3DROID
3DROID provides renderable 3D Gaussian reconstructions of 114 DROID scenes, together with the camera parameters used for reconstruction, rendered metric depth, and per-scene reliability measurements.
DROID source videos and robot states are not redistributed. Each scene records its
DROID uuid, which should be used to join the release with the original dataset. The
directory name is provided for readability but is not a reliable programmatic key.
Dataset layout
scenes/<episode>/
gaussians.ply four-view 3D Gaussian reconstruction
cameras.json intrinsics, extrinsics, preprocessing, and normalization
depth.npz metric depth, alpha coverage, cameras, and metric scale
metrics.json released-condition measurements, gate results, and flags
measurements.json measurements for all four experimental conditions
pipeline/ reconstruction, calibration-gate, and robot-probe code
evaluation/ table aggregation and paired statistical analyses
third_party/ZipSplat/ vendored ZipSplat source; pretrained weights are not included
manifest.json sizes and SHA-256 hashes of per-scene release files
reproduction.json per-scene numerical reproduction record
requirements.txt pinned pip environment
environment.yml pinned conda environment
LICENSES/ upstream licenses and third-party notices
Quick Start
This example downloads one scene and visualizes its released metric depth maps. It does not require a GPU, DROID source videos, or a reconstruction checkpoint.
1. Download one scene
pip install -U huggingface_hub numpy matplotlib
hf download wonguen/3DROID \
--repo-type dataset \
--include "scenes/IPRL__Wed_Oct_18_23-32-56_2023/*" \
--local-dir 3DROID
2. Load and visualize the depth maps
from pathlib import Path
import json
import matplotlib.pyplot as plt
import numpy as np
scene = Path(
"3DROID/scenes/IPRL__Wed_Oct_18_23-32-56_2023"
)
with open(scene / "cameras.json") as f:
cameras = json.load(f)
data = np.load(scene / "depth.npz")
depth = data["depth_m"] # (4, 252, 252), camera-axis depth in metres
alpha = data["alpha"] # (4, 252, 252), rendered coverage
valid = alpha > 0.5
print("Episode:", cameras["episode"])
print("UUID:", cameras["uuid"])
print("Views:", cameras["views"])
print("Depth shape:", depth.shape)
print("Model-to-metre scale:", float(data["s_m"]))
fig, axes = plt.subplots(1, 4, figsize=(14, 3))
for i, view in enumerate(cameras["views"]):
visible_depth = np.where(valid[i], depth[i], np.nan)
axes[i].imshow(visible_depth, cmap="turbo")
axes[i].set_title(view)
axes[i].axis("off")
plt.tight_layout()
plt.show()
The four depth maps follow the view order stored in cameras.json:
ext1L, ext1R, ext2L, ext2R
Pixels with alpha <= 0.5 are not treated as valid depth measurements.
Per-scene files
| File | Contents |
|---|---|
gaussians.ply |
Four-view reconstruction in the INRIA 3DGS PLY convention. This is the scene representation used for geometric evaluation. |
cameras.json |
Camera-to-world extrinsics for both calibration sources, DROID intrinsics, image preprocessing, and the normalization needed for rendering. |
depth.npz |
depth_m and alpha for the four exterior views, metric camera poses in cams, and the model-to-meter factor s_m. Pixels with alpha > 0.5 are treated as covered. |
metrics.json |
Measurements and quality flags for the released pw_inj reconstruction. |
measurements.json |
Photometric and geometric measurements for all four pose/extrinsic conditions. |
The four conditions are:
| Tag | Extrinsics | Pose conditioning |
|---|---|---|
orig_free |
DROID shipped | No |
orig_inj |
DROID shipped | Yes |
pw_free |
PointWorld refined | No |
pw_inj |
PointWorld refined | Yes; released configuration |
Reproduce the calibration and robot-surface diagnostics
The diagnostic driver is separate from reconstruction. The calibration gate runs before reconstruction, while the robot-surface probe evaluates rendered depth after reconstruction.
export DROID_INTRINSICS_ALL=/path/to/droid/full-intrinsics.json
# Reproduce the 225-candidate calibration gate.
python pipeline/run_diagnostics.py extract gate
# Recompute the robot-surface geometry measurements after the depth stage.
python pipeline/run_diagnostics.py measure
# Compare available diagnostic outputs with the released references.
python pipeline/run_diagnostics.py check
pipeline/episodes_eligible.txt contains the 225 gate candidates, and
pipeline/episodes.txt contains the 114 scenes that pass under the refined extrinsics.
pipeline/gate_expected.json records the expected gate outputs. The two lists and the
gate record agree exactly: 114 scenes pass under refined extrinsics and 80 under shipped
extrinsics.
