Blame for README.md

---
language:
- en
tags:
- 3d
- depth
- segmentation
- voxel-occupancy
pretty_name: PartProbe3D compact reproduction data
---

# PartProbe3D reproduction data

This release supports SS, generative SLAT segmentation, and 32³ SLAT shape training.
The default transfer is **271.79 GB**, including observations, SS targets, and available shape targets.
Default downloads exclude optional and deprecated folders.

<!-- partprobe3d-required-files:start -->
## Mandatory files

These paths refer to the Hugging Face repository root.
Files are mandatory for the task specified in each row.
The default command restores training files.
The benchmark also requires the separate download with `--dataset evaluation`.

| File or directory | Required task | Contents |
| --- | --- | --- |
| `training/index.json` | Training restoration | Component paths, sizes, monitoring counts, and source identities |
| `training/observations/<dataset>/` | Training on the selected dataset | Archive index, sample NPZ files, calibration, object mappings, and scene voxel targets |
| `training/ss/<dataset>/` | Direct SS training after target publication | Archive index, native SS latents, and codec provenance |
| `training/shape/32/<dataset>/` | SLAT shape training where targets are available | Archive index and object shape targets with explicit exclusions |
| `evaluation/index.json` | Benchmark restoration | Archive list and SHA-256 checksums |
| `evaluation/metadata.tar` | Benchmark evaluation | Portable manifests, source selections, and dataset documentation |
| `evaluation/picasso-*.tar` | Benchmark evaluation | All 100 PICASSO images, masks, previews, and GT meshes |
| `evaluation/messykitchen-*.tar` | Benchmark evaluation | All 1,503 MessyKitchens images, masks, previews, and 100 shared GT meshes |
| `evaluation/3dfront-*.tar` | Benchmark evaluation | The selected 1,000 3D-FRONT test images, masks, previews, and GT meshes |

Restore every archive listed in the selected component's `index.json`.
The evaluation bundle has 33 mandatory archives and requires 29.313 GB.
Older checkouts also require the corresponding restore patch under `code/`.
<!-- partprobe3d-required-files:end -->

<!-- partprobe3d-tasks:start -->
## Download by task

Select a task to avoid unnecessary data transfers.
The [task catalog](tasks/index.json) records component paths, sizes, and the data revision.
The [task guide](tasks/README.md) lists downloads for individual datasets.

| Task | Required data | Default transfer, GB |
| --- | --- | ---: |
| [`eval`](tasks/eval/index.json) | RGB, masks and GT meshes | 29.313 |
| [`ss`](tasks/ss/index.json) | Observations and native SS targets | 265.667 |
| [`slat-seg`](tasks/slat-seg/index.json) | Observations only | 221.950 |
| [`slat-shape`](tasks/slat-shape/index.json) | Observations and 32³ object shape targets | 233.780 |
| [`all-training`](tasks/all-training/index.json) | All published training components | 277.497 |

Use the current project checkout for these commands.
Preview the transfer before downloading:

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task ss --dry-run
```

Select the task required by this server:

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task eval
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task ss
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task slat-seg
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task slat-shape
```

Select individual datasets with `--dataset`, using commas for multiple datasets.

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task eval --dataset picasso
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task eval --dataset picasso,messykitchen
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task slat-seg --dataset foundationpose
```

Task defaults follow the published component inventory.
Explicit shape requests require matching object targets before archive transfer.

SS uses the published native targets without another target encoding pass.
Segmentation computes its instance codes and image features during training.

Online shape conditioning avoids a separate image-feature cache.

Evaluation metadata excludes the project README. Dataset provenance documents remain in the archive.

Model weights remain separate. Download only the assets required by the selected model.

Old commands without `--task` still restore all training components.
<!-- partprobe3d-tasks:end -->

## Restore tools

Use project Git revision `76437eb951160603d0bfd4797212094aea67a4fa` or later.
This revision includes the restore tools and the online SS defaults.

