ZhewenZheng/Spring-Gaus-recon

Spring-Gaus reconstructions — 2026-09-20

12 completed UniPhy/MASIV reconstruction and preparation runs, organized into Synthetic/ (7 cases) and Real/ (5 cases). These are learned reconstructions, filled volume particles and fitted correspondence, not ground-truth geometry or fitted Stage 2 material models. No future physics predictions are included.

Subset	Cases	Observed / total frames per case	Input cameras	Frame interval
Synthetic	7	20 / 30	10	0.03 s
Real	5	13 / 20	3	1/120 s

Synthetic uses the released 20/10 split. Real uses a user-defined floor(2N/3) observed prefix, not an official temporal prediction split. Future images do not supervise reconstruction, filling or correspondence. Each case completed 40,000 reconstruction updates and 10,000 correspondence updates. Completing those budgets is not a claim of convergence or of successful Stage 2 trajectory prediction. Source cameras, image framing, full-frame videos and timestamps are retained.

Files

Synthetic/
  archives/Spring-Gaus-Synthetic-<case>.tar.gz
  previews/Spring-Gaus-Synthetic-<case>/
    comparison_camera0.mp4
    reconstruction_grid.mp4
    metrics.json
Real/
  archives/Spring-Gaus-Real-<case>.tar.gz
  previews/Spring-Gaus-Real-<case>/
    comparison_camera0.mp4
    reconstruction_grid.mp4
    metrics.json
manifest.json
quality.json
VIDEOS.md

Each archive retains Spring-Gaus-<subset>-<case>/recon/<experiment>/ and contains:

• point_cloud/iteration_40000/point_cloud.ply and deform/iteration_40000/deform.pth: canonical appearance Gaussians and the image-reconstruction deformation network.
• gaussians.pt, initial_positions.npy, initial_velocities.npy: reconstructed first state.
• training_state.pt: saved matching reconstruction model/optimizer/RNG state.
• volume/frame*.npz, volume/gaussians.pt, volume/volume.json: filled per-frame particles, opacity, surface indices and sampling metadata. Volume Gaussians have neutral appearance; they are not the RGB appearance model.
• deformation/best.pt: selected model and canonical points for correspondence; this selected file is weights-only. deformation/training_state.pt contains its own saved training state. Do not combine selected weights with another optimizer.
• settings.json, reconstruction.json, initialization.json, runtime.json, prepared.json, cfg_args and histories: settings and provenance.
• eval/: full-frame source/reconstruction RGB and mask/alpha videos, source-camera grid videos, snapshots, per-image metrics and an interactive reconstruction.rrd. The Rerun recording shows learned 3D Gaussian centers and the fixed dataset cube.

Per-frame volume particle counts vary and do not share particle identities. Use deformation/best.pt for learned correspondence. Interior accuracy and mass conservation are not established by good image reconstruction metrics. Redundant intermediate gs/ exports and img/ previews are omitted. Source code, credentials and original input datasets are not bundled.

Banana repair

Only banana uses experiment temporal-20260920; the other cases use prestage2-20260919. Its original run lost the object from most cameras starting at observed frame 12, producing extreme positions and an impossible dense volume allocation. This release includes only the repaired, fully prepared result.

The repair sets reconstruction.temporal_growth_interval=500: after the existing 3,000-update first-frame warm-up, two frames become eligible and one more is added every 500 updates; all 20 are eligible at update 12,000. Initialization, losses, cameras, split, timesteps and volume sampling remain unchanged. No coordinate clamp or GT geometry was used in this repair. The final mean mask IoU is 0.98947 and final-frame IoU is 0.98472. All 20 volume frames and 10,000 correspondence updates completed. The discarded foreground-reweighted trial is not included.

Quality checks and limitations

All 12 saved final reconstruction and correspondence training states have the expected completed step counts and finite network tensors. Every one of the 205 observed volume frames has finite positions/opacity, the recorded particle count, and valid surface indices. Retained comparison videos were visually inspected at uniform times through the last observed frame and around low-IoU times.

Synthetic reconstructions preserve the main motion and shape with some smoothing. Real reconstructions preserve the main object and motion, but show small silhouette spikes, slightly enlarged boundaries, alpha speckle and softened detail; burger's layer detail is less sharp. The weakest individual Real view is potato frame 0, camera 2 (IoU 0.8701). This is not a claim of flawless masks or interior geometry.