The robot-surface probe uses the included Franka URDF and meshes, DROID joint states,
and the rendered four-view depth maps. It reproduces the geometry_arm_surface block.
See pipeline/README.md for stage dependencies, output paths, comparison tolerances,
and troubleshooting details.
Reconstruction backbones and checkpoints
Three feed-forward Gaussian backbones were evaluated in the paper. The released scenes and executable reconstruction path use ZipSplat; YoNoSplat and AnySplat are auxiliary baselines identified by pinned upstream revisions and checkpoint hashes.
| Model | Checkpoint source and revision | SHA-256 | Included |
|---|---|---|---|
| ZipSplat | veichta/zipsplat @ a0c9c8c1f886d1600e64837c2fcc6be6cc3b3694 |
e2906acdfadeb4f7b3311f398ede9e3a2fde734932a1a017c28a28597a5f5cd6 |
Source only; weights excluded |
| YoNoSplat | botaoye/YoNoSplat @ 5c26f8b38b67d8c8878c021102cc424a4f8dcb1f |
b72b979c9607984b238ab3e8e414e9576be68716e97c9688423e17dea217208b |
No |
| AnySplat | lhjiang/anysplat @ d2e8c343672646041ad4ea518184968f94362f01 |
1c4de2ba5a29c540b899af901bf02107395b5f0617655d347e262f814b4c0c7c |
No |
Download and verify the ZipSplat checkpoint required by the release pipeline:
hf download veichta/zipsplat zipsplat-da3g-252p.tar \
--revision a0c9c8c1f886d1600e64837c2fcc6be6cc3b3694 \
--local-dir weights
echo "e2906acdfadeb4f7b3311f398ede9e3a2fde734932a1a017c28a28597a5f5cd6 weights/zipsplat-da3g-252p.tar" \
| shasum -a 256 -c -
Checkpoint licenses are separate from source-code licenses. Consult
LICENSES/THIRD-PARTY-NOTICES.md before downloading or using them.
Rebuild the released reconstructions
The executable release path uses ZipSplat and performs inference only. Run the following commands from the bundle root:
conda env create -f environment.yml
conda activate 3droid
pip install -e third_party/ZipSplat
The pipeline requires externally obtained DROID episodes, PointWorld camera files, and the official ZipSplat checkpoint:
export DROID_ROOT=/path/to/droid
export POINTWORLD_CAMERAS=/path/to/pointworld/cameras
export ZIPSPLAT_CHECKPOINT=/path/to/weights/zipsplat-da3g-252p.tar
export PIPELINE_WORK_DIR=/path/to/work
python pipeline/run_pipeline.py all
To rebuild only selected scenes:
python pipeline/run_pipeline.py all --episodes \
AUTOLab__Fri_Jul__7_09-43-39_2023 \
CLVR__Sun_May_21_19-26-15_2023
The reconstruction driver performs four resumable stages:
| Stage | Output |
|---|---|
extract |
First frame from each camera stream |
reconstruct |
Three-view reconstruction and held-out ext2R rendering |
depth |
Four-view reconstruction, PLY, depth, and alpha |
check |
Numerical comparison with the released reconstruction |
The three-view and four-view scenes are different reconstructions. Held-out photometry
uses ext1L, ext1R, and ext2L as input and evaluates ext2R; the released PLY and
geometric measurements use all four exterior views.
Reconstruction verification
run_pipeline.py check treats numerical agreement as the reproduction criterion. It
compares depth_m, alpha, cams, and s_m with their released counterparts. The PLY
SHA-256 is reported separately as a diagnostic because atomic rasterization may not be
byte-identical across hardware.
