Pith. sign in

REVIEW 3 major objections 4 minor 48 references

DigitCode shows that the anatomical unit a hand-motion symbol spans—bone, finger, or whole hand—matters more than which quantizer is used, and that finger-level tokens turn a code into an editable handle.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

DigitCode tokenizes hand motion by anatomical units, showing the token span (bone/finger/hand) matters more than the quantizer family, and reduces symbolic reconstruction error by about three quarters.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Solid, well-controlled case that anatomical unit beats quantizer family for hand-motion tokenization; task-unit ordering needs cross-dataset support. the 3 major comments →

arxiv 2608.03127 v1 pith:VP6CCV3D submitted 2026-08-04 cs.RO cs.AI

DigitCode: Symbolic Tokenization of Hand Motion by Anatomical Units

Haoyu Gu , Haotian Lu , Jingrun Du , Xiao-Ping Zhang This is my paper
classification cs.RO cs.AI
keywords hand motion tokenizationsymbolic representationanatomical unitsquantizationrate-distortionfinger editingrobot retargetingLabanotation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper tries to establish that the decisive design choice in discrete hand-motion representation is the anatomical unit a symbol covers, not the quantizer family. Building on Hand Labanotation's T×40 grid of per-bone direction symbols, DigitCode fits, groups, and layers the alphabet along the bone–finger–hand hierarchy, cutting held-out quantization error from 14.71° to 3.26° at essentially the same rate. At a fixed unit, training-free k-means and learned strong quantizers are interchangeable; moving the unit down the anatomy reshapes the rate–distortion frontier. The hierarchy tracks what downstream tasks need: bone-level codes for dynamics, finger-level for interaction, whole-hand for identity. Because a finger is enumerable, one per-finger token serves as a training-free handle for detecting and repairing malformed generated hands and for compiling robot retargeting into lookup tables.

Core claim

The central claim is that a symbolic hand-motion code should be designed by choosing what a token spans, and that the hand's own anatomy supplies the right hierarchy. Finger bones move as coupled chains, so quantizing them independently spends rate on redundancy no quantizer can recover; one per-finger token captures the coupling for free. DigitCode implements this as three changes to HL: a data-adaptive spherical codebook per bone (DigitCode-A), joint per-finger quantization (DigitCode-F), and a coarse per-finger token with a per-bone residual layer (DigitCode-H). Held-out angular error falls from HL-26's 14.71° to 3.26° at comparable bits. The same unit hierarchy predicts which representat

What carries the argument

The central object is the anatomical unit a token spans along the bone–finger–hand hierarchy, implemented on HL's grid of 20 bone-direction vectors per hand. DigitCode-A fits a spherical k-means codebook to the data; DigitCode-F concatenates a finger's four bone directions into a 12-D vector and quantizes them jointly; DigitCode-H layers a per-finger coarse token with a per-bone residual direction captured by rotating each bone's residual toward the canonical pole. The key identity is that a finger's bones form a coupled chain, so joint quantization captures the coupling without extra rate, while a per-bone residual decouples a corrupted bone. The claims are carried by rate–distortion sweeps

Load-bearing premise

The load-bearing premise is that the task-unit order found on InterHand2.6M—bone for dynamics, finger for interaction, hand for identity—reflects hand anatomy rather than that particular dataset, since the cross-dataset evidence covers reconstruction only.

