cubec/PartProbe3D-preprocessed

PartProbe3D preprocessed datasets

Saved 64³ training and evaluation inputs for PartProbe3D preserve calibrated observations and voxel targets. The 3D-FRONT, SAGE-10k, and YCB-V bundles also retain posed meshes. FoundationPose contains an export for training without meshes. Targets describe triangle surfaces, including hidden surfaces and overlapping object channels.

Dataset	Split	Views	Scenes
sage	test	1,952	1,000
sage	train	17,456	8,931
3dfront	test	10,066	935
3dfront	train	39,930	3,703
ycbv	test	900	900
foundationpose	train	42,256	42,256
foundationpose	test	4,692	4,692

3D-FRONT uses the pinned last 1,000 MIDI rooms for testing and all other rooms for training; all views of a room stay together. There is no validation split. YCB-V is evaluation only. The viewer contains metadata indices; binary training artifacts are in the archives.

Upload status

Dataset	Packaging complete	Upload complete	Packaged GiB	Transferred GiB
sage	True	True	475.80	475.80
3dfront	True	True	273.53	273.53
ycbv	True	True	3.60	3.60
foundationpose	True	True	83.19	83.19

SAGE-10k keeps every room and view of a source layout in the same split, using the saved SHA-256 layout split with a 10% test fraction and no validation split. The completed extent dataset retains 19,408 training/evaluation views; 592 source captures were rejected by target preparation.

Archive inventories and SHA-256 checksums are in bundles/<dataset>/index.json. The extent archives exclude diagnostic subsets, raw downloads, model weights, and model-specific feature caches. The FoundationPose export contains 46,948 views in 86 verified archives.

FoundationPose on another server

Use code revision 92ca3ee or a compatible later revision on slat-shape. The FoundationPose bundle preserves original RGB, depth, visible masks, cameras, object mapping, and complete per-object voxel occupancy at 64³. Its source manifest has SHA-256 2c4dcd1cd1e2e56a0eb08f2f92da320f299729c8ce7a2f52828c272cbc1fa0ad. The extracted files occupy 87.82 GB before filesystem overhead. The verified inventory contains 86 archives totaling 89.33 GB. The export excludes raw archives, meshes, model weights, image features, SLAT latents, and GT occlusion relations.

The loaders resolve input paths relative to the restored dataset. Historical absolute paths remain provenance entries. SS preparation creates the union of complete occupancy for objects with supplied masks and excludes other object voxels.

Restore the data

Use Python 3.12 and a compatible PyTorch/CUDA installation. Run these commands from the repository root.

1. Install the project dependencies.

   python -m pip install -e '.[preprocess,ss-multiobj-depth,tracking]'
   

2. Restore the prepared observations and voxel targets.

   python -m preprocess.huggingface download \
     --repo cubec/PartProbe3D-preprocessed --dataset foundationpose --cleanup \
     --revision 073988c5b5e206570da67f57aada4a2f0ad94e95
   

3. Initialize the pinned Pixal3D source.

   git submodule update --init third_party/Pixal3D
   

--cleanup retains extracted data and removes each verified temporary archive. Allow at least 100 GB for data restoration, plus space for weights and model features. Use --revision <Hub commit> to retain one dataset snapshot.

Acquire SS weights

All required weights for the following SS configuration are available from Hugging Face. The commands pin revisions with the same configuration and weight bytes as the verified local files. DINOv3 requires approved access to its gated model repository.

1. Authenticate with an account that can download DINOv3.

   hf auth login
   

2. Download the native SS flow and decoder from Pixal3D.

   hf download TencentARC/Pixal3D \
     --revision b0cb2e1b794cab9aa0ac38a95d794a4d9337437f \
     --local-dir checkpoints/Pixal3D \
     --include 'ckpts/ss_flow_img_dit_1_3B_64_bf16.*' 'ckpts/ss_dec_conv3d_16l8_fp16.*'
   

3. Download the SS encoder and paired decoder from TRELLIS.

   hf download microsoft/TRELLIS-image-large \
     --revision 25e0d31ffbebe4b5a97464dd851910efc3002d96 \
     --local-dir checkpoints/TRELLIS \
     --include 'ckpts/ss_enc_conv3d_16l8_fp16.*' 'ckpts/ss_dec_conv3d_16l8_fp16.safetensors'
   

