Humanscore Eval

humanscore-eval

HumanScore: Benchmarking Human Motions in Generated Videos — Fang et al. (2026) (arXiv:2604.20157, 2026)

What this evaluates

Evaluates the biomechanical plausibility and realism of human motion in AI-generated videos by measuring anatomical, kinematic, and kinetic correctness. It also assesses how well these automated metrics correlate with human preference judgments.

Datasets

• HumanScore Benchmark — total ?; splits: test (-1)

Metrics

• Anatomy Correctness — range: [0, 100]
  • Measures anatomical plausibility of reconstructed 3D poses, penalizing violations of joint limits and bone-length stability.
• Kinematic Correctness — range: [0, 100]
  • Evaluates temporal stability and motion smoothness, penalizing frame-wise jitter and motion drift in the recovered 3D trajectory.
• Kinetic Correctness (primary) — range: [0, 100]
  • Assesses overall biomechanical consistency including self-collision and physical plausibility, serving as the primary composite score.

Input / output format

Input: Video clips (generated or real) containing a single person performing a specific motion.

Output: Numerical scores (0–100 scale) for each biomechanical dimension and a composite HumanScore.

Scoring recipe

def evaluate_video(video):
    poses, meshes = recover_3d_from_video(video)
    valid_frames = filter_by_confidence(poses, meshes)
    anatomy_score = penalize_joint_limits_and_bone_length(valid_frames)
    kinematic_score = penalize_jitter_and_motion_drift(valid_frames)
    kinetic_score = penalize_self_collision_and_physical_violations(valid_frames)
    humanscore = average(anatomy_score, kinematic_score, kinetic_score)
    return humanscore

Common pitfalls

• Metrics rely on monocular 3D reconstruction, making them sensitive to occlusion, motion blur, and background clutter.
• Real-world videos do not achieve perfect scores due to inherent depth ambiguity and reconstruction noise causing mild frame-wise jitter.
• Atypical or extreme human poses in real videos may be penalized by conservative literature-derived biomechanical bounds.

Evidence (verbatim from paper)

The models are evaluated using our proposed biomechanics-informed metrics across three major dimensions: anatomy correctness, kinematic correctness, and kinetic correctness.

Citation

@misc{fang2026humanscore,
  title={HumanScore: Benchmarking Human Motions in Generated Videos},
  author={Fang et al. (2026)},
  year={2026},
  note={arXiv:2604.20157}
}

• arXiv: 2604.20157