You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This repository contains data derived from ScanNet++ (undistorted-ready DSLR views resized to 256x256, with COLMAP poses and intrinsics). ScanNet++ may not be redistributed. Access is granted only to individuals who already hold a ScanNet++ license from the ScanNet++ team and agree to use this data under the ScanNet++ Terms of Use (non-commercial research only, no redistribution).
Log in or Sign Up to review the conditions and access this dataset content.
static-scene-test
ScanNet++ DSLR views packed for novel-view-synthesis training: 256x256 JPEG (quality 95, 4:4:4), stretched to square, with camera-to-world poses, normalized intrinsics, per-view train/test split codes and the COLMAP fisheye camera model so views can be undistorted without the raw dataset.
Layout: train/ (856 scenes, official ScanNet++ NVS train split) and val/
(50 scenes). Each split holds the same scenes in five shard formats, 24 scenes
per shard: wds/ (WebDataset tar), parquet/, zarr/, hdf5/, mds/
(MosaicML Streaming). <split>/manifest.json lists shards, sizes and view
counts per format; <split>/index/shard-XXXX.json lists the scenes per shard.
Readers and packers live in the baselines/hfstream module of the owner's
scene-representation-learning repository.
Terms. Derived from ScanNet++. Use is subject to the ScanNet++ Terms of Use; access is approved manually and only for holders of a ScanNet++ license. Do not redistribute.
- Downloads last month
- 16