4. Download DINOv3.

   hf download facebook/dinov3-vitl16-pretrain-lvd1689m \
     --revision ea8dc2863c51be0a264bab82070e3e8836b02d51 \
     --local-dir checkpoints/DINOv3 --include config.json model.safetensors
   

Prepare features and train SS

The FoundationPose configuration uses GT depth, depth_embedding voxel conditioning, global cross-attention, and rank-8 NoRA. It creates features under outputs/ss_multiobj_depth/foundationpose/features and preserves the saved training/test splits. It starts a new run with W&B online by default. Online logging requires W&B authentication. The setting training.wandb_mode="disabled" permits a local run without W&B.

1. Prepare local features and encoded SS targets.

   python -m ss_multiobj_depth.prepare --config ss_multiobj_depth/configs/foundationpose.json
   

2. Start training.

   python -m ss_multiobj_depth.train --config ss_multiobj_depth/configs/foundationpose.json
   

This configuration needs no source archives, meshes, or additional dataset downloads. The occlusion_object variant additionally requires a matching GT occlusion cache. That cache requires mesh geometry during preparation and is absent from this FoundationPose bundle. A matching GT occlusion overlay can transfer separately without another download of the observation archives. New SLAT shape encoding and mesh-based evaluation also require geometry or separately transferred targets. Exact resume of an existing run requires its matching checkpoint, feature manifests, source hashes, and settings.

FoundationPose previews

These previews show saved inputs and GT voxel targets. They use no learned checkpoint.

GT occupancy, train (GLB)

GT occupancy, test (GLB)

The publication report records checks, dataset identities, and limitations.

GT occlusion relations

Separate archives under precomputed_occlusion/ preserve the completed GT occlusion caches. The caches contain directed [front, back] relations from calibrated visible masks and posed GT mesh silhouettes. Each archive preserves the original matrices, completion records, and manifest bytes.

Dataset	Training views	Test views	Archives	Size, MB
3D-FRONT	39,930	10,066	3	240.01
SAGE-10k	17,456	1,952	3	103.45
Total	57,386	12,018	6	343.46

MB means decimal megabytes. The inventories at precomputed_occlusion/<dataset>/index.json record archive sizes and SHA-256 checksums. They also preserve the source policy and checksum of the matching extent manifest.

Download GT occlusion caches

Restore the matching extent datasets before these overlays. Run these commands from the repository root.

python -m preprocess.huggingface download \
  --repo cubec/PartProbe3D-preprocessed --bundle-root precomputed_occlusion \
  --dataset 3dfront --revision 88cdf8f0e880a33e7cafa580ee9722776f16aa28 --cleanup
python -m preprocess.huggingface download \
  --repo cubec/PartProbe3D-preprocessed --bundle-root precomputed_occlusion \
  --dataset sage --revision 88cdf8f0e880a33e7cafa580ee9722776f16aa28 --cleanup

The commands restore datasets/<dataset>/extent/point_cloud/occlusion/. The downloader verifies each archive and refuses to replace different existing files. The dataset reader checks the extent manifest identity before reading GT relations.

SS training with occlusion-aware object cross-attention reads these relations. SLAT shape tokens already have instance identities and require no occlusion matrix. Frozen image features and SLAT shape latents remain separate artifacts.

Publish GT occlusion caches

Package the verified caches from the prepared extent roots:

python -m preprocess.huggingface pack --artifact-kind gt-occlusion \
  --dataset 3dfront=datasets/3dfront/extent/point_cloud \
  --dataset sage=datasets/sage/extent/point_cloud \
  --output outputs/huggingface/gt_occlusion
python -m preprocess.huggingface upload \
  --repo cubec/PartProbe3D-preprocessed --bundle-root precomputed_occlusion \
  --dataset 3dfront --dataset sage --output outputs/huggingface/gt_occlusion

Packaging checks every matrix checksum and its completion record. The uploader publishes complete inventories after it verifies all remote archives.