```bash
python -m pip install -e '.[preprocess,ss-multiobj-depth,slat-segvigen,slat-multiobj-shape]'
```

## Default download

Run this command from the patched project root:

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed
```

The command restores observations and SS targets for 3D-FRONT, SAGE-10k, and FoundationPose.
It also restores available 32³ object shape targets.
It verifies each archive and removes temporary archives after restoration.
The downloader pins one Hub revision before reading any archive.
Add `--revision <Hub commit>` to select a fixed release.
Add `--keep-archives` only when you need temporary transfer files.

| Dataset | SS/segmentation training views | Shape training views | Monitoring views | Transfer, GB |
| --- | ---: | ---: | ---: | ---: |
| 3dfront | 39,930 | 39,930 | 16 | 123.01 |
| sage | 17,456 | 17,453 | 16 | 78.73 |
| foundationpose | 42,256 | 42,256 | 16 | 75.75 |

The default includes FoundationPose for SS and segmentation.
FoundationPose shape targets remain unavailable.
Use this command to restore FoundationPose alone:

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset foundationpose
```

## Contents

Sample NPZ files retain RGB, metric depth, visible labels, complete object occupancy, and the original transforms.
Calibration, scene voxel targets, object mappings, and original extent manifests remain available.
The compact export omits separate RGB, masks, depth copies, overlays, and duplicate voxel exports.
Every training view remains available to SS and segmentation.
The first 16 test views per dataset support existing eight-view and sixteen-view training monitors.
All other test observations require an optional download.

Object shape arrays retain their original values and identities.
Shape manifests explicitly exclude the three failed SAGE views used by the existing experiment.
SS and segmentation retain those views.
The export omits combined scene latents because the current shape objective uses object targets.
Model weights and image-feature caches remain separate dependencies.

## Optional files

These files support additional datasets, test views, and target generation.
The mandatory evaluation bundle already includes its RGB images, masks, and GT meshes.
Some SS configurations require the matching optional occlusion component.

| File or directory | Purpose | Transfer, GB |
| --- | --- | ---: |
| `optional/evaluation/ycbv/` | YCB-V mesh evaluation | 1.670 |
| `optional/geometry/3dfront/` | Mesh scoring and new target encoding | 6.252 |
| `optional/geometry/sage/` | Mesh scoring and new target encoding | 303.693 |
| `optional/held_out/3dfront/` | Remaining test observations | 22.440 |
| `optional/held_out/foundationpose/` | Remaining test observations | 6.825 |
| `optional/held_out/sage/` | Remaining test observations | 8.183 |
| `optional/occlusion/3dfront/` | GT relations for SS object attention | 0.240 |
| `optional/occlusion/sage/` | GT relations for SS object attention | 0.103 |

Restore observations before geometry, occlusion, or test overlays.
YCB-V evaluation includes its own observations and geometry.
Select each optional component explicitly:

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset sage --optional geometry
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset 3dfront --optional occlusion
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset sage --optional held-out
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset ycbv --optional evaluation
```

## Train all three stages

Obtain the configured model weights separately.
The restored SS targets support online SS training directly.

```bash
export CUDA_VISIBLE_DEVICES=4,6
export CUDA_DEVICE_ORDER=PCI_BUS_ID
python -m ss_multiobj_depth.train
python -m slat_segvigen.train --config slat_segvigen/configs/object_mean_image_nora_coverage_gate.json
python -m preprocess.slat.prepare --config preprocess/slat/configs/front3d_sage_32_assigned_object.json --stage features --devices cuda:0 cuda:1
python -m slat_multiobj_shape.train --config slat_multiobj_shape/configs/front3d_sage_32_assigned_object.json
```

Use the saved configuration and pretrained assets for a specific experiment.
Feature preparation creates local caches without meshes or another dataset download.
For lower local storage, generate an online shape configuration:

```bash
python -m preprocess.reproduction configure-shape --output outputs/reproduction/shape_online.json
python -m slat_multiobj_shape.train --config outputs/reproduction/shape_online.json
```

The generated configuration computes image conditioning during training and avoids separate image-feature caches.
It preserves the selected preset's model and optimization settings.
Use the original cached preset when reproducing a cached-feature experiment.
Data and decoder provenance remain in the restored manifests.
Exact checkpoint resume requires the original matching feature and data identities.

## Deprecated uploads

`_deprecated/20261020/` contains superseded bundles and catalogs.
The project downloader never selects these paths.
An unfiltered Hugging Face snapshot download includes every folder.
Use the project downloader for minimal transfer.
The [original release](https://huggingface.co/datasets/cubec/PartProbe3D-preprocessed/tree/d13e577a7e71291ed94fa40e4297f54106577fbd) remains available by revision.

