KCBtheone/splatatlas-core

SplatAtlas Core Bundle

This is the lightweight core bundle for the anonymous NeurIPS 2026 submission of SplatAtlas, a 3D Gaussian Splatting benchmark over 31 methods × 31 scenes.

Single landing page:

• https://huggingface.co/KCBtheone

Companion runs bundle:

• https://huggingface.co/datasets/KCBtheone/splatatlas-runs

This core bundle is intended for quick reviewer inspection. It contains aggregated CSV/JSON source data for paper figures and tables, audit summaries, Croissant metadata, wrapper templates, and manifests that map large run artifacts to the companion runs bundle.

Large PLY/checkpoint artifacts are not duplicated here. They are stored in the runs bundle as archive pairs:

outputs/<method_key>_<scene_key>_point_cloud.tar
outputs/<method_key>_<scene_key>_rest.tar

The full 31 × 31 archive mapping is:

manifests/runs_archive_manifest.tsv

Bundle split

Bundle	Purpose	Expected use
Core bundle	Lightweight CSV/JSON/code/metadata/manifests	Review figures/tables, audit claims, wrapper templates
Runs bundle	Large per-experiment archives	Inspect selected trajectories, checkpoints, Gaussian fields

Runs archive layout

Each canonical method-scene cell in the runs bundle has two archives.

Archive suffix	Contents
_point_cloud.tar	Exported Gaussian-field / PLY point-cloud artifacts, including checkpointed point-cloud directories when available.
_rest.tar	TensorBoard scalar trajectories, JSON/YAML metadata, metrics, timing files, audit logs, configs, and method-native non-PLY checkpoints.

Example:

Method	Scene	Point-cloud archive	Rest archive
pgsr	truck	outputs/pgsr_truck_point_cloud.tar	outputs/pgsr_truck_rest.tar
2dgs	chair	outputs/2dgs_Chair_point_cloud.tar	outputs/2dgs_Chair_rest.tar

Figure/table source mapping

Paper item	Source file(s)	Columns / fields used	Notes
Figure 1	csvs/task_4_1_seed_variance_raw_reconstructed.csv; csvs/task_4_1_seed_variance_summary_reconstructed.csv; selected Truck archives in runs bundle	method, scene, seed, PSNR/SSIM/LPIPS, summary statistics	Bonsai/Lego seed rankings and Truck render-vs-representation examples.
Figure 2	csvs/per_cell_metrics_long.csv; phase3 JSON summaries under csvs_and_phase_metadata/	method, scene, PSNR, opacity, anisotropy, scale, sh_energy_ratio	Variance attribution bar chart.
Figure 3	csvs/sh_loss_28x31.csv	method, scene, final_sh_mean, final_training_loss	SH vs. loss scatter over 868 cells.
Figure 4	phase3 CSV/JSON files under csvs_and_phase_metadata/	CCF lag, adaptability index, densification-window effect size	Three views on dynamical structure.
Figure 5	phase3 trajectory-feature CSV/JSON files under csvs_and_phase_metadata/	trajectory features, UMAP coordinates, cluster id, literature category	UMAP of trajectory features.
Figure 6	csvs/task_2_1_global_profile.csv; csvs/task_2_2_fig11_matrix.csv	method, PSNR, channel medians, relative IQR ratios	Saturation-zone IQR ratios.
Figure 7	csvs/task_2_2_fig11_matrix.csv; csvs/task_2_1_global_profile.csv	method, PSNR, SH net, SH corruption rate, opacity net, opacity pathology rate, coverage, coverage error fraction, residual	Counterfactual profile heatmap.
Table 1	N/A	N/A	Inclusion tiers are defined in the paper text.
Table 2	N/A	N/A	Channel definitions are defined in the paper text.
Table 3	csvs/task_2_3_table5.csv; csvs/task_2_3_appE_filtered_P01_P04.csv	pair_id, variant, channel, median delta, IQR, direction agreement, p-value, status	Paired validation summary.
Table 4	N/A	N/A	Hypothesis mapping is interpretive.
Table 5	phase2 reliability JSON files under csvs_and_phase_metadata/	noise floor, rasteriser rank consistency, additivity, SH-depth sensitivity	Reliability checks.
Table 6	wrapper_templates/; method inventory files if present	method key, display name, category, tier, notes	Method inventory.
Table 7	scene inventory file if present under csvs_and_phase_metadata/	scene, dataset, train/test/val counts, resolution, type	Scene inventory.
Table 8	csvs/task_2_3_appE_full.csv	pair_id, variant, scene, delta_PSNR, delta_SH, delta_opacity, delta_coverage, delta_residual	Appendix C per-scene paired deltas.
Table 9	csvs/task_4_1_seed_variance_summary_reconstructed.csv	method, scene, seed-variance summary fields	Appendix D seed variance summary.
Table 10	csvs/task_4_2_kendall_tau_checked.csv	scene, Kendall tau, ranking statistics	Appendix D Kendall tau.
Appendix E PSNR grid	csvs/appendixE_psnr_grid.csv	method × scene PSNR	Metric grid.
Appendix E SSIM grid	csvs/appendixE_ssim_grid.csv	method × scene SSIM	Metric grid.
Appendix E LPIPS grid	csvs/appendixE_lpips_grid.csv	method × scene LPIPS	Metric grid.

Important manifests

Manifest	Purpose
manifests/runs_archive_manifest.tsv	Full mapping from method-scene cell to large run archives.
manifests/runs_archive_missing.tsv	Missing archive report. Expected to contain only the header.
manifests/runs_archive_manifest_summary.txt	Summary of runs archive coverage and size.
manifests/core_file_manifest.tsv	File manifest for this core bundle.

Adding a new method

Wrapper templates are provided under:

wrapper_templates/
├── core/
├── methods/
└── configs/

Adding a new method typically requires:

1. create a method wrapper under wrapper_templates/methods/;
2. add one YAML configuration under wrapper_templates/configs/;
3. map the method entry point to the common train/eval/checkpoint interface;
4. verify metric logging, TensorBoard trajectories, checkpoint export, and audit logging.

The intended implementation size is roughly one method wrapper plus one YAML configuration entry.

Known caveats

• erankgs × LEGO is a known method-failure cell.
• analyticsplatting × LEGO has high seed variance and is treated as an outlier in seed-variance analysis.
• Some methods have unavailable entries in Appendix E metric grids where method output was not comparable or failed under the unified protocol.
• Counterfactual profiling uses a unified gsplat backend; native-rasteriser rank consistency is treated as a reliability check rather than a claim of pixel-identical equivalence.
• This core bundle is not a one-command reproduction package. It is a navigation and audit bundle mapping paper figures/tables to source files and mapping large run artifacts to the companion runs bundle.