Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Server error while post-processing the rows. Please report the issue.
Error code:   RowsPostProcessingError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Stereo NIR synthetic hands: GigaHands motion, hand layers

Synthetic stereo near-infrared renders of two hands (MANO meshes, posed with motion from GigaHands), as seen by a headset's pair of downward fisheye tracking cameras, with 3D and 2D joint labels.

There are no backgrounds in this set. Each sample holds only the rendered hand and arm layer, its optical coverage (alpha), the instance mask and the labels. To train on it, composite the layer over a background of your choice: composite = alpha * gain * fg + (1 - alpha) * background.

Some frames also carry a translated hand layer: the same hand re-shaded by a render-to-real appearance model, pasted on top of the composite with its own alpha.

Layout

giga_layers_v1/{train,val}/<chunk>-wNN-NNNNNN.tar, WebDataset-style tars. Per sample KEY, views left and right:

member content
KEY.fg.<view>.png hand and arm layer alone, uint8
KEY.alpha.<view>.png optical coverage, round(alpha × 255)
KEY.mask.<view>.png instance mask (1, 2 hands; 3 arm skin, 4 sleeve, 5 watch, 6 ring; 7, 8 props)
KEY.xhand.<view>.png, KEY.xalpha.<view>.png translated hand and its paste alpha (where present)
KEY.labels.npz joints (world, camera, pixel), visibility, MANO shape and pose, boxes, and so on
KEY.meta.json scalar labels, clip, frame index, split

giga_layers_v1/chunks/cNNN.done marks a fully uploaded chunk.

Licence and credits

CC BY-NC 4.0, non-commercial use only.

  • Hand motion: GigaHands (Fu et al., CVPR 2025), CC BY-NC 4.0. https://github.com/brown-ivl/GigaHands
  • Translated layers: produced by an appearance model trained on Frametop Hands (DeeJanuz/frametop-hands, CC BY-NC 4.0, DeeJanuz and contributors). No Frametop images are included in this set.
  • Hand model: MANO (https://mano.is.tue.mpg.de/). The MANO model files are not included.
Downloads last month
105