Interedit Eval

interedit-eval

InterEdit: Navigating Text-Guided Multi-Human 3D Motion Editing — Yang et al. (2026) (arXiv:2603.13082, 2026)

What this evaluates

Evaluates a model's ability to edit two-person 3D motions according to text instructions, balancing semantic modification (instruction adherence) with content preservation (source fidelity) and motion realism.

Datasets

• InterEdit3D — total 5161; splits: test (-1); repo https://github.com/YNG916/InterEdit

Metrics

• Recall@1 (primary) — range: percent
  • Percentage of generated motions where the ground truth target (g2t) or source (g2s) motion is ranked in the top 1 by cosine similarity in an L2-normalized learned motion embedding space.
• Recall@2 — range: percent
  • Percentage of generated motions where the ground truth target or source motion is ranked in the top 2 by cosine similarity in the same embedding space.
• Recall@3 — range: percent
  • Percentage of generated motions where the ground truth target or source motion is ranked in the top 3 by cosine similarity in the same embedding space.
• FID — range: other
  • Fréchet Inception Distance measuring the distribution distance between the feature embeddings of generated motions and real target motions in the same learned space. Lower values indicate better realism.

Input / output format

Input: Source 3D multi-human motion sequence and a natural language text instruction describing the desired edit.

Output: Edited 3D multi-human motion sequence.

Scoring recipe

def compute_recall(generated_feats, test_feats, gt_indices, K):
    sims = cosine_similarity(generated_feats, test_feats)
    top_k = np.argsort(sims, axis=1)[:, -K:]
    return np.mean([gt in row for gt, row in zip(gt_indices, top_k)]) * 100

def compute_fid(gen_feats, real_feats):
    mu_g, cov_g = np.mean(gen_feats, axis=0), np.cov(gen_feats, rowvar=False)
    mu_r, cov_r = np.mean(real_feats, axis=0), np.cov(real_feats, rowvar=False)
    diff = mu_g - mu_r
    cov_mean = (cov_g + cov_r) / 2
    sqrt_cov = sqrtm(cov_g @ cov_r)
    return np.sum(diff**2) + np.trace(cov_g + cov_r - 2 * sqrt_cov)

Common pitfalls

• Confusing g2t (instruction adherence) with g2s (source preservation); high g2t with low g2s indicates over-editing or global drift.
• Failing to use the exact InterGen text-to-motion retrieval model for feature extraction, which invalidates the embedding space used for both retrieval and FID.
• Not averaging over 20 independent runs with DDIM (50 steps, η=0) and CFG (scale=3.5), leading to high variance in reported metrics.

Evidence (verbatim from paper)

We use retrieval-based metrics as primary measures following MotionFix. For a 3D multi-human motion editing result, we evaluate: (i) generated-to-target retrieval (g2t) and (ii) generated-to-source retrieval (g2s) in a learned motion embedding space, reporting Recall@K (K∈{1,2,3}). We use the InterGen text-to-motion retrieval model as the feature extractor, where motions are L2-normalized and ranked by cosine similarity against the full test set. g2t measures instruction adherence, while g2s reflects source preservation. We also report FID in the same embedding space to assess motion realism, measuring the distribution distance between generated and real target motions (lower is better). All methods are evaluated for 20 independent runs, and we report the mean with 95% confidence intervals.

Citation

@misc{yang2026interedit,
  title={InterEdit: Navigating Text-Guided Multi-Human 3D Motion Editing},
  author={Yang et al. (2026)},
  year={2026},
  note={arXiv:2603.13082}
}

• arXiv: 2603.13082