What would settle it

Take a hand-motion dataset with very different statistics (e.g., sign language with unusually independent finger articulations or musical fingering), fit DigitCode variants on it, and run the same downstream battery; if per-finger coding no longer wins interaction tasks or per-bone coding no longer wins forecasting, the claimed hierarchy is dataset-specific. Alternatively, train the learned RVQ-VAE on a much larger hand-motion corpus; if it overtakes DigitCode-H at matched rate, the quantizer-null claim is scoped to current data volume.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single code can carry bone, finger, and whole-hand units at once, giving symbolic hand motion continuous-comparable fidelity—3.26° angular error at HL-26's rate—while keeping tokens readable and editable.
  • Per-finger tokens provide training-free detection, localization, and repair of corrupted fingers, plus compiled O(1) retargeting tables for robot hands, operations continuous parameterizations cannot do locally.
  • At a fixed anatomical unit, training-free k-means and tuned learned VQ are interchangeable within a small margin, so simpler quantizers suffice for hand-motion reconstruction.
  • The unit hierarchy predicts downstream behavior: dynamics need bone-level detail, interaction needs finger-level grouping, and identity needs whole-hand tokens, with a single whole-hand token matching a 20-token per-bone code at a fraction of the rate.
  • Data-adaptive bone codebooks transfer across capture regimes and beat the fixed HL-26 alphabet on reconstruction by 46–61% on unseen datasets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the unit-first principle transfers, tokenizers for other compositional signals (whole body, face, hand–object interaction) should vary token span against quantizer family on the same harness; the paper's HandTok makes that test possible.
  • The 'coarse for identity, fine for fidelity' reversal under noise suggests an adaptive design: switch token granularity by noise regime, using coarse whole-hand tokens for recognition and fine bone residuals for reconstruction, a routing rule the paper hints at but does not fully develop.
  • A testable extension is to use the per-finger codeword residual as a general, training-free anomaly score for hand pose estimates in the wild; the paper demonstrates detection on MediaPipe and generated video but does not claim a comprehensive evaluation.
  • The quantizer-null claim is scoped to current data volume; if a much larger hand-motion corpus closes the learned system's overfitting gap, learned tokenizers could overtake unit-aligned geometric codes at matched rate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. DigitCode proposes a family of symbolic hand-motion tokenizers built on Hand Labanotation's T×40 grid. It varies the anatomical span of a token (bone, finger, hand) and compares this against quantizer-family changes. The method proceeds in three steps: a data-adaptive codebook (DigitCode-A), per-finger joint quantization (DigitCode-F), and a hierarchical finger-code plus per-bone residual (DigitCode-H). On held-out InterHand2.6M, the held-out angular error is reduced from 14.71° (HL-26) to 3.26° at essentially the same bit rate. The paper argues that at a fixed unit, k-means and a tuned learned VQ are interchangeable, while changing the anatomical unit is what shifts the rate–distortion frontier. Downstream experiments on forecasting, generation, retrieval, denoising, editing, and classification are used to claim that dynamics read the bone, interaction the finger, and identity the whole hand. The paper also demonstrates the use of per-finger tokens as editable handles for repairing malformed hands and for robot-hand retargeting, and it releases the HandTok evaluation harness.

Significance. The reconstruction core of the paper is strong and carefully controlled. The grouping control (App. D), the strict unit control for the hierarchical code (App. E, Table 11), the matched-rate quantizer sweep (App. I), and the bootstrap/permutation significance tests (App. B) make the central unit-vs-quantizer finding credible. The release of HandTok as a reproducible testbed is a valuable contribution. If the unit-first principle holds, it is a practically useful design rule for discrete hand representations. The main reservation concerns the general claim that the hierarchy tracks what downstream tasks need: the three-way ordering is supported almost entirely on a single dataset, InterHand2.6M, and one of the six tasks (generation) does not resolve the ordering. The paper's own scoping statements in §7 are more careful than the abstract, and this gap needs to be addressed.

