Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

CoMPAS3D: A Dataset and Benchmark for Interactive Motion

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that interactive motion generation should be judged by whether the motion is legible within a shared move vocabulary and matched to the partner's skill level, and that current models fail both tests.

desk verdict Genuine new dataset, but the legibility/proficiency metrics are unvalidated proxies and the headline failure claim rests on them. read the letter →

arxiv 2507.19684 v2 pith:VS5DOXV3 submitted 2025-07-25 cs.LG cs.AIcs.CLcs.CV

Bermet Burkanova , Yasaman Etesam , Payam Jome Yazdian , Trinity Evans , Chuxuan Zhang , Zoe Stanley , Paige Tuttösí , Angelica Lim This is my paper · ORCID
classification cs.LGcs.AIcs.CLcs.CV
keywords interactivemotiongenerationcapturedatasetsalsapartnerdancemovelegibilityproficiencyestimationvision-languagemodelsfollowerbenchmarksexpertannotation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CoMPAS3D is a motion-capture dataset of improvised partner salsa: 3 hours of dancing by 18 people in nine pairs, spanning beginner, intermediate, and professional levels, with over 2,800 expert-annotated 8-beat segments covering 30 move types, execution errors, and styling. The paper uses it to argue that evaluating interactive motion generation by kinematics alone is insufficient: a follower's motion must also be legible within the shared move vocabulary and appropriate to the partner's proficiency. To measure that, it fine-tunes vision-language models to classify moves and to estimate proficiency, then applies them to generated follower motion from Duolando and InterGen. Those classifiers score generated motion far below ground truth on both move legibility ($F_1$ 6.70–7.69 vs. 53.55 for ground truth) and proficiency appropriateness ($F_1$ 22.47–37.12 vs. 51.64), while kinematic metrics show comparable scores across methods and cannot detect the gap. If the paper is right, future interactive-motion systems should be held to semantic standards, not just physical-realism scores, and datasets need annotations and skill variation to support that.

What carries the argument

The load-bearing mechanism is the pairing of a richly annotated corpus with trained recognition models used as evaluators. The corpus is CoMPAS3D: 3 hours of 120 fps motion capture of improvised salsa duets, segmented into 8-beat cycles aligned to the salsa clave rhythm and labeled by an expert dancer with 30 move categories, styling annotations, and error classes. The benchmark derives three tasks from a dialogue analogy—move classification (transcription), proficiency estimation (fluency assessment), and follower generation (dialogue response)—so that generated motion can be scored semantically. Fine-tuned vision-language models (Qwen2.5-VL, LLaVA-NeXT-Video, InternVL3) act as the recognizers: the move classifier measures whether a generated follower clip is recognizable as one of the shared moves, and the proficiency estimator measures whether its skill level matches the target. These two $F_1$ scores, computed on single-person follower clips to avoid leaking leader information, are the new objective metrics that reveal the gap between generated and ground-truth motion.

What would settle it

Show expert salsa dancers short clips of the generated follower motion that the classifiers scored, ask them to pick one move label per clip, and compute per-clip agreement between humans and the VLM. If humans reliably recognize moves that the VLM labels as failures, or if the VLM's scores do not track human recognition accuracy across clips, then the legibility metric is measuring render shift rather than semantic illegibility.

Watch

Extended reading notes

Core claim

The central claim is that CoMPAS3D is the first openly available improvised partner-dance dataset with expert move transcriptions across three proficiency levels, and that those annotations support two objective evaluation dimensions that kinematics cannot: move legibility and proficiency appropriateness. On ground-truth sequences, fine-tuned vision-language models classify moves and proficiency with enough accuracy to validate the annotations as learnable signal—dyadic proficiency estimation reaches 84.63% accuracy. Applied to follower motion generated by Duolando and InterGen, the same classifiers return move $F_1$ scores of 6.70 and 7.69 and proficiency $F_1$ scores of 22.47 and 37.12, respectively, against ground-truth values of 53.55 and 51.64. The paper reads these low scores as evidence that current generative models produce follower motion that is not legible within the salsa vocabulary and not appropriate to the leader's skill level, a failure that kinematic metrics do not expose. Human raters, using six competition-based judging dimensions, also rate ground truth significantly higher than both generative methods on every dimension ($p < 0.001$).

