boazkg/gs-pool-paslcd

GS-POOL on PASLCD: everything the paper measured

The products behind GS-POOL: Object-Pooled Change Detection in the Primitive Space of 3D Gaussian Splatting, on the ten scenes of PASLCD: 10 scenes x 2 instances x 2 substrates (FastPGSR, vanilla 3DGS) x 2 camera frames x 3 reconstruction seeds (22, 23, 24) = 240 reconstructed members with their evidence, and the photograph-only caches every member was computed with. Roughly 200 GB.

The benchmark's own data — the photographs, its joint COLMAP solve and the annotated masks — is not re-hosted here. It is downloaded from PASLCD itself, and our runner does that for you, pinned to the revision these products were computed from (aef858fe).

Layout

One directory per camera frame, each self-contained. joint is the benchmark's own solve over both visits; anchored is ours, with the inference visit registered into the reference poses, which are never re-optimised.

<joint|anchored>/
  caches/<Scene>/Instance_<n>/feat_raw1184/{T1,T2}/*.npy   DINOv3 ViT-L/16 patch grids, long side 1184
  caches/<Scene>/Instance_<n>/pca32.pkl                    the cell's PCA to 32 dimensions, over both visits
  caches/<Scene>/Instance_<n>/sam2/{T1,T2}/*.png           SAM2 label maps of the keyframes, ids linked across them
  caches/<Scene>/Instance_<n>/sam2_auto/{T1,T2}/*.png      the automatic SAM2 maps the links were made from
  members/<Scene>/Instance_<n>/<fastpgsr|vanilla>/seed<22|23|24>/
      t1/model.pt  t2/model.pt                       each visit's reconstruction with its distilled semantic
                                                     field (a gsplat checkpoint: splats + a 32-d feats tensor)
      scoring/prim_scores.npz                        per-primitive evidence: the four carriers (geometric,
                                                     colour, semantic, photographic), coverage, drift scales
      scoring/pool_T{1,2}.npz                        the object pool (an identity per primitive)
      scoring/view_stats.npz                         the photographic pass, per identity: observing views,
                                                     hot fraction, displacement, feature agreement

joint/paslcd_repairs/Pots/Instance_1/sparse/0/       the benchmark's one non-metric solve (a 70,497x gauge
                                                     against its own sibling), rescaled; GAUGE_REPAIRED
                                                     records the factor
anchored/solves/<Scene>/Instance_<n>/t{1,2}/sparse/0/  our two per-visit solves in one frame

T1 is the reference visit, T2 the inference visit (PASLCD's _test_ images). Undistortion follows each frame's own cameras, so the two frames carry separate caches and separate members. The caches depend on the photographs alone, so a new reconstruction of a cell reuses them. Every solve here names PASLCD's files with the benchmark's own spelling.

Use

git clone https://github.com/BoazKG93/gs-pool && cd gs-pool
uv venv .venv --python 3.12 && source .venv/bin/activate
uv pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
uv pip install -e ".[sam2]"

python -m gspool paslcd --scene Cantina --instance 1 --seed 22     # downloads what it needs, decides, scores
python -m gspool paslcd --substrate vanilla --frame anchored       # a whole row
python -m gspool paslcd --from field                               # re-derived as in the paper's campaign

--from verdict|evidence|field|scratch chooses the stage a member is reproduced from: decide on the banked evidence, recompute it from the banked reconstruction, re-distil the field onto the frozen primitives, or train from the photographs. --rung 1..9 and --oracle reproduce the paper's ablation ladder and its ceiling: rungs 1 to 5 build the method a brick at a time (the geometric kernel alone, then colour, then semantics, then the photographs — the four carriers, the core — then the corroboration rule), and rungs 6 to 9 take one rule or one clause back off the full system. The receipts the paper's tables are means of are in the code repository under results/, and results/check_paper_numbers.py recomputes every quoted number from them.

Nothing here is written by the decision stage: these are the photographs' caches, the solves and the reconstructions with their per-primitive evidence, the pool and the view statistics. A change to how the verdict reads them re-decides these same files, which is what every campaign in the paper does.

Licence and attribution

MIT, the licence of the benchmark these products are derived from.

Every file here is computed from the photographs of PASLCD (Galappaththige, Lai, Windrim, Dansereau, Sunderhauf and Miller; MIT), and is a derivative work of them: the reconstructions were trained on those photographs, the grids and label maps describe them, the solves pose them. Use of these files is use of PASLCD, so cite the benchmark alongside our paper.

Multi-View Pose-Agnostic Change Localization with Zero Labels
C. J. Galappaththige, J. Lai, L. Windrim, D. Dansereau, N. Sunderhauf, D. Miller
CVPR 2025.  https://chumsy0725.github.io/MV-3DCD/

Citation

@article{kerengil2026gspool,
  title   = {GS-POOL: Object-Pooled Change Detection in the Primitive Space of 3D Gaussian Splatting},
  author  = {Keren-Gil, Boaz},
  year    = {2026}
}