title: Hand Annotation Worker
emoji: 🤲
colorFrom: blue
colorTo: yellow
sdk: docker
app_port: 7860
license: other
license_name: mixed-upstream-licenses
license_link: README.md#licenses
models:
- nvidia/EgoHand-v1
- nvidia/RT-DETR-Hand-Detector-v1
- nvidia/SOMA-X
Hand annotation worker
Public HF Jobs image for LeRobot Annotation Studio's EgoHand/SOMA hand annotations, with DROID-W + MoGe Small retained. The image includes the frozen model package; Jobs need no private model-repository access. Sign in to launch billed HF Jobs or read private datasets.
The Space build fetches the three large weights anonymously from their public
upstream repositories at model-sources.json's exact revisions and checksums.
model-assets/ contains the remaining frozen support files and licenses, keeping
the Git repository below the Space's 1-GB limit. Every one of the resulting 1,768
model/support files must match model-assets/assets.json before the image builds.
The inference-only EgoHand checkpoint preserves all 817 inference tensors;
optimizer/trainer state is omitted. Checkpoint container metadata and tensor
hashes are pinned in egohand-checkpoint-template.zip, without tensor values.
The builder streams matching tensor records with fixed-size buffers and never
unpickles the source checkpoint. Model weights, precision, native camera/hand
code, tracking and annotation settings are unchanged. Corresponding native
sources, licenses and compiler provenance remain in the source archives,
camera-wheels/ and build-provenance/. The prior native runtime acceptance is
recorded by Studio; this packaging build verifies weights and imports on CPU.
Camera TensorRT plans are compiled on the selected Job GPU.
Licenses
EgoHand and its detector retain NVIDIA OpenMDW-1.1. SOMA/MHR retain Apache-2.0.
The camera and other dependencies retain their included component-specific
licenses and notices. See model-assets/licenses/, model-notices/, the retained
model cache and source archives. This package does not claim a single license
for all components. It contains no MANO, user recordings, predictions, reference
cameras or GPU-specific TensorRT plans.