Load-bearing premise

The assumption that a vision-language classifier trained on ground-truth rendered motion, applied to generated renders, measures what a human would perceive as move legibility; the paper itself does not test whether its objective metrics correlate with human judgment.

Editorial extensions

If this is right

  • Follower-generation models can now be scored on whether their output is recognizable as a move and matched to the partner's skill, not just on physical realism.
  • The move classifier and proficiency estimator are reusable evaluation tools: future models trained or fine-tuned on CoMPAS3D can be compared on legibility and proficiency $F_1$ against the reported baselines.
  • Because dyadic move classification outperforms single-person classification, leader motion carries information about the move; scoring generated followers in single-person view is the conservative way to avoid that leakage.
  • Human ratings on the six competition dimensions confirm the objective-metric gap: ground-truth dancing is rated significantly higher than both generated methods on all dimensions ($p<0.001$), so the benchmark provides a subjective counterpart to the automatic scores.
  • The open area of research the paper identifies is interactive motion generation that is both legible and proficiency-appropriate, with CoMPAS3D as the first public benchmark on which to measure progress.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: if these metrics become training objectives, a natural next step is to optimize generated motion directly against classifier scores (e.g., as a differentiable reward), in the same way ASR-derived scores are used to improve text-to-speech systems.
  • Beyond the paper: the single-person scoring design suggests a testable extension—measuring whether a generated follower is legible to humans only when the leader's cue is visible, since the dyadic classifier's advantage implies moves are partially encoded in the leader.
  • Beyond the paper: the half of the dataset left unannotated is a clean test set for automatic move segmentation, which the paper names as future work; a segmentation-free transcription baseline would complete the speech-recognition analogy.
  • Beyond the paper: the metric-correlation study the paper leaves open could be run at low cost by reusing the human-evaluation setup: ask raters to identify moves in generated clips and compare their accuracy with the VLM's $F_1$ score.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces CoMPAS3D, a motion capture dataset of improvised partner salsa dance consisting of 3 hours of leader-follower motion from 18 dancers across beginner, intermediate, and professional levels, with 2,803 expert-annotated 8-beat segments covering 30 move types, execution errors, and styling. The authors define three benchmark tasks—move classification, proficiency estimation, and follower generation—and propose objective evaluation metrics for generated follower motion: move legibility (measured by move classification F1) and proficiency appropriateness (measured by proficiency estimation F1). They fine-tune three vision-language models (Qwen2.5-VL, LLaVA-NeXT-Video, InternVL3) on the annotations, report classification results, and apply the best classifiers to follower motion generated by Duolando and InterGen. The results show that both generators score far below ground truth on the two proposed metrics, whereas kinematic metrics (FID, diversity, beat alignment) show mixed or comparable performance. A subjective human study along six salsa competition dimensions also rates ground-truth motion significantly higher than both generators.

Significance. CoMPAS3D is a potentially valuable community resource: it is the first openly available improvised partner-dance dataset with expert move transcriptions and proficiency variation, and the authors provide the data, annotation files, benchmark code, and baseline results publicly. The analogy between partner dance and spoken dialogue is productive, and the two proposed objective metrics address a real gap in interactive motion evaluation, which currently relies on kinematic measures. The reported inter-annotator agreement (Cohen's Kappa 0.752 on 5% of move annotations) strengthens the dataset's credibility. If the objective metrics are properly validated, the framework could become a reusable evaluation tool for social motion generation. However, the significance of the paper's headline finding—that current generators produce illegible and proficiency-inappropriate follower motion—depends entirely on the validity of those metrics, which is exactly where the manuscript needs additional work.

