Motion-to-Text · HumanML3D

Static leaderboard for motion-to-text captioning on the official HumanML3D test population. All methods use the same 4,400 samples, three-reference TM2T language protocol, candidate groups of 32 for semantic retrieval, and one deterministic evaluation pass.

The all-case explorer exposes the input motion, all human references, and every TM2T, MotionGPT, MotionGPT3, VerMo, and UniMuMo prediction for all 4,400 evaluated clips. The input is shown as an animated neutral SMPL mesh next to all five caption outputs on one page. Large motion chunks are loaded on demand from ZeyuLing/Motius-Leaderboard-Cases. The metric table names both BERTScore scales explicitly: raw cosine similarity and the English layer-17 baseline-rescaled score used by the TM2T protocol.

GT is reported as a reference row under the same protocol. It is excluded from method rankings, best/second-best styling, and comparison charts.

The static Space consists of:

• index.html: responsive leaderboard and protocol layout.
• leaderboard.js: sorting, ranking, chart, and qualitative audit logic.
• m2t_results.json: versioned metric snapshot.
• cases/: all-case SMPL mesh viewer and manifest; binary assets live in the public dataset above.