## Source terms and previews

Source dataset terms continue to apply to these derived observations.
The prepared data does not replace access agreements for original source assets.

![3D-FRONT RGB](previews/3dfront/train-rgb.png)

![3D-FRONT visible masks](previews/3dfront/train-overlay.png)

[Posed object geometry](previews/3dfront/train-geometry.glb)

<!-- partprobe3d-frozen-evaluation:start -->
## Required evaluation dataset

The required [evaluation dataset](evaluation/README.md) contains 2,603 images with visible masks and GT object meshes.
It includes all 100 PICASSO images, all 1,503 MessyKitchens images, and 1,000 random 3D-FRONT test views.
The PartProbe3D benchmark requires this dataset.
Benchmark inference and scoring require geometry meshes.
Model comparisons preserve native geometry and omit independent texture generation.
The transfer is 29.313 GB.
MessyKitchens shares 100 GT scene meshes across its views.
MessyKitchens uses reviewed GT-assisted SAM3 masks for all 1,503 views.
The bundle retains exact independent masks, registration records, and all 13 audit flags.
Native SAM3 quality resolves overlap for evaluation labels, with smaller object IDs breaking ties.
See the [mask release](https://github.com/cube-c/PartProbe3D/blob/main/reports/messykitchen_registered_release_20261024.md) for the pinned revision and migration paths.
Default training downloads exclude this bundle.

Use the current project checkout for the task commands.

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task eval
```
<!-- partprobe3d-frozen-evaluation:end -->

<!-- partprobe3d-ss-targets:start -->
## Encoded SS targets

The standard download includes encoded SS targets for the datasets listed below.
These targets remove GPU encoding from the restore procedure.
They retain the original cache manifests, latent tensors, and monitoring targets.
Their native latent shape is `8×16×16×16`.
Model weights remain separate dependencies.

| Dataset | Target views | Transfer, GB |
| --- | ---: | ---: |
| 3dfront | 49,996 | 31.249 |
| sage | 19,408 | 4.175 |
| foundationpose | 46,948 | 8.292 |

Use the current project checkout and download the selected datasets:

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task ss
```

SS training can then use the default online configuration without a preparation command.
The downloader checks archive hashes before extraction.
The training loader checks dataset and frozen codec identities.
The image encoders compute features during training.
<!-- partprobe3d-ss-targets:end -->

<!-- partprobe3d-scenes:start -->
## Voxel-aligned scene latents

These optional caches contain one token per saved reference voxel.
Overlapping objects share that token and retain separate membership labels.
The `latent_valid` flag identifies cells with a native encoder output.
Invalid cells retain their reference coordinates and contain zero latent values.
The archives preserve the original NPZ files, calibration transforms, and codec provenance.

| Dataset | Grid | Training views | Transfer, GB |
| --- | ---: | ---: | ---: |
| [3dfront](optional/scene/32/3dfront/index.json) | 32³ | 39,930 | 4.590 |
| [sage](optional/scene/32/sage/index.json) | 32³ | 17,456 | 1.176 |
| [foundationpose](optional/scene/32/foundationpose/index.json) | 32³ | 42,256 | 5.046 |
| [3dfront](optional/scene/64/3dfront/index.json) | 64³ | 39,778 | 19.744 |
| [sage](optional/scene/64/sage/index.json) | 64³ | 17,365 | 4.252 |
| [foundationpose](optional/scene/64/foundationpose/index.json) | 64³ | 42,223 | 19.809 |
| [omni6dpose](optional/scene/64/omni6dpose/index.json) | 64³ | 44,788 | 5.469 |

Restore the matching observation bundles before reading scene targets.
The `slat-seg` task downloads those observations without SS or object latent targets.