SLAT shape: upload and download specification

The repository contains prepared extent datasets and separate SLAT-shape targets. Combined 32³ SLAT finetuning requires both artifact types for both training datasets. The 6.84 GB SLAT upload contains latent targets and metadata. It excludes RGB observations, depth maps, masks, and frozen image features.

The sizes and coverage below describe data snapshot 0e4623fb0c8d9bad7cbe1dfb161a993ac165b6aa. GB means decimal gigabytes.

Bundle	Contents	Training archives and metadata, GB	All archives, GB
bundles/3dfront	Calibrated observations, extent geometry, and occupancy targets	234.67	293.70
bundles/sage	Calibrated observations, extent geometry, and occupancy targets	460.59	510.88
precomputed_slat/3dfront	32³ object latents and available scene latents	5.57	5.57
precomputed_slat/sage	32³ object latents and recorded failures	1.27	1.27
Total	Both datasets and both artifact types	702.10	811.42

Training requires each extent bundle's metadata.tar and train-*.tar archives. Evaluation also requires its test-*.tar archives. The current download command restores every archive in a bundle, including available test archives. It has no split filter. The four commands below therefore download about 811.42 GB.

Download and restore

Use the updated PartProbe3D checkout with latent generation in preprocess/slat/. Install its preprocessing dependencies and huggingface-hub. Authenticate with a Hugging Face account that can access this dataset repository. Run these commands from the repository root.

1. Restore the prepared 3D-FRONT extent dataset.

   python -m preprocess.huggingface download \
     --repo cubec/PartProbe3D-preprocessed --bundle-root bundles \
     --dataset 3dfront --revision 0e4623fb0c8d9bad7cbe1dfb161a993ac165b6aa --cleanup
   

2. Restore the prepared SAGE-10k extent dataset.

   python -m preprocess.huggingface download \
     --repo cubec/PartProbe3D-preprocessed --bundle-root bundles \
     --dataset sage --revision 0e4623fb0c8d9bad7cbe1dfb161a993ac165b6aa --cleanup
   

3. Restore the 3D-FRONT shape targets.

   python -m preprocess.huggingface download \
     --repo cubec/PartProbe3D-preprocessed --bundle-root precomputed_slat \
     --dataset 3dfront --revision 0e4623fb0c8d9bad7cbe1dfb161a993ac165b6aa --cleanup
   

4. Restore the SAGE-10k shape targets.

   python -m preprocess.huggingface download \
     --repo cubec/PartProbe3D-preprocessed --bundle-root precomputed_slat \
     --dataset sage --revision 0e4623fb0c8d9bad7cbe1dfb161a993ac165b6aa --cleanup
   

Dataset	Prepared observations and extent targets	SLAT targets
3D-FRONT	datasets/3dfront/extent/point_cloud/	datasets/3dfront/slat/32/targets/
SAGE-10k	datasets/sage/extent/point_cloud/	datasets/sage/slat/32/targets/

The updated downloader maps historical SLAT archive paths under outputs/ to these dataset locations. It preserves target bytes and verifies every archive's size and SHA-256 checksum. It refuses to replace different existing files. --cleanup retains one temporary archive at a time and removes it after successful restoration. Temporary archives reside under datasets/<dataset>/archives/huggingface/. Extracted files and future feature caches need additional disk space.

Preserve the original extent manifests without edits. The SLAT manifests bind their source datasets through manifest checksums. Raw dataset archives, rendering again, and SS feature caches are unnecessary for this workflow.

Complete preparation before training

The published targets cover 39,930 3D-FRONT training views and 17,456 SAGE-10k training views. Three SAGE-10k views retain one failed object target each. These failures block strict combined training until object encoding succeeds. The snapshot contains scene latents for 6,720 3D-FRONT views and no SAGE-10k scenes. The current object loss does not require completed scene encoding.

1. Obtain the configured Pixal3D, DINOv3, and NAF weights separately.
2. Resolve the three failed SAGE-10k object targets.
3. Prepare frozen conditioning features and immutable object targets.
4. Start finetuning after feature preparation succeeds.

