SilHand Data

Processed data for Hand in Silhouette: Synthetic-to-Real Multi-View Hand Reconstruction via Binary Masks.

This package contains the synthetic hand geometry, preprocessed real-world datasets, and camera bank used by SilHand. It preserves the original packed rows, split indices, and view selections. Geometry and calibration are stored as NumPy arrays; Clean input masks are rendered from the reference geometry at inference time, retaining object occlusion.

Contents

Directory	Data	Train	Validation	Test
synthetic_pca/	Synthetic free-hand poses	60,000	0	6,000
synthetic_dexonomy/	Synthetic grasping poses	46,775	5,197	0
arctic/	ARCTIC-MV	135,469	0	17,393
dexycb/	DexYCB-MV	25,387	1,412	4,951
ho3d/	HO3D-MV	9,087	0	2,706
interhand/	InterHand-MV	210,021	0	85,255
oakink/	OakInk-MV	14,673	0	4,978

camera_bank/ contains 210,100 camera rigs and their sampling weights for synthetic observations. The package contains 629,304 hand samples in total and occupies approximately 11.87 GB before compression.

SilHand was trained on the 106,775 synthetic training samples. The five real-world test splits contain 115,283 samples. Real-world train/validation splits are also preserved in this data release; the accompanying inference code reads the real-world test splits.

Use with SilHand

Place the directories and files from this package under silhand/data/, so that paths such as data/ho3d/test_indices.npy and data/oakink/vertices.npy exist. Follow the code repository's README for CUDA/DRTK installation and the separate MANO download.

From the code repository root, run the five Clean benchmarks:

for dataset in dexycb ho3d oakink arctic interhand; do
  CUDA_VISIBLE_DEVICES=0 python infer.py \
    "datamodule.real_dataset_dirs=[data/${dataset}]" \
    test.save_predictions=false \
    "log_dir=outputs/paper_clean/${dataset}" || break
done

Array format

All indices are zero-based. Coordinates and translations are in meters, and all hand samples represent right hands. Matrices are flattened in row-major order. N is the number of packed hand samples and V the stored number of camera slots.

File	Shape	Meaning
vertices.npy	[N, 778, 3]	Hand vertices in world coordinates
joint_frames_12d.npy	[N, 21, 12]	Joint-to-world frames: translation first, then a flattened 3×3 rotation
amano_hand_params.npy	[N, 43]	21 anatomical pose parameters, 10 shape parameters, and a 12D wrist-to-world transform
face_data.npz	Named arrays	Hand triangle indices, wrist-cap face indices, and optional face labels
<split>_indices.npy	[N_split]	Packed row IDs for train, val, or test, in split order
camera_params.npy	[N, V, 25]	Flattened 4×4 world-to-camera extrinsics followed by 3×3 intrinsics
source_image_sizes.npy	[N, V, 2]	Full-frame image height and width
camera_valid_mask.npy	[N, V]	Valid camera slots
<split>_view_mask.npy	[N_split, V]	Views selected for each row of the corresponding split indices

Synthetic geometry uses the separate camera bank. Its arrays have the same camera layout, with one row per rig; camera_sampling_weights.npy stores the rig sampling weights. Synthetic free-hand samples have no object table.

For datasets with objects, object_offsets.npy gives each hand sample's range in the object-instance table. object_mesh_indices.npy selects a mesh per instance; object_scales.npy and tfs_object_to_world.npy supply scale and pose. Meshes are packed into object_mesh_vertices.npy and object_mesh_faces.npy, indexed by their corresponding object_mesh_vertex_offsets.npy and object_mesh_face_offsets.npy. Face indices are local to each mesh. World vertices are obtained by scaling local vertices, applying the object rotation, and adding its translation. InterHand retains empty object tables.

The view mask is indexed by split position, while geometry and calibration are indexed by the packed row ID. Preserve both the split indices and their order when loading data.

Citation

@article{cui2026silhand,
  title   = {Hand in Silhouette: Synthetic-to-Real Multi-View Hand Reconstruction via Binary Masks},
  author  = {Cui, Leiyao and Ruan, Lecheng and Zhao, Zihang},
  journal = {arXiv preprint arXiv:2611.16339},
  year    = {2026}
}

The real-world portions derive from ARCTIC, DexYCB, HO3D, InterHand, and OakInk; synthetic grasping data derive from DEXonomy. Please acknowledge the corresponding source datasets when using their processed data.