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task slat-seg
```

Restore the selected scene caches:

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset 3dfront --bundle-root optional/scene/32 --cleanup
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset sage --bundle-root optional/scene/32 --cleanup
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset foundationpose --bundle-root optional/scene/32 --cleanup
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset 3dfront --bundle-root optional/scene/64 --cleanup
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset sage --bundle-root optional/scene/64 --cleanup
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset foundationpose --bundle-root optional/scene/64 --cleanup
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset omni6dpose --bundle-root optional/scene/64 --cleanup
```

The restored caches reside under `datasets/<dataset>/slat/<resolution>/scene_targets/`.
`dataloader.slat_scene.SceneTargetDataset` reads them with their matching saved observations.
Raw meshes and source datasets are unnecessary for this reader.
Scene downloads remain separate from object targets and default task downloads.
<!-- partprobe3d-scenes:end -->

## Deprecated MessyKitchens masks

Historical MessyKitchens mask archives now reside under [_deprecated/messykitchen/original_masks/evaluation](https://huggingface.co/datasets/cubec/PartProbe3D-preprocessed/tree/main/_deprecated/messykitchen/original_masks/evaluation).
Current evaluation downloads exclude these archives.

<!-- partprobe3d-omni6dpose:start -->
## Omni6DPose SOPE

This release contains 3 selected frames per SOPE sequence.
It contains 55,848 usable views from 18,617 sequences.
The visibility requirement excludes 3 source views.
The official splits contain 44,788 training views and 11,060 test views.
All selected frames use one dataset and the same component paths.

Every retained mask covers strictly more than 0.1% of the native image.
At 640 by 480 pixels, each retained object covers at least 308 pixels.
Each view contains at least two retained objects.
The observation, object, SS, and scene components use this selection.

Native RGB, perfect metric depth, instance labels, and calibration come from SOPE.
The observation bundle stores these arrays in sample NPZ files.
Targets contain surface voxel occupancy at resolution 64.
PointCloudExtent uses near and far trim fractions of 0.0525, far margin 0.125, and padding 1.31.
Object latents and scene latents use resolution 32.
SS targets contain native latents with shape `8×16×16×16`.

| Component | Views | Transfer, GB |
| --- | ---: | ---: |
| [training/observations/omni6dpose](training/observations/omni6dpose/index.json) | 44,804 | 70.945 |
| [training/shape/32/omni6dpose](training/shape/32/omni6dpose/index.json) | 44,788 | 1.739 |
| [training/ss/omni6dpose](training/ss/omni6dpose/index.json) | 55,848 | 145.935 |
| [optional/held_out/omni6dpose](optional/held_out/omni6dpose/index.json) | 11,044 | 17.792 |
| [optional/geometry/omni6dpose](optional/geometry/omni6dpose/index.json) | 55,848 | 12.039 |
| [optional/scene_mask_area/32/omni6dpose](optional/scene_mask_area/32/omni6dpose/index.json) | 44,788 | 1.573 |

The core observations include all training views and 16 test views for monitoring.
The held-out component adds the remaining 11,044 test views.
The SS component contains all 55,848 views.
Optional geometry contains canonical GLBs and per-frame instance transforms.
Filtered scene latents occupy the same namespace as the other filtered datasets.

Restore all training components with this command:

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --task all-training --dataset omni6dpose --cleanup
```

Restore optional components with these commands:

```bash
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset omni6dpose --bundle-root optional/held_out --cleanup
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset omni6dpose --bundle-root optional/geometry --cleanup
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset omni6dpose --bundle-root optional/scene_mask_area/32 --cleanup
```

Use project Git revision `513895b85af6e56716e1ed51d9352639a27e8e8d` or a later revision.
The committed code patch is available at [code/omni6dpose.patch](code/omni6dpose.patch).
Apply this patch to project Git revision `32d92a1e1b9e99eafaac9dda13c43e676a7aeeb3` when necessary.
Model weights remain separate from this dataset.
Component manifests preserve model hashes, source hashes, and generation settings.

The shape encoder is `TRELLIS.2 shape_enc_next_dc_f16c32_fp16`.
Its safetensors SHA-256 is `f37c5ff5b983b68e9946060000f09bc131f3e84318a2c8b7430a81e4b4636c41`.
The paired Pixal3D decoder SHA-256 is `e3b718d3e43e4f8780e9a24ac6fff231811a67e3b058e336e10fe654c911d581`.

Preview: [RGB, masks, extent, meshes, and voxels](previews/omni6dpose/extent_voxels.png).
The [release inventory](releases/omni6dpose/index.json) records source identities and component checksums.
Dataset credit: [Omni6DPose](https://jiyao06.github.io/Omni6DPose/).
Source documentation: [Omni6DPoseAPI](https://github.com/Omni6DPose/Omni6DPoseAPI).
<!-- partprobe3d-omni6dpose:end -->