Metrics below average full observed frames and source cameras per case. Foreground PSNR uses target foreground pixels. Mask IoU thresholds alpha at 0.5. These are input-view fitting metrics, not unseen-view or future prediction metrics.

Subset / case	Mean mask IoU	Final-frame mask IoU	Foreground PSNR	Worst frame/view IoU
Real-bun	0.9782	0.9887	28.51 dB	0.9534
Real-burger	0.9600	0.9561	26.47 dB	0.9089
Real-dog	0.9716	0.9828	26.82 dB	0.9270
Real-pig	0.9746	0.9854	27.27 dB	0.9411
Real-potato	0.9651	0.9811	28.77 dB	0.8701
Synthetic-apple	0.9975	0.9974	39.73 dB	0.9957
Synthetic-banana	0.9895	0.9847	31.59 dB	0.9599
Synthetic-chess	0.9959	0.9955	33.33 dB	0.9898
Synthetic-cream	0.9950	0.9945	33.62 dB	0.9873
Synthetic-cross	0.9966	0.9964	36.39 dB	0.9946
Synthetic-toothpaste	0.9960	0.9955	34.83 dB	0.9910
Synthetic-torus	0.9933	0.9910	32.39 dB	0.9799

Download and extract

Authenticate with hf auth login while the repository is private.

hf download ZhewenZheng/Spring-Gaus-recon --repo-type dataset --local-dir springgaus-recon-release

From the UniPhy repository:

from pathlib import Path
import tarfile

for archive in Path('springgaus-recon-release').glob('*/archives/*.tar.gz'):
    with tarfile.open(archive) as bundle:
        bundle.extractall('runs', filter='data')

For source-image evaluation or Stage 2 fitting, install UniPhy and its CUDA extensions using its repository README, and separately obtain original inputs:

python -m scripts.download --datasets Spring-Gaus-Synthetic Spring-Gaus-Real --root data

Use the canonical infer_material.py entrypoint. Select --reconstruction-name prestage2-20260919 for all cases except banana, for which use --reconstruction-name temporal-20260920. Keep the saved numerical settings and temporal split when reusing results. Synthetic starts from the recorded configuration box; Real uses the released static cloud with the recorded registration for initialization. Do not label this entire release as a scene-prior-free initialization experiment.

Relocate metadata before using the saved models

Machine-specific paths were replaced by ${PROJECT_ROOT}, ${DATA_ROOT} and ${RUNS_ROOT} in JSON, saved checkpoint metadata and cfg_args. Tensor values, optimizer state and RNG state were retained. Set these paths to the new project, official dataset root and extraction root. This matters because the pipeline checks that source paths and reconstruction settings match, including checkpoint metadata. The following only changes path strings in the extracted copy:

import json
from pathlib import Path
import torch

roots = {'${PROJECT_ROOT}': Path('.').resolve().as_posix(),
         '${DATA_ROOT}': Path('data').resolve().as_posix(),
         '${RUNS_ROOT}': Path('runs').resolve().as_posix()}

def relocate(value):
    if isinstance(value, str):
        for token, path in roots.items():
            value = value.replace(token, path)
        return value
    if isinstance(value, dict):
        return {k: relocate(v) for k, v in value.items()}
    if isinstance(value, list):
        return [relocate(v) for v in value]
    if isinstance(value, tuple):
        return tuple(relocate(v) for v in value)
    return value

for directory in Path('runs').glob('Spring-Gaus-*/recon/*'):
    for path in directory.rglob('*'):
        if path.suffix == '.json':
            path.write_text(json.dumps(relocate(json.loads(path.read_text())), indent=2))
        elif path.suffix == '.pt':
            state = torch.load(path, map_location='cpu', weights_only=True)
            torch.save(relocate(state), path)
        elif path.name == 'cfg_args':
            path.write_text(relocate(path.read_text()))

Use the retained settings.json as the record of numerical choices. Do not silently replace these settings with newer defaults when resuming or fitting physics. Exact training continuation also depends on the compatible UniPhy implementation and CUDA environment; their recorded revision/runtime are in runtime.json.

Provenance

Derived research outputs from Spring-Gaus processed with UniPhy/MASIV. Original data are distributed separately. This release grants no additional rights to upstream data, renderers or code; retain their applicable terms and cite the underlying work. Source configuration links and runtime revisions remain in the saved metadata. manifest.json records the file inventory and particle counts; quality.json records this check; VIDEOS.md links all 24 videos.