The TRELLIS.2 shape encoder and paired decoder are also required when regenerating shape targets. Image features and immutable training snapshots are not included in the uploaded target bundles. The commands below do not automatically schedule scene encoding or training during feature preparation.

python -m preprocess.slat.prepare \
  --config preprocess/slat/configs/front3d_sage_32.json \
  --stage features --devices cuda:1 cuda:2
python -m slat_multiobj_shape.train \
  --config slat_multiobj_shape/configs/front3d_sage_32.json

Feature preparation writes object_targets/ and object_features/ under each dataset's slat/32/ directory. It retains clean GT depth for fresh augmentation during training. Testing uses MoGe-3 depth and calibration. Training results reside under outputs/, and model checkpoints reside under checkpoints/.

Upload contract

bundles/<dataset>/index.json describes prepared extent archives. precomputed_slat/<dataset>/index.json describes SLAT target archives. Each inventory records archive sizes, SHA-256 checksums, and the restore destination. SLAT inventories also record source manifests, codec provenance, coverage, and encoding failures.

An inventory's complete=true means packaging covers the snapshot files. The uploader publishes a complete inventory after every archive is present with matching size. It also compares remote SHA-256 checksums when Hub file metadata provides them. source_complete separately records whether source preprocessing finished. An upload can therefore finish while source object or scene encoding remains incomplete. The upload preserves failure records and pending scene statuses.

Package shape targets with this command:

python -m preprocess.huggingface pack \
  --artifact-kind slat-shape-targets \
  --dataset 3dfront=datasets/3dfront/slat/32/targets \
  --dataset sage=datasets/sage/slat/32/targets \
  --output outputs/huggingface/slat_targets_publish

Publish those archives to an existing dataset repository with this command:

python -m preprocess.huggingface upload \
  --repo cubec/PartProbe3D-preprocessed \
  --bundle-root precomputed_slat --dataset 3dfront --dataset sage \
  --output outputs/huggingface/slat_targets_publish

The target upload excludes worker jobs, raw assets, model weights, and frozen image feature caches. Use a fresh packaging directory when source manifests change. Publish matching extent and target snapshots before documenting a new reproducible data revision.

Restore on another server

Use the PartProbe3D checkout containing preprocess/huggingface.py and install its preprocessing dependencies plus huggingface-hub. Run from the repository root:

python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset sage --cleanup
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset 3dfront --cleanup
python -m preprocess.huggingface download --repo cubec/PartProbe3D-preprocessed --dataset ycbv --cleanup

Downloads resume through Hugging Face. Restoration verifies every archive SHA-256 and refuses to replace different existing data. --revision <commit> pins a release. The extracted roots are datasets/<dataset>/extent/point_cloud. --cleanup downloads and restores one archive at a time, deleting each only after checksum verification and successful extraction. Temporary archives use datasets/<dataset>/archives/huggingface/. Without --cleanup, all downloaded archives are retained for reuse.

The loaders resolve sample, camera, geometry and target paths relative to the dataset root. Historical absolute file and raw asset paths remain provenance; loading saved samples and geometry does not require those paths or raw assets.

Prepare the model-specific features on the destination server, then train:

python -m ss_multiobj_depth.prepare --config ss_multiobj_depth/configs/front3d.json
python -m ss_multiobj_depth.train --config ss_multiobj_depth/configs/front3d.json

The mixed 32³ SLAT training config reads the saved 64³ datasets and pools grids online:

python -m slat_segmentation.train --config slat_segmentation/configs/front3d_sage_32.json

Provenance

Original manifests, artifact completion markers, checksums, identities and split policies are retained byte-for-byte inside the archives. Generation settings and code hashes are also available under metadata/<dataset>/conversion_policy.json. The extent upload does not include model weights. SLAT target inventories record the frozen encoder provenance. Source datasets retain their original licenses and access terms.

Input and geometry previews

sage: test

View: 0078d732045c32e0e3c2.

Posed prompted GT geometry (GLB)

sage: train

View: 0002c10ce97a86f789fd.

Posed prompted GT geometry (GLB)

3dfront: test

View: 0001acce61c945bf4470.

Posed prompted GT geometry (GLB)

3dfront: train

View: 0000b002995b4131d534.

Posed prompted GT geometry (GLB)

ycbv: test

View: 003c09039641a2b2efa1.

Posed prompted GT geometry (GLB)