major comments (3)
  1. [Abstract, §4.3, Table 3, App. M/N/P] The abstract and contribution 3 state that 'the hierarchy tracks what downstream tasks need,' and §4.3 presents the ordering bone→dynamics, finger→interaction, hand→identity. The evidence for this ordering, however, is essentially InterHand2.6M-specific. Forecasting (Table 19), retrieval (Table 20), and editing (App. L) are evaluated only on InterHand2.6M; the denoising row has an additional FreiHAND cleanup experiment (Table 23), and classification has an ASL noise-reversal (App. N), but neither tests the full three-way ordering on another dataset. Appendix M shows cross-dataset transfer only for reconstruction, and the robot experiments in App. P measure encoding residual and screening, not the same task-unit ordering. In addition, generation (Table 24) does not separate bone-level from finger-level codes, so the 'six downstream settings' actually include one unresolved task. To suppor
  2. [§7, Table 18, Abstract] The effect decomposition in Table 18 shows that replacing HL-26's fixed cube with a data-fit codebook at the same K=26 and matched rate changes error by 6.26°, while the strict unit control (changing the coarse stage from bone to finger) changes it by 1.63° at a lower-error operating point. On the relative scale the two are comparable (43% vs. 46%). The paper reports both scales, but the abstract's sentence 'The lever is the unit, not the quantizer family' and the corresponding introduction wording can be read as claiming the unit is the dominant design decision overall. In absolute degrees at the compared operating points, the alphabet-content effect is larger. Please restrict the 'lever' claim to the unit-vs-quantizer-family contrast, and state explicitly that codebook content is an equally important (and in absolute terms larger) design lever at the compared point.
  3. [Abstract, §4.2, Table 2] The equivalence claim 'training-free and learned strong quantizers are interchangeable on reconstruction' is the second central claim. The evidence supports this for a single-layer k-means versus a learned VQ with EMA and dead-code revival; RVQ, FSQ, PQ, and BSQ all underperform at matched rate. Since RVQ is also a learned quantizer, the unqualified phrase 'strong quantizers' is misleading. Please define 'strong' at first use (e.g., high-utilization single-layer codebooks) and carry that qualifier into the abstract and conclusion, or state the claim as 'at fixed unit and rate, a single-layer k-means and a single-layer learned VQ are interchangeable.'
minor comments (4)
  1. [§1, Contributions] The phrase 'every token still a readable direction' is inaccurate for DigitCode-F and DigitCode-H, whose tokens are codebook indices over 12-D finger poses or coarse-plus-residual layers. They decode to per-bone directions, but they are not themselves directions. Suggest rewording to 'every token decodes to readable per-bone directions.'
  2. [§7] The term 'training-free k-means' can be confusing: k-means is fit on data, even if it has no learned encoder. Consider using 'geometric k-means' or 'k-means without a learned encoder' throughout.
  3. [Table 3] The column 'DigitCode' does not specify which variant is used in each row. Please label each row with the exact code (e.g., DigitCode-A for forecasting, DigitCode-F for the interaction tasks, DigitCode-H or whole-hand token for classification). Also, the generation row is marked 'unresolved' and should not be counted as evidence for any unit label.
  4. [§4.2, Table 2 caption] The main text explains that 'the null is over strong quantizers, not all of them,' but the Table 2 caption does not. Adding a one-sentence scope note to the caption would prevent readers from misinterpreting the four underperforming columns as contradicting the equivalence claim.

Circularity Check

0 steps flagged

No significant circularity: the core unit-vs-quantizer claims are controlled, held-out, and not reducible to fitted parameters or self-citations.

full rationale

I walked the derivation chain and found no step that reduces to its own inputs by construction or by self-citation. The central reconstruction claims use capture-disjoint held-out splits: adaptive codebooks are fit on the training partition only, and every headline comparison is evaluated on held-out captures with matched rate. The unit-vs-quantizer conclusion is supported by controlled contrasts that hold mechanism, depth, and rate fixed while varying only the anatomical grouping (Table 11) or only the quantizer family (Table 18), so it is not a fitted parameter renamed as a prediction. The downstream task ordering is empirical, though its cross-dataset generality is limited to reconstruction; that is a scoping/generalization caveat, not circularity. Results marked 'exact by construction' (e.g., off-finger editing drift of 0°) are explicitly disclosed as structural, not presented as predictions. The only notable self-citation is HL (Li et al. 2024), which shares a co-author, but it is not load-bearing: HL-26 is redefined in §3 from public 3D joints and recomputed as a baseline in this paper, so the paper does not rely on the citation for its numbers. The mutual-information analysis motivates the finger grouping, but the reported accuracy improvements are measured on held-out data, so the design choice does not force the evaluation result. No circular step meets the evidence bar.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The codebooks are trained on data, and the design choices (grouping, layering) are motivated by empirical statistics on the same dataset family. No new physical entities are postulated.

free parameters (2)
  • Codebook sizes (K, K1, K2) = K=26/64/128; K1=128; K2=8/16/32
    Chosen by matched-rate comparison; codebook entries themselves are fit on training data via spherical k-means.
  • Equivalence margin delta = 0.5 degrees
    Declared margin for quantizer-interchangeability null test, set as one third of the smallest unit effect; chosen after observing effects.
axioms (5)
  • domain assumption Twenty bone-direction vectors per hand fully represent the hand pose for reconstruction and the downstream tasks studied.
    Used throughout §3; ignores absolute position, scale, and global rotation.
  • domain assumption The MI structure computed on InterHand2.6M is representative of hand anatomy across datasets.
    Used to motivate finger grouping (§3) and to explain the unit hierarchy (§4.3); only reconstruction transfer is verified cross-dataset.
  • standard math Spherical k-means minimizes mean angular error for direction quantization.
    Assumed in DigitCode-A/F/H; no formal optimality proof, but standard.
  • domain assumption The residual direction after rotation to the canonical pole is uniform enough for a shared codebook.
    Used in §3.3 for the per-bone residual codebook.
  • domain assumption HL-26 is a valid baseline for rate-distortion comparison.
    The paper acknowledges HL-26 is designed for readability, not reconstruction; used as reference.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DigitCode: Symbolic Tokenization of Hand Motion by Anatomical Units." pith.science (2026). https://pith.science/paper/VP6CCV3D

@misc{pith2026260803127,
  author       = {Pith},
  title        = {Pith review of: DigitCode: Symbolic Tokenization of Hand Motion by Anatomical Units},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VP6CCV3D}},
  note         = {Machine review of arXiv:2608.03127}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Hand motion carries the finest-grained information in human activity, yet the representations behind hand generation, understanding, and robot learning are overwhelmingly continuous--joint angles or MANO parameters. These are accurate but unstructured: a finger cannot be indexed or edited as a symbol, and nothing marks a pose as anatomically valid. Discrete symbolic representations supply exactly this structure, and Hand Labanotation (HL) has shown they are feasible for the hand, writing motion as a T x 40 grid of one fixed direction symbol per bone. Building on this grid, we ask the question underneath it: the anatomical unit a symbol should span--bone, finger, or whole hand. DigitCode answers it by adapting, grouping, and layering HL's alphabet along the hand's unit hierarchy within one code, cutting the symbolic representation's quantization error by three quarters. The lever is the unit, not the quantizer family: at a fixed unit, training-free and learned strong quantizers are interchangeable on reconstruction, while moving down the anatomical hierarchy is what shifts accuracy. The hierarchy also tracks what downstream tasks need. Because a finger is a genuine, enumerable unit, one per-finger token doubles as a training-free, editable handle for jobs a continuous representation cannot address--repairing malformed generated hands, and retargeting them onto robots. We release HandTok, a reproducible testbed, so hand tokenizers can be compared unit-for-unit. Project page: https://digitcode-demo.github.io.

Figures

Figures reproduced from arXiv: 2608.03127 by Haotian Lu, Haoyu Gu, Jingrun Du, Xiao-Ping Zhang.

Figure 1
Figure 1. Figure 1: Overview of DigitCode. Fitting, grouping, and layering HL’s alphabet change the anatomical [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Rate–distortion design space (held-out Inter [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A hand gesture from InterHand2.6M as a 3D joint skeleton at five stages (per-finger colors; the wrist root in black). [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Empirical distribution of hand-bone directions on [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The HL-26 alphabet: 26 directions on a cube surface (6 face centers, 12 edge midpoints, 8 vertices)—{−1, 0, 1} 3 \ {0} normalized to S 2—shown in the hand-canonical frame (X palm normal, Y forward/fingertips, Z ulnar). HL-26 quantizes each bone direction to the nearest of these by argmax cosine. Scheme K entr. b. Ang. (◦ ) HL-26 (cube surface) 26 3.29 14.71 Spherical θϕ uniform 5×8 40 3.64 16.79 Spherical … view at source ↗
Figure 6
Figure 6. Figure 6: The HL-26 token grid as a “hand pianoroll”: [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Change map of the same window: black marks a bone whose token differs from the previous frame. Only [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Per-bone reconstruction error. DigitCode domi [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: DigitCode-F per-finger codebook (index finger): five codes spanning the curl range, each decoded to a whole-finger [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Token editing (single hand). Swapping one finger’s [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 44 canonical work pages · 2 internal anchors

  1. [1]

    S.; and Baek, S

    Cha, J.; Kim, J.; Yoon, J. S.; and Baek, S. 2024. Text2HOI: Text-Guided 3D Motion Generation for Hand-Object Interaction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2024

  2. [2]

    Choensawat, W.; Nakamura, M.; and Hachimura, K. 2015. GenLaban: A Tool for Generating Labanotation from Motion Capture Data. Multimedia Tools and Applications, 74: 10823--10846

    work page 2015

  3. [3]

    Copet, J.; Kreuk, F.; Gat, I.; Remez, T.; Kant, D.; Synnaeve, G.; Adi, Y.; and D \'e fossez, A. 2023. Simple and Controllable Music Generation. In Advances in Neural Information Processing Systems (NeurIPS)

    work page 2023

  4. [4]

    C.; and Zanchettin, C

    de Amorim, C. C.; and Zanchettin, C. 2022. ASL-Skeleton3D and ASL-Phono: Two Novel Datasets for the American Sign Language. arXiv preprint arXiv:2201.02065

    Pith/arXiv arXiv 2022

  5. [5]

    H.; and Peucker, T

    Douglas, D. H.; and Peucker, T. K. 1973. Algorithms for the Reduction of the Number of Points Required to Represent a Digitized Line or Its Caricature. The Canadian Cartographer, 10(2): 112--122

    work page 1973

  6. [6]

    G.; Wang, S.; and Cheng, L

    Guo, C.; Mu, Y.; Javed, M. G.; Wang, S.; and Cheng, L. 2024. MoMask: Generative Masked Modeling of 3D Human Motions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2024

  7. [7]

    Guo, C.; Zuo, X.; Wang, S.; and Cheng, L. 2022. TM2T: Stochastic and Tokenized Modeling for the Reciprocal Generation of 3D Human Motions and Texts. In European Conference on Computer Vision (ECCV)

    work page 2022

  8. [8]

    Hsiao, W.-Y.; Liu, J.-Y.; Yeh, Y.-C.; and Yang, Y.-H. 2021. Compound Word Transformer: Learning to Compose Full-Song Music over Dynamic Directed Hypergraphs. In Proceedings of the AAAI Conference on Artificial Intelligence

    work page 2021

  9. [9]

    Huang, M.; Chen, F.-J.; Lin, Y.-H.; et al. 2025. HOIGPT: Learning Long-Sequence Hand-Object Interaction with Language Models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2025

  10. [10]

    Huang, Y.; Wan, W.; Yang, Y.; Callison-Burch, C.; Yatskar, M.; and Liu, L. 2024. CoMo: Controllable Motion Generation through Language Guided Pose Code Editing. In European Conference on Computer Vision (ECCV)

    work page 2024

  11. [11]

    ScaleMoGen: Autoregressive Next-Scale Prediction for Human Motion Generation

    Hwang, I.; Jang, H.; Zhou, B.; Wang, J.; Kim, Y. M.; and Guo, C. 2026. ScaleMoGen: Autoregressive Next-Scale Prediction for Human Motion Generation. arXiv preprint arXiv:2605.11704

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  12. [12]

    J \'e gou, H.; Douze, M.; and Schmid, C. 2011. Product Quantization for Nearest Neighbor Search. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(1): 117--128

    work page 2011

  13. [13]

    Jiang, B.; Chen, X.; Liu, W.; Yu, J.; Yu, G.; and Chen, T. 2023. MotionGPT: Human Motion as a Foreign Language. In Advances in Neural Information Processing Systems (NeurIPS)

    work page 2023

  14. [14]

    Y.; Chen, J.; Xiang, J.; and Chen, M

    Jiang, J.; Au, H. Y.; Chen, J.; Xiang, J.; and Chen, M. 2024. Motion Part-Level Interpolation and Manipulation over Automatic Symbolic Labanotation Annotation. In International Joint Conference on Neural Networks (IJCNN)

    work page 2024

  15. [15]

    Y.; Xiang, J.; and Chen, J

    Jiang, J.; Au, H. Y.; Xiang, J.; and Chen, J. 2026. LaMoGen: Language to Motion Generation Through LLM-Guided Symbolic Inference. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2026

  16. [16]

    Li, L.; Yang, W.; Xing, J.; Yu, X.; and Zhang, X.-P. 2024. Translating Motion to Notation: Hand Labanotation for Intuitive and Comprehensive Hand Movement Documentation. In Proceedings of the 32nd ACM International Conference on Multimedia (ACM MM)

    work page 2024

  17. [17]

    Li, Y.; Li, Z.; Hou, R.; Chen, Y.; Chang, H.; Liu, H.; and Shan, S. 2026. AnyMo: Scaling Any-Modality Conditional Motion Generation with Masked Modeling. arXiv preprint arXiv:2605.29488

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  18. [18]

    Liu, S.; Zhou, Y.; Yang, J.; Gupta, S.; and Wang, S. 2023. ContactGen: Generative Contact Modeling for Grasp Generation. In IEEE/CVF International Conference on Computer Vision (ICCV)

    work page 2023

  19. [19]

    Lloyd, S. P. 1982. Least Squares Quantization in PCM. IEEE Transactions on Information Theory