On the reference environment—a single NVIDIA A100 80GB with CUDA 12.1—all 114 scenes
matched the released depth, alpha, cameras, and scale exactly. The per-scene results are
recorded in reproduction.json.
After model loading and one-time CUDA kernel compilation, reconstruction took a median of 0.77 seconds per scene over the 114 released scenes. See the accompanying paper for the complete compute report.
Rendering the released PLY
gaussians.ply is stored in the model's normalized frame. Construct the render cameras
as described by cameras.json:normalisation:
ref = inv(E[0])
Er = ref @ E
Er[:, :3, 3] /= s_norm
depth.npz:cams instead contains metric camera-to-world extrinsics for back-projection.
Do not use those matrices directly as normalized render cameras.
The reported photometric scores cannot be reproduced by rendering ext2R from the
released four-view PLY because ext2R was an input to that reconstruction. Photometric
evaluation uses the separate three-view reconstruction produced by the reconstruct
stage.
Verify the reported aggregates
The released per-scene measurements can be aggregated without a GPU, network access, or DROID source videos:
pip install numpy scipy
python evaluation/make_tables.py --bundle .
python evaluation/paired_stats.py --bundle .
make_tables.py reproduces the ZipSplat entries in the paper's photometric and
geometric tables, including the 109/110 sample-count split. paired_stats.py reproduces
the ZipSplat paired effect sizes, Wilcoxon p-values, improve/worsen counts, and
difference percentiles. These commands re-aggregate stored measurements; they do not
rerun reconstruction or re-measure geometry.
YoNoSplat and AnySplat serve as auxiliary baselines in the paper. Their code, outputs, and measurements are not included, so their table rows cannot be derived from this bundle.
Known upstream data issues
The release exposes upstream camera defects rather than silently hiding them. Each case
is marked in metrics.json:flags.
- Four scenes contain all-zero DROID intrinsics. Their reconstructions retain those values and are flagged as degenerate. The gate uses the documented serial-median substitution to evaluate the extrinsics, so a gate pass does not certify these reconstructions.
- One scene has no DROID intrinsic entry and uses the median camera matrix from 501 other episodes recorded with the same physical camera serial.
Interpreting the measurements
- Report depth errors together with coverage. An error computed from very few valid queries is not representative of the full scene.
- SGBM depth is a classical stereo reference, not absolute ground truth.
- Robot-referenced comparisons across shipped and refined extrinsics are not independent because the refined calibration uses a related robot-depth signal. Within-source comparisons and held-out photometry avoid this direct cross-source comparison.
License and provenance
3DROID-authored code is released under Apache-2.0. Generated Gaussian, depth, and measurement assets are offered under CC BY-NC 4.0 to the extent that the authors hold the relevant rights, and remain subject to applicable upstream terms.
| Component | Terms |
|---|---|
pipeline/ and evaluation/ authored for 3DROID |
Apache-2.0; see LICENSE |
| Vendored ZipSplat source | Apache-2.0; see third_party/ZipSplat/LICENSE |
| Franka URDF and meshes | Apache-2.0; see pipeline/assets/franka_panda/LICENSE.txt |
| Generated scene and measurement files | CC BY-NC 4.0 plus applicable upstream terms |
| Refined left-eye extrinsics | PointWorld-DROID NVIDIA License |
| Derived right-eye extrinsics | PointWorld terms, with the original within-rig stereo transform composed onto the refined left-eye pose |
| DROID intrinsics and shipped extrinsics | DROID terms |
No DROID images or robot states are redistributed. Refined left-eye extrinsics come
from PointWorld-DROID; the right-eye poses are derived by composing DROID's within-rig
stereo transform. See LICENSES/THIRD-PARTY-NOTICES.md for source revisions,
attribution, and component-specific restrictions.
No new camera calibration was performed for this release.
Citation
If you use 3DROID, please cite the accompanying paper, 3DROID: A Renderable 3D Gaussian Dataset with Measured Per-Scene Reliability. Complete BibTeX will be added after de-anonymization. Please also cite DROID and PointWorld for the redistributed camera metadata.
- DROID: https://droid-dataset.github.io/
- PointWorld: https://github.com/NVlabs/PointWorld
- Downloads last month
- 125