ZhewenZheng/PAC-NeRF-recon

PAC-NeRF reconstruction — 2026-09-20

45 completed UniPhy reconstruction/preparation runs on PAC-NeRF: 10 elastic, 10 plasticine, 5 sand, 10 Newtonian and 10 non-Newtonian sequences. This release contains learned Gaussian reconstructions, filled volume particles, fitted deformation correspondence, training states, metrics and retained media. It does not contain fitted Stage 2 material fields or future physics predictions. The particles and deformation are model estimates, not ground-truth trajectories.

Temporal split and numerical settings

Each sequence has 14 source frames at 24 fps. Only the first 9 frames (indices 0–8, 0–1/3 seconds) were used for reconstruction and preparation; the remaining 5 are held out. This is a user-defined floor(2N/3) split, not an official PAC-NeRF temporal prediction protocol. Image evaluation covers all 9 observed frames and 11 input cameras. The reconstruction budget was 40,000 updates; correspondence fitting used 10,000 updates. A completed budget is not a convergence claim. Simulation settings retain 200 substeps per observation, dt = 1/4800. Volume-filling lattice spacing is a separate quantity, recorded in each volume/volume.json; it is not the MPM grid spacing.

Initialization uses uniformly sampled points in the recorded configuration box. initialization.json contains the actual bounds and seed; the original record labels bounds provenance as unspecified. Do not interpret this release as an ablation proving that initialization is independent of all scene priors. Foreground supervision was produced by the PAC-NeRF preprocessing/matting path; one union foreground mask is used, without instance-separated supervision.

Contents

Each archives/PAC-NeRF-data-<case>.tar.gz extracts into:

PAC-NeRF-data-<case>/recon/prestage2-20260919/
  settings.json, reconstruction.json, initialization.json, runtime.json
  prepared.json, history.jsonl, cfg_args
  point_cloud/iteration_40000/point_cloud.ply  # canonical appearance Gaussians
  deform/iteration_40000/deform.pth           # image-reconstruction deformation
  gaussians.pt                               # appearance at first observation
  initial_positions.npy, initial_velocities.npy
  training_state.pt                          # saved reconstruction training state
  volume/
    frame0000.npz ... frame0008.npz           # filled positions, opacity, surface indices
    gaussians.pt                             # neutral volume proxies, not RGB appearance
    volume.json, history.json
  deformation/
    best.pt                                 # selected model/canonical points; weights-only
    training_state.pt                       # optimizer/model/RNG state for correspondence
    complete.json, history.json
  eval/
    comparison_camera0.mp4                   # source RGB/recon and union mask/alpha
    reconstruction_grid.mp4                  # all 11 source cameras
    reconstruction.rrd                      # interactive 3D learned Gaussian centers
    comparison_*.png, grid_*.png
    metrics.json, visualization.json

Per-frame volume particles have varying counts and no shared particle indexing. Use deformation/best.pt for learned correspondence. Its selected weights are not a resumable optimizer checkpoint. The separate training_state.pt files preserve their own matching saved training state; never combine them with selected weights. Intermediate gs/ PLY exports and img/ training snapshots are omitted as redundant. Original source image archives, credentials, environment files and project source archives are not included. Small videos and metrics are also exposed under previews/ so they can be inspected without downloading training states.

Download and extract

Authenticate with hf auth login while this repository is private.

hf download ZhewenZheng/PAC-NeRF-recon --repo-type dataset --local-dir pacnerf-recon-release

From a Python session in the UniPhy repository:

from pathlib import Path
import tarfile

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

Archives retain the experiment directory name so they can be selected by --reconstruction-name prestage2-20260919 in the canonical infer_material.py entrypoint. Install UniPhy using its repository README; CUDA and its renderer extensions are required for reconstruction rendering/training. Raw PAC-NeRF inputs and the matting model are separate dependencies for rerendering against source images or Stage 2:

python -m scripts.download --datasets PAC-NeRF --root data

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('PAC-NeRF-*/recon/prestage2-20260919'):
    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.

Image reconstruction quality

The table averages per-image metrics within each case and then across cases. Foreground PSNR restricts RGB error to the target foreground, reducing the influence of the large black background. IoU thresholds target and reconstructed alpha at 0.5. These are input-view, observed-frame fitting metrics, not held-out-view or future prediction metrics, and do not establish interior geometry or correspondence accuracy.

Material	Cases	Mean mask IoU	Mean foreground PSNR
Elastic	10	0.9956	33.31 dB
Plasticine	10	0.9912	31.80 dB
Sand	5	0.9896	27.78 dB
Newtonian	10	0.9953	33.95 dB
Non-Newtonian	10	0.9959	34.41 dB

Visual spot checks found good overall motion/silhouette agreement, with local contact-region differences and softened surface detail. Some source foregrounds contain ground-adjacent artifacts. The worst individual plasticine-7 view/frame has IoU 0.808 despite a high case mean. See per-image metrics for these exceptions.

Provenance and usage

These are derived research outputs from PAC-NeRF processed with the UniPhy/MASIV reconstruction pipeline. Original PAC-NeRF data are distributed separately. This release does not grant additional rights to upstream data, renderers or code; retain applicable upstream terms and cite the underlying work when using results. Upstream configuration links and the MASIV source revision are retained in metadata.

See manifest.json for the archive inventory, particle counts and per-case metrics, and VIDEOS.md for all 45 comparison and multi-camera videos.