Apppublic

vitrus/gen-humanego-semantic-field

sourceHugging Facecc-by-sa-4.0updated 1d agoView on Hugging Face

0likes

App README

VITRUS Gen-HumanEgo Semantic Field

Visual search and synchronized-view inspection for the verified 114-episode Gen-HumanEgo industrial canary.

Run locally

bash

python3 server.py --port 4202

Open http://127.0.0.1:4202/.

Current index

—114 MCAP episodes
—684 decoded previews: six synchronized RGB cameras per episode
—zero extraction failures
—annotation-aware search across factory, station, activity, episode description, task, and active subtask
—source episode path, timestamp, duration, success label, and modality inventory retained in data/index.json

The compact index is a derivative preview. The original MCAP episodes in the Gen-HumanEgo dataset remain the source of truth.

Source dataset: genrobot2025/Gen-HumanEgo, licensed under CC-BY-SA-4.0. This public Space contains a compact set of decoded preview frames derived from the 114-episode industrial sample and links each view back to its source episode metadata.

Rebuild

Run extract_canary.py inside an environment containing mcap, mcap-protobuf-support, and Pillow. FFmpeg is required to decode H.264 camera windows.

The current ranker searches the supplied hierarchy and temporal annotations. It does not yet claim learned image-to-text retrieval. The intended next stage is to replace the ranker with precomputed LVL-1.1 image/episode embeddings while preserving the same review interface and source receipts.