major comments (4)
  1. [Section 7 and Table 4] The legibility and appropriateness metrics are not validated against human perception on the identical clips. The paper's own Section 7 concedes that no direct correlation with human judgment is measured, and the human study in Figure 4 asks only six holistic quality dimensions (timing, musicality, technique, difficulty, partner coordination, originality), never asking 'what move is being performed' or 'at what proficiency level.' Because the classifiers in Table 4 are fine-tuned on ground-truth rendered videos and then applied to generated renders, the large drop in F1 could reflect render distribution shift (e.g., different mesh proportions, jitter, camera statistics) rather than semantic illegibility. The ground-truth follower-only move classification F1 of 53.55 (Table 2) is itself low, so the metric is a weak transcriber even in-distribution. To support the headline claim that Duolando and InterGen produce illegible motion, the authors should run a human move-identification/proficiency-rating study on the same generated clips and report agreement with the classifier, or otherwise demonstrate that classifier scores are robust to render-level distribution shift.
  2. [Appendix C] The text states: 'Models were trained for 3 epochs, and we report results from the epoch with the best performance on the test set.' Selecting the best epoch on the test set is a form of test-set model selection that inflates the reported classification accuracy and F1 in Table 2 and the ground-truth reference values in Table 4. This also biases the comparison between methods because each model's test performance is an upper bound rather than an unbiased estimate. The authors should select the epoch on a held-out validation split (for example, the held-out pairs used for proficiency estimation) and report the corresponding test numbers.
  3. [Section C.3] The assignment of ground-truth move labels to generated follower clips by 'maximum temporal overlap with the leader segment on which the follower motion was conditioned' is a strong assumption that creates a risk of circularity. It presumes that the generated follower should perform exactly the move cued by the leader, and that the temporal overlap between generated follower motion and the conditioning leader segment is meaningful. For a generated follower that performs a different yet legible move (or a legible variation), the label would be wrong and the F1 would be penalized, incorrectly scoring the motion as illegible. The authors should justify this labeling rule, provide evidence that generated follower motion is temporally aligned with the conditional leader segment, or report a label-free alternative such as comparing the predicted move distribution against the leader's move distribution.
  4. [Section 6.1 and Table 2] The best single-person move classification F1 is only 53.55 on ground-truth follower-only videos, meaning that even the best classifier mislabels about half of the test segments. The paper should explicitly discuss what this level of performance implies for the legibility metric when applied to generated motion. Specifically, a generated clip scoring F1 of 6.70 could be genuinely illegible, or it could be using a movement pattern that the classifier maps to a different class than the one assigned by the overlap rule. Reporting the confusion matrices in Figure 9 for follower-only settings and discussing the most confusable class pairs would give readers a more concrete interpretation of the legibility numbers.
minor comments (4)
  1. [Appendix C.1, Figure 13 caption] The caption spells the model name as 'Dualando' while the rest of the paper uses 'Duolando'; please correct the typo.
  2. [Section 2, Related Work] The sentence 'The salsa dataset of Senecal et al. include skill-level variation' has a subject-verb agreement error and should read 'includes.'
  3. [Section 5, Objective metrics] There is a typo in 'appropriatenessto partner characteristics'—a missing space between 'appropriateness' and 'to.'
  4. [Figure 2] The move names in the bar chart are very small and difficult to read at printed resolution; a higher-resolution figure or a rotated label layout would improve legibility.

Circularity Check

2 steps flagged · score 3.0 of 10

Mild circularity: the objective metrics are validated by their own classifier accuracy, and the ground-truth reference is epoch-selected on the test set, but the dataset contribution itself remains independent.

