Changelog

Dataset: 1000+ scenes released with 600+ new scenes.

Please re-download and replace the old version of the dataset using the updated download script, since v2 includes notable improvements to v1 scene data.

• Longer-tail semantics with more unique objects annotated



• Balanced diversity of scene types: more apartments, workshops, recreational areas, and many new types. Scene type metadata is released (scene_types.json).

Statistic	v1	v2
Scenes	380	1006
DSLR frames	228K	1.1M
iPhone Frames	3.27M	10M
iPhone Duration	15 hours	46 hours
Unique Annotated Labels	1.7k	3.4k
Annotated Instances	30k	111k
Download size	400 GB	1.5TB

• Higher DSLR and iPhone coverage with lesser motion blur in new scenes
• Improved DSLR and iPhone poses and anonymization for all scenes (including updates to v1 scenes)
• iPhone video changed to mkv format for more accurate anonymization

Benchmarks:

• New scenes are added to the train set. The other splits remain the same, while annotations/poses are improved across all splits.
• New NVS-small benchmark on a subset of 12 scenes for faster evaluation. See details under documentation. The NVS-full benchmark is unchanged
• Please re-download and replace the old version of the dataset using the updated download script, since v2 includes notable improvements to v1 scene data.

380 scenes released along with NVS and Semantic benchmarks, several improvements to the dataset

• Added nvs_test and sem_test scenes, benchmark details
• Improved DSLR metadata (nerfstudio/transforms.json) listing the clean and blurry images, additional DSLR images in some scenes
• Semantic annotation on segments (segments.json, segments_anno.json) changed
• Filtering of images in iphone/colmap improved
• Several assets zipped for faster downloads