boazkg/gs-pool-paslcd

GS-Pool on PASLCD

This dataset is built on PASLCD, the Pose-Agnostic Scene-Level Change Detection dataset by Galappaththige et al. (dataset, paper, CVPR 2025). Its photographs, camera solve and change masks are theirs; please cite their work (below) whenever you use this data.

The reconstructions, evidence and results of GS-Pool (paper, code) on the ten scenes of PASLCD, as GS-Pool 1.3.1 computed them: 240 members, each a scene instance in one camera frame (joint or anchored), reconstructed by one trainer (FastPGSR or vanilla 3DGS) with one seed (22, 23, 24). PASLCD's photographs, solve and masks are not included; the code downloads them from PASLCD.

Structure

members/<Scene>/Instance_<n>/
├── <joint|anchored>/<fastpgsr|vanilla>/seed<22|23|24>/
│   ├── t1/, t2/                   the two reconstructions (model.pt with the distilled field, model.ply)
│   ├── evidence/                  object pools, carrier scores, view statistics, and the per-photograph rows
│   │                              the view statistics are computed from (photo_trace)
│   └── paper_results/
│       ├── change_detection/      receipt.json, pool.json, decision trace, change masks; extras.json (joint),
│       │                          the score of every readout of the paper's supplement
│       ├── extras/<readout>/      each readout (joint): the ladder's rungs, the comparisons, the oracle and the
│       │                          sweeps, with its receipt.json, pool.json and scored masks (pred/); the SAM2
│       │                          and view-door sweeps also keep the pools and view statistics they decided on
│       │                          (evidence/). paslcd-extras regenerates the rest (per-visit masks, change PLYs,
│       │                          decision trace) from the released evidence
│       ├── unchanged/vs_seed<s>/  against another reconstruction of the same capture (joint): result, receipt,
│       │                          decision trace, masks and the pair's evidence
│       └── lighting/              against the same unchanged scene under other lighting (joint, Instance_1),
│                                  kept as the unchanged pairs are, with the masks at Instance_2's reference
│                                  cameras (reference_masks/)
└── baselines/<mv3dcd|oscd>/lighting/   MV-3DCD and O-SCD on the same lighting test (Instance_1)
inputs/<Scene>/Instance_<n>/<joint|anchored>/
├── undistorted/                   PASLCD's photographs undistorted, with their cameras
├── sam2/                          SAM2 label maps (automatic, linked)
├── sam2_stability<v>/             the same at SAM2's swept stability thresholds 0.8, 0.85 and 0.95 (joint)
├── dinov3/                        DINOv3 features (raw, pca)
├── solve/                         our per-visit camera solves (anchored only)
└── repair/                        a repaired joint solve (Pots, Instance_1)

How to run

With the code installed, one member, starting from its released evidence:

bash run.sh paslcd fastpgsr joint Porch 1 22 --from decision
python -m gspool paslcd-extras --scene Porch --instance 1 --seed 22 --substrate fastpgsr     # every supplement readout
python -m gspool paslcd-fp --test unchanged --scene Porch --instance 1 --substrate vanilla   # the same-capture pairs

The runner downloads only what that member needs. --from carriers and --from field start from the released reconstructions, --use-caches uses the released inputs/, and --data DIR points at a local copy of this dataset.

Citation

PASLCD:

@inproceedings{galappaththige2025multi,
  title     = {Multi-View Pose-Agnostic Change Localization with Zero Labels},
  author    = {Galappaththige, Chamuditha Jayanga and Lai, Jason and Windrim, Lloyd and Dansereau, Donald and Sunderhauf, Niko and Miller, Dimity},
  booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages     = {11600--11610},
  year      = {2025}
}

GS-Pool:

@article{kerengil2026gspool,
  title   = {GS-Pool: Object-Level Change Detection in 3D Gaussian Splatting},
  author  = {Keren-Gil, Boaz and Gain, James and Marais, Patrick},
  journal = {arXiv preprint arXiv:2610.06688},
  year    = {2026}
}