  1. self definitional [Section 1 (Introduction), Section 5 (Evaluation Metrics), Section 7 (Limitations)]
    "The objective metrics operate in two stages: we first validate that fine-tuned vision language models (VLMs) accurately classify moves and proficiency on ground-truth sequences (Table 2), establishing them as meaningful measures; we then apply them to generated sequences to assess whether follower motion is legible and proficiency-appropriate."

    Legibility is defined as the F1 of the move classifier and appropriateness as the F1 of the proficiency classifier. The only evidence that these F1 scores measure legibility or appropriateness is the classifiers' own accuracy on ground-truth clips in Table 2. This is the metric validating itself: high classifier F1 shows that the classifier matches its training labels, not that F1 tracks human move identification or proficiency judgment.

  2. fitted input called prediction [Appendix C (Classification Details), used by Tables 2 and 4]
    "Models were trained for 3 epochs, and we report results from the epoch with the best performance on the test set. This selection is motivated by our intention to reuse the fine-tuned model as an evaluation tool on generated data."

    The ground-truth reference values in Table 4 (Move F1 53.55, Proficiency F1 51.64) are computed with a model whose epoch was selected on the very test set used to compute those values, and the same model is then applied to generated clips without any such selection. The reported gap between ground truth and generated motion is therefore partly a test-set fitting artifact: the reference classifier is optimized for the ground-truth test distribution, inflating the baseline against which Duolando and InterGen are judged. The 'strong performance' used to justify the metric's meaningfulness is a fitted maximum over epochs, not an unbiased estimate, and the subsequent comparison to generated output inherits that bias.

full rationale

CoMPAS3D is primarily a dataset and benchmark contribution, and its core assets — 3 hours of improvised salsa motion capture, expert move/error/styling annotations with a reported Cohen's kappa of 0.752, and three proficiency levels — are externally grounded and not circular. The paper does not rely on a load-bearing self-citation chain, a uniqueness theorem, or an ansatz smuggled in via citation. The circularity is confined to the validation and use of the two objective metrics. First, the claim that the metrics are 'meaningful measures' is supported only by the classifiers' own accuracy on ground-truth clips; since legibility and appropriateness are operationally defined as those classifier F1 scores, the validation is self-referential rather than anchored to human move identification or proficiency judgments. The paper itself concedes in Section 7 that no correlation with human judgment was measured. Second, the ground-truth reference values used as the comparison baseline are inflated by selecting the best epoch on the test set, so part of the observed failure of Duolando and InterGen reflects test-set overfitting in the reference model. These issues weaken the specific headline that generated motions are illegible or proficiency-inappropriate, but they do not make the dataset, the annotations, or the kinematic evaluations circular. A score of 3 reflects partial circularity in the objective-metric validation without reducing the central dataset contribution to its own inputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the validity of the salsa move ontology, the transferability of classifier scores from ground-truth renders to generated renders, and the adapted competition criteria. The dataset also relies on hand-chosen proficiency boundaries and a frequency threshold for move classes. No new physical entities, forces, or mediators are introduced.

free parameters (2)
  • Proficiency level boundaries = Beginner: 3-12 months; Intermediate: 1-3 years; Professional: >3 years
    Recruitment thresholds chosen by the authors to define the three skill classes; the paper reports no objective validation that these labels correspond to measurable skill differences.
  • Minimum move frequency threshold = 20 instances
    Appendix C.1 retains only moves appearing at least 20 times in the training set, collapsing the remaining annotated classes into 'Other'. This choice shapes the classification task and the legibility metric.
assumptions (4)
  • domain assumption The 30-class salsa move vocabulary, based on expert knowledge and an external salsa dictionary, is a valid shared movement vocabulary for evaluating follower motion legibility.
    Section 3 and Appendix A.3 define the ontology; if this vocabulary is not the right semantic level for salsa interaction, the move classification benchmark and legibility metric lose meaning.
  • domain assumption Competition judging criteria (timing, musicality, technique, difficulty, partnering/coordination, originality) transfer to improvised social dance evaluation as adapted in Appendix B, where showmanship is omitted and choreography is renamed originality.
    The subjective evaluation relies on these criteria being meaningful for improvised salsa duets, not just choreographed competition performances.
  • domain assumption Rendered SMPL-X videos preserve the motion information required for VLM classification, so a classifier trained on ground-truth renders can be applied to generated renders without distribution shift invalidating the scores.
    The objective metrics in Sections 5 and 6.3 depend on this transfer; the paper does not test whether rendered video artifacts confound classification of generated motion.
  • ad hoc to paper Generated follower clips can be assigned ground-truth move labels by maximum temporal overlap with the conditioned leader segment.
    Appendix C.3 states this assignment rule; it is an unvalidated assumption about alignment between generated follower motion and the leader conditioning segment.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CoMPAS3D: A Dataset and Benchmark for Interactive Motion." pith.science (2026). https://pith.science/paper/VS5DOXV3

@misc{pith2026250719684,
  author       = {Pith},
  title        = {Pith review of: CoMPAS3D: A Dataset and Benchmark for Interactive Motion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VS5DOXV3}},
  note         = {Machine review of arXiv:2507.19684}
}
read the original abstract

Socially interactive humanoid robots must engage with humans through their bodies, adapting in real time to a partner's movement, intent, and abilities. This requires models that understand not just how bodies move, but what movement means in a shared social context. Yet evaluation frameworks for interactive motion generation do not measure whether generated follower motion is legible within a shared movement vocabulary, nor whether it is appropriate to the partner's proficiency level. This gap has two causes: existing frameworks rely on kinematic metrics such as FID and beat alignment that cannot measure either property, and existing datasets lack the move annotations and proficiency variation needed. Salsa is well-suited as an evaluation domain: improvised, dyadic, and governed by a move vocabulary and judging criteria covering timing, musicality, technique, difficulty, partnering, and originality. We present CoMPAS3D, a motion capture dataset of improvised partner salsa paired with an evaluation framework covering kinematic quality, two objective metrics (move legibility and proficiency appropriateness), and six competition-based subjective dimensions. The dataset includes 3 hours of improvisation by 18 dancers spanning beginner, intermediate, and professional levels, with over 2,800 expert-annotated segments covering move types, errors, and stylistic elements. We define three benchmarks: move classification (analogous to transcription), proficiency estimation (fluency assessment), and follower generation (dialogue response). Fine-tuned vision-language models perform strongly on objective metrics applied to ground-truth motion sequences. Applied to Duolando and InterGen, the metrics reveal failures that kinematic metrics miss. Human evaluations confirm the gap between generated and ground-truth motion. CoMPAS3D, annotations, benchmark code, and baseline results are publicly available.

Figures

Figures reproduced from arXiv: 2507.19684 by the authors.

Figure 1
Figure 1. CoMPAS3D comprises 3 hours of improvised salsa dance with beginner (top), intermediate [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Distribution over the 30 move classes (sorted by beginner move frequency) in CoMPAS3D [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Proposed benchmark tasks for the CoMPAS3D dataset: (1) move classification (dyadic and [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Human evaluation study results. Ratings are on a 5-point Likert scale across six salsa [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: ELAN annotation tool used for segmenting and labeling dance moves in animated SMPL-X [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: The annotation validation task of the salsa move identification built in Gorilla. Each trial [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: The human evaluation interface presented to Prolific participants. Participants rated each [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Sample frames from the two video configurations used in our experiments: (1) follower-only [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Confusion matrices for move classification using fine-tuned models, reported in Table 2. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Confusion matrices for move classification using zero-shot models, reported in Table 2. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Confusion matrices for proficiency classification using fine-tuned models, reported in [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Confusion matrices for proficiency classification using zero-shot models, reported in [PITH_FULL_IMAGE:figures/full_fig_p021_12.png]
Figure 13
Figure 13. Figure 13: Confusion matrices for legibility using both zero-shot and fine-tuned InternVL3 models, [PITH_FULL_IMAGE:figures/full_fig_p022_13.png]
Figure 14
Figure 14. Figure 14: Confusion matrices for appropriateness using both zero-shot and fine-tuned Qwen2.5-VL [PITH_FULL_IMAGE:figures/full_fig_p023_14.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. OMG: Omni-Modal Motion Generation for Generalist Humanoid Control

    cs.RO 2026-06 unverdicted novelty 5.0 of 10

    OMG is a diffusion model for omni-modal whole-body humanoid motion generation that uses language, audio, and reference motions after large-scale data curation to achieve state-of-the-art performance and adaptation.

Reference graph

Works this paper leans on

38 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    Generating robot/agent backchannels during a storytelling experiment

    Sames Al Moubayed, Malek Baklouti, Mohamed Chetouani, Thierry Dutoit, Ammar Mahd- haoui, J-C Martin, Stanislav Ondas, Catherine Pelachaud, Jérôme Urbain, and Mehmet Yilmaz. Generating robot/agent backchannels during a storytelling experiment. In2009 IEEE Interna- tional Conference on Robotics and Automation, pages 3749–3754. IEEE, 2009

    work page 2009

  2. [2]

    The nxt-format switchboard corpus: a rich resource for investigating the syntax, semantics, pragmatics and prosody of dialogue.Language resources and evaluation, 44:387–419, 2010

    Sasha Calhoun, Jean Carletta, Jason M Brenier, Neil Mayo, Dan Jurafsky, Mark Steedman, and David Beaver. The nxt-format switchboard corpus: a rich resource for investigating the syntax, semantics, pragmatics and prosody of dialogue.Language resources and evaluation, 44:387–419, 2010

    work page 2010

  3. [3]

    Rules, judging criteria & definitions

    Canada Salsa & Bachata Congress. Rules, judging criteria & definitions. https://www.canadasalsacongress.com/rules, 2026

    work page 2026

  4. [4]

    Participatory sense-making: An enactive approach to social cognition.Phenomenology and the Cognitive Sciences, 6(4):485–507, 2007

    Hanne De Jaegher and Ezequiel A Di Paolo. Participatory sense-making: An enactive approach to social cognition.Phenomenology and the Cognitive Sciences, 6(4):485–507, 2007

    work page 2007

  5. [5]

    Qlora: Efficient finetuning of quantized llms.Advances in neural information processing systems, 36:10088– 10115, 2023

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms.Advances in neural information processing systems, 36:10088– 10115, 2023

    2023

  6. [6]

    Reconstructing three-dimensional models of interacting humans.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Mihai Fieraru, Mihai Zanfir, Elisabeta Oneata, Alin-Ionut Popa, Vlad Olaru, and Cristian Smin- chisescu. Reconstructing three-dimensional models of interacting humans.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    work page 2025

  7. [7]

    Reduced nonverbal interpersonal synchrony in autism spectrum disorder independent of partner diagnosis: a motion energy study.Molecular autism, 11:1–14, 2020

    Alexandra Livia Georgescu, Sevim Koeroglu, A F de C Hamilton, Kai V ogeley, Christine M Falter-Wagner, and Wolfgang Tschacher. Reduced nonverbal interpersonal synchrony in autism spectrum disorder independent of partner diagnosis: a motion energy study.Molecular autism, 11:1–14, 2020

    work page 2020

  8. [8]

    Remos: 3d motion-conditioned reaction synthesis for two-person interactions

    Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Remos: 3d motion-conditioned reaction synthesis for two-person interactions. InEuropean Conference on Computer Vision, pages 418–437. Springer, 2024

    work page 2024

Show all 38 references
  1. [9]

    Multi-person extreme motion prediction

    Wen Guo, Xiaoyu Bie, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Multi-person extreme motion prediction. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13053–13064, 2022

    2022

  2. [10]

    Mdd: A dataset for text-and-music conditioned duet dance generation

    Prerit Gupta, Jason Alexander Fotso-Puepi, Zhengyuan Li, Jay Mehta, and Aniket Bera. Mdd: A dataset for text-and-music conditioned duet dance generation. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13932–13941, October 2025

    2025

  3. [11]

    University of Chicago Press, 1987

    Judith Lynne Hanna.To dance is human: A theory of nonverbal communication. University of Chicago Press, 1987

    1987

  4. [12]

    Lora: Low-rank adaptation of large language models.Proceedings of the 10th International Conference on Learning Representations (ICLR), 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models.Proceedings of the 10th International Conference on Learning Representations (ICLR), 1(2):3, 2022

    2022

  5. [13]

    Interdance: Reactive 3d dance generation with realistic duet interactions

    Ronghui Li, Youliang Zhang, Yachao Zhang, Yuxiang Zhang, Mingyang Su, Jie Guo, Ziwei Liu, Yebin Liu, and Xiu Li. Interdance: Reactive 3d dance generation with realistic duet interactions. arXiv preprint arXiv:2412.16982, 2024

    2024 arXiv

  6. [14]

    Intergen: Diffusion-based multi-human motion generation under complex interactions.International Journal of Computer Vision, 132:3463–3483, 2024

    Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions.International Journal of Computer Vision, 132:3463–3483, 2024

    2024

  7. [15]

    Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding.IEEE transactions on pattern analysis and machine intelligence, 42(10):2684–2701, 2019

    Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding.IEEE transactions on pattern analysis and machine intelligence, 42(10):2684–2701, 2019

    2019

  8. [16]

    Synergy and synchrony in couple dances.arXiv preprint arXiv:2409.04440, 2024

    V ongani Maluleke, Lea Müller, Jathushan Rajasegaran, Georgios Pavlakos, Shiry Ginosar, Angjoo Kanazawa, and Jitendra Malik. Synergy and synchrony in couple dances.arXiv preprint arXiv:2409.04440, 2024. 10

    2024 arXiv

  9. [17]

    ELAN (version 7.1) [computer software]

    Max Planck Institute for Psycholinguistics, The Language Archive. ELAN (version 7.1) [computer software]. https://archive.mpi.nl/tla/elan, 2026. Retrieved from https: //archive.mpi.nl/tla/elan

    2026

  10. [18]

    Single-shot multi-person 3d pose estimation from monocular rgb

    Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In2018 international conference on 3D vision (3DV), pages 120–130. IEEE, 2018

    2018

  11. [19]

    Hamza Mughal, Rishabh Dabral, Kiran Chhatre, Christian Theobalt, Libin Liu, Stefan Kopp, Rachel McDonnell, Michael Neff, Taras Kucherenko, Youngwoo Yoon, and Gustav Eje Henter

    Rajmund Nagy, Hendric V oss, Thanh Hoang-Minh, Mihail Tsakov, Teodor Nikolov, Zeyi Zhang, Tenglong Ao, Sicheng Yang, Shaoli Huang, Yongkang Cheng, M. Hamza Mughal, Rishabh Dabral, Kiran Chhatre, Christian Theobalt, Libin Liu, Stefan Kopp, Rachel McDonnell, Michael Neff, Taras ...

    2026

  12. [20]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318, 2002

    2002

  13. [21]

    Super linguistics: an introduction.Linguistics and Philosophy, 46(4):627–692, 2023

    Pritty Patel-Grosz, Salvador Mascarenhas, Emmanuel Chemla, and Philippe Schlenker. Super linguistics: an introduction.Linguistics and Philosophy, 46(4):627–692, 2023

    2023

  14. [22]

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. InProceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pag...

    2019

  15. [23]

    The mi-motion dataset and benchmark for 3d multi-person motion prediction.arXiv preprint arXiv:2306.13566, 2023

    Xiaogang Peng, Xiao Zhou, Yikai Luo, Hao Wen, Yu Ding, and Zizhao Wu. The mi-motion dataset and benchmark for 3d multi-person motion prediction.arXiv preprint arXiv:2306.13566, 2023

    2023 arXiv

  16. [24]

    Salsa dancing dictionary

    Salsa is Good. Salsa dancing dictionary. https://www.salsaisgood.com/dictionary/Salsa_dictionary.htm, n.d. Accessed: 2025-04-07

    2025

  17. [25]

    Motion analysis and classi- fication of salsa dance using music-related motion features

    Simon Senecal, Niels A Nijdam, and Nadia Magnenat Thalmann. Motion analysis and classi- fication of salsa dance using music-related motion features. InProceedings of the 11th ACM SIGGRAPH Conference on Motion, Interaction and Games, pages 1–10, 2018

    2018

  18. [26]

    Classification of salsa dance level using music and interaction based motion features

    Simon Senecal, Niels Alexander Nijdam, and Nadia Magnenat-Thalmann. Classification of salsa dance level using music and interaction based motion features. InVISIGRAPP (1: GRAPP), pages 100–109, 2019

    2019

  19. [27]

    Theorizing fundamental music/dance interactions in salsa.Music Theory Spectrum, 41(1):74–103, 2019

    Rebecca Simpson-Litke and Chris Stover. Theorizing fundamental music/dance interactions in salsa.Music Theory Spectrum, 41(1):74–103, 2019

    2019

  20. [28]

    Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment

    Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. InInternational Conference on Learning Representations, volume 2024, pages 810–829, 2024

    2024

  21. [29]

    Bailando: 3d dance generation by actor-critic gpt with choreographic memory

    Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11050–11...

    2022

  22. [30]

    Confidence intervals for asr-based tts evaluation

    Jason Taylor and Korin Richmond. Confidence intervals for asr-based tts evaluation. In Interspeech 2021, pages 2791–2795. ISCA, 2021

    2021

  23. [31]

    Spatio-temporal detection of fine-grained dyadic human interactions

    Coert Van Gemeren, Ronald Poppe, and Remco C Veltkamp. Spatio-temporal detection of fine-grained dyadic human interactions. InHuman Behavior Understanding: 7th International Workshop, HBU 2016, Amsterdam, The Netherlands, October 16, 2016, Proceedings 7, pages 116–133. Springe...

    2016

  24. [32]

    Qwen2-vl: Enhancing vision- language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision- language model’s p...

    2024 arXiv

  25. [33]

    Inter-x: Towards versatile human-human interac- tion analysis

    Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter-x: Towards versatile human-human interac- tion analysis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pa...

    2024

  26. [34]

    Regennet: Towards human action-reaction synthesis

    Liang Xu, Yizhou Zhou, Yichao Yan, Xin Jin, Wenhan Zhu, Fengyun Rao, Xiaokang Yang, and Wenjun Zeng. Regennet: Towards human action-reaction synthesis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1759–1769, 2024

    2024

  27. [35]

    Hi4d: 4d instance segmentation of close human interaction

    Yifei Yin, Chen Guo, Manuel Kaufmann, Juan Jose Zarate, Jie Song, and Otmar Hilliges. Hi4d: 4d instance segmentation of close human interaction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17016–17027, 2023

    2023

  28. [36]

    Generating human motion from textual descriptions with discrete representations

    Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. Generating human motion from textual descriptions with discrete representations. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (...

    2023

  29. [37]

    Llava-next: A strong zero-shot video understanding model, April 2024

    Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-next: A strong zero-shot video understanding model, April 2024

    2024

  30. [38]

    Man Styling

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, et al. Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models.arXiv preprint arXiv:2504.10479, 2025. A CoMPAS3D: Additio...

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.