syscon3d / README.md
syscon3d-neurips26's picture
Hide Croissant metadata from dataset card
9aea33c verified
---
pretty_name: SysCON3D
license: other
tags:
- 3d
- multiview
- benchmark
---
# SysCON3D
Portable release bundle for the SysCON3D benchmark and demo.
## Contents
- `mipnerf360_calibration_splits.json`: consistent-scene calibration splits.
- `mipnerf360_impossible_splits.json`: precomputed SysCON3D benchmark samples.
- `archives/syscon3d_mipnerf360_*.tar`: tar shards containing `mipnerf360/...` payload files.
## Release summary
- scenes: bicycle, bonsai, counter, flowers, garden, kitchen, room, stump, treehill
- copy mode: referenced-only
- storage mode: tar-shards
- copied files: 5059
- copied size (GiB): 0.606
- archive files: 1
## Notes
- The manifests in this folder use `dataset_root: mipnerf360`, so they are portable across local runs and Spaces.
- If this release uses `archives/`, extract those tar files before running tools that expect the raw `mipnerf360/` directory.
## Extracting Files
Extract the full payload:
```bash
mkdir -p tmp/syscon3d_release
for shard in tmp/syscon3d_release/archives/*.tar; do
tar -xf "$shard" -C tmp/syscon3d_release
done
```
List files without extracting:
```bash
tar -tf tmp/syscon3d_release/archives/syscon3d_mipnerf360_000.tar | less
```
Extract one file:
```bash
tar -xf tmp/syscon3d_release/archives/syscon3d_mipnerf360_000.tar \
-C tmp/syscon3d_release \
mipnerf360/syscon3d_scene_types/noise_gaussian/k09/noise_gaussian_k09_000/images_4/view_000.png
```
Extract one sample directory:
```bash
tar -xf tmp/syscon3d_release/archives/syscon3d_mipnerf360_000.tar \
-C tmp/syscon3d_release \
mipnerf360/syscon3d_scene_types/patched_gaussian/k09/patched_gaussian_k09_000
```
The paths in `mipnerf360_impossible_splits.json` are relative to
`tmp/syscon3d_release/mipnerf360/` after extraction.
## License and Source Data
SysCON3D is a derived benchmark bundle for research evaluation. It contains
referenced Mip-NeRF 360 scene images plus deterministic materialized stress-test
images. Users should follow the terms of the upstream source data and cite the
source dataset and SysCON3D paper when using this benchmark.