commercial research or education: checkbox
  I will not redistribute DL3DV source content: checkbox

DL3DV-OVS: Open-Vocabulary 3D Scene Understanding Dataset for Large, Complex Indoor–Outdoor Scenes

DL3DV-OVS is a four-scene dataset and evaluation benchmark for open-vocabulary 3D scene understanding in large, complex indoor and outdoor environments. It was introduced with LightSplat.

Dataset contents

Component	Park	Shop	Road	Office	Total	Source
RGB/COLMAP frames	314	407	317	378	1,416	DL3DV
Annotated frames	5	3	3	3	14	Ours
GT masks	12	17	18	11	58	Ours
SAM/CLIP pairs	314	407	317	378	1,416	Ours
Text embeddings	5	8	7	5	22 unique	Ours
RGB 3DGS checkpoints	1	1	1	1	4	Ours

The 58 masks provide ground truth for 14 evaluation frames; the 1,416 SAM/OpenCLIP pairs cover all scene frames.

Original DL3DV RGB frames, cameras, and COLMAP caches are not included. Obtain them from the official DL3DV repositories after accepting their terms. Source images are 960×540. The companion setup applies COLMAP undistortion, and the provided features and masks match its output.

Access requirements

Before accessing these files, request access to both official DL3DV source repositories:

The DL3DV maintainers state that requesting access accepts the DL3DV-10K Terms of Use. Access here does not replace that upstream agreement.

Files

label/<scene>/gt/<frame>/<query>.jpg
reference/<scene>/language_features/<frame>_s.npy
reference/<scene>/language_features/<frame>_f.npy
reference/<scene>/checkpoint/chkpnt30000.pth
reference/text/text_features.json
metadata/annotations.json
metadata/scenes.json

Public scene names are park, shop, road, and office. The masks follow the LERF-OVS evaluation layout and use 8-bit grayscale JPEG, with foreground defined as pixel > 10. Each _s.npy stores per-pixel SAM mask IDs and the matching _f.npy stores one OpenCLIP image embedding per mask. The shared JSON contains normalized OpenCLIP text embeddings for all 22 benchmark queries. The encoder is OpenCLIP ViT-B-16 (laion2b_s34b_b88k).

Setup

Use the companion DL3DV-OVS repository to assemble the complete benchmark:

dl3dv-ovs setup data/dl3dv-ovs

License

DL3DV-OVS masks, features, checkpoints, and benchmark metadata are provided under CC BY-NC 4.0, subject to the DL3DV-10K Terms of Use. Original DL3DV data is not redistributed. See LICENSE.md.

Citation

If you use DL3DV-OVS, cite both LightSplat and DL3DV-10K.

@inproceedings{bang2026lightsplat,
  title={Lightsplat: Fast and memory-efficient open-vocabulary 3d scene understanding in five seconds},
  author={Bang, Jaehun and Kim, Jinhyeok and Kim, Minji and Jeong, Seungheon and Joo, Kyungdon},
  booktitle={2026 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={19812--19821},
  year={2026},
  organization={IEEE}
}

@inproceedings{ling2024dl3dv,
  title={Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision},
  author={Ling, Lu and Sheng, Yichen and Tu, Zhi and Zhao, Wentian and Xin, Cheng and Wan, Kun and Yu, Lantao and Guo, Qianyu and Yu, Zixun and Lu, Yawen and others},
  booktitle={2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={22160--22169},
  year={2024},
  organization={IEEE}
}