RobertRussell/Articulation-XL2-human-pose-augmented

Human pose augmentation for Articulation XL2

This dataset contains historical human-focused training variants and the associated base, validation, and test collections derived from Seed3D Articulation XL2.0. They preserve the original NumPy object-array format for compatibility with Puppeteer-style skeleton training.

Files

File	Samples	File size	Contents
train_human_final_augmented.npz	69,534	38,226,475,998 bytes	Historical augmented variant; count equals the 34,767-sample base plus 34,767 additions
train_human_final_hard4_aug.npz	173,835	70,880,586,260 bytes	Base plus four hard pose augmentations per base sample
train_human.npz	32,855	25,073,678,050 bytes	Earlier human training collection; not assumed identical to the final base
train_human_final.npz	34,767	27,340,400,445 bytes	Final unaugmented training base
val_human.npz	3,864	2,948,658,404 bytes	Preserved historical validation collection
test_human.npz	1,836	1,265,922,892 bytes	Preserved historical test collection

Both variants include the same base training collection. They are alternative training variants, not disjoint dataset splits; concatenating them repeats base examples. The four additional NPZs preserve historical split files without reconstructing or reshuffling them. File names identify their intended roles; UUID-level disjointness and overlap between earlier and final training versions have not been newly audited. Do not concatenate the earlier training collection with the final base as though they were independent data.

Processing

The base collection was constructed by filtering human-related categories from the upstream main and diverse-pose collections and forming a training split. Pose augmentation uses skeleton-joint rotations, forward kinematics, linear blend skinning, and resampled point clouds. The upstream mesh and rig fields are retained.

The hard variant uses difficulty=hard and four augmentations per sample. The current generator's default seed is 42. Exact historical dependency versions are not recorded. For the older augmented variant, the sample count is known but its exact historical augmentation settings are not fully recorded. Do not assume it is a strict subset of the hard variant.

Loading

import numpy as np

with np.load("train_human_final_augmented.npz", allow_pickle=True) as archive:
    samples = archive["arr_0"]

The files contain pickled Python objects: only load trusted files. Accessing arr_0 loads a large object array; compressed NPZ object arrays do not provide effective memory-mapped sample access. Use a sufficiently large-memory machine. A tabular dataset viewer is not expected for this format.

Attribution

The original data and rig annotations are credited to the authors of MagicArticulate and Puppeteer, including Chaoyue Song and collaborators. This repository adds human-focused selection and pose augmentation; it is not an official Seed3D release. The upstream dataset declares CC BY 4.0. Retain upstream attribution and applicable underlying asset notices when reusing or redistributing the data.

Augmentation code and recipes

The original augmentation implementation is included in scripts/augment_poses.py. Only machine-specific example paths and the default visualization directory have been made portable; the augmentation algorithm is unchanged. It supports both dataset generation and headless comparison PNGs. Dependencies are listed in scripts/requirements.txt; no GPU is required.

The implementation applies these steps to each eligible rigged sample:

1. Sample a random rotation axis and an angle uniformly between zero and a depth-dependent maximum. Base maximum angles are 5°, 20°, 45°, and 75° at depths 0, 1, 2, and 3 or greater. Difficulty multiplies these by 0.5, 1.0, or 1.5 for easy, medium, or hard. Disconnected joints receive no rotation.
2. Update joint positions by forward kinematics and deform mesh vertices with linear blend skinning using the existing sparse skinning weights.
3. Recenter and uniformly scale the posed mesh to maximum absolute coordinate 0.5; apply the same transform to the posed joints.
4. Sample 8,192 surface points and face normals. Recenter and uniformly scale point coordinates to maximum absolute coordinate 0.9995; normalize normals and store the six point features as float16.
5. Preserve original records first, then append successful augmentations in source-sample order. Original topology, skinning data, and other fields are retained; joints, vertices, point cloud, and UUID are updated. Failed augmentations are counted and skipped, so output counts can vary on other inputs.

For hard4, the maximum rotation angles are 7.5°, 30°, 67.5°, and 112.5°. The command below implements its recorded difficulty and augmentation count:

python scripts/augment_poses.py \
  --input_npz train_human_final.npz \
  --mode augment --difficulty hard --augments_per_sample 4 --seed 42 \
  --output_npz regenerated_hard4.npz

Seed 42 is the available script's default, not a verified historical run seed. The exact historical dependency versions are unavailable; point sampling can also depend on the trimesh version. This is the generation method, not a promise of byte-identical regeneration. The older augmented artifact has one additional record per base record by count, but its exact difficulty is unspecified. recipes.json keeps these facts separate from default settings.

The separate train_human_final.npz is now distributed directly. The generator also keeps the original base collection as the first 34,767 records. As an alternative to downloading the separate base, recover that prefix from augmented file with:

python scripts/extract_base.py \
  --input_npz train_human_final_augmented.npz \
  --output_npz recovered_train_human_final.npz

Extraction loads the entire object array into RAM. The uncompressed arr_0 member is about 75.1 GB for augmented and 121.2 GB for hard4; Python object overhead and generation buffers require additional memory. Extraction does not reconstruct the original upstream category-filtering or train/validation split. Both augmented variants contain related versions of the same base objects; split by source object, not by individual augmentation, when constructing new train/test partitions.

Sample fields and integrity

arr_0 is a one-dimensional object array of sample dictionaries. The fields used by the augmentation code are:

Field	Layout	Meaning
uuid	string	Original object ID, or original ID plus _aug0, _aug1, etc.
joints	(J, 3)	Joint positions; regenerated values are float64
bones	(B, 2)	Parent–child joint index pairs
root_index	scalar	Root joint index
vertices	(V, 3)	Mesh vertices; regenerated values are float64
faces	(F, 3)	Triangle vertex indices, retained from the source
pc_w_norm	(8192, 6) for regenerated samples	XYZ coordinates followed by normal XYZ, float16
skinning_weights_value	sparse entries	Skinning weights
skinning_weights_row / skinning_weights_rows	sparse row indices	Vertex indices
skinning_weights_col / skinning_weights_cols	sparse column indices	Joint indices
skinning_weights_shape	(V, J)	Sparse skinning matrix shape

Original records retain their source dtypes and point counts. Fields not updated by augmentation are copied unchanged; do not assume every ancillary source field was recomputed for the posed mesh. UUID augmentation suffixes are local to each source record, not globally unique IDs across variants or repeated source IDs. An additional allow_pickle archive member is a historical NumPy save artifact, not another dataset split. Always read arr_0 explicitly.

dataset_metadata.json records sizes, counts, schema, and SHA-256 hashes. SHA256SUMS can check downloaded NPZs:

sha256sum -c SHA256SUMS

Source revision and transformation details are recorded in provenance.json.

Selected historical checkpoint backups, their full-validity records and preview images are available in Puppeteer human skeleton selected. The 1,836-sample and older 1,412-sample evaluation cohorts are reported separately.