You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This is a derived (resized and cropped) copy of DL3DV-10K (https://github.com/DL3DV-10K/Dataset). Access is restricted under the DL3DV-10K Terms of Use and the CC BY-NC 4.0 license: you may use the data only for non-commercial research and educational purposes and you may not redistribute it. Please also request access to the original DL3DV-10K dataset and accept its terms before using this copy.
Log in or Sign Up to review the conditions and access this dataset content.
DL3DV-480P, 224 x 224 Lanczos centre crop
A pre-processed copy of DL3DV-10K (the 480P release, images_8, 480 x 270),
resized and cropped to 224 x 224 for feed-forward novel-view-synthesis / 3DGS training and evaluation at 224 resolution.
Processing
For every frame of every scene:
- PIL
Image.LANCZOSresize keeping the aspect ratio so that the short side is 224 (480 x 270 -> 398 x 224). - Fixed centre 224 x 224 crop (no random crop, no flip).
- Pinhole intrinsics scaled by the resize factor and shifted by the crop offset;
w2cunchanged. The 480P images are not undistorted (as in the original release). - Saved as lossless PNG.
The result is bit-identical to an on-the-fly loader that applies the same Lanczos cover resize + centre crop to the original 480P frames (verified on random frames of every split: pixels, intrinsics and poses).
Contents
| split | scenes | shards |
|---|---|---|
train |
10,350 | train/train-000.tar ... train/train-103.tar (100 scenes each) |
test |
140 | test/test-000.tar, test/test-001.tar |
Each tar holds <split>/<scene_id>/opencv_cameras.json and <split>/<scene_id>/images_224/frame_XXXXX.png.
opencv_cameras.json keeps the original structure; every frame has file_path, w, h (224), fx, fy, cx, cy
(adjusted) and w2c (OpenCV convention), plus a marker "precropped_fs": [224, 224].
Metadata at the repository root:
train_scenes.txt,test_scenes.txt: scene ids, in shard order.eval_ctx{6,12,24}_tgt8.json: context / target frame indices of the test scenes for 6 / 12 / 24 input views and 8 target views per scene.
Usage
huggingface-cli download HyebinKim/DL3DV_224_Lanczos_cropped --repo-type dataset --local-dir DL3DV_224
cd DL3DV_224 && for f in train/*.tar test/*.tar; do tar -xf "$f"; done
License and attribution
Derived from DL3DV-10K by Ling et al., "DL3DV-10K: A Large-Scale Scene Dataset for Deep Learning-based 3D Vision" (CVPR 2024). Modified (resized and cropped, see above). Licensed under CC BY-NC 4.0 and subject to the DL3DV-10K Terms of Use: non-commercial research and education only; no redistribution.
@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={CVPR},
year={2024}
}
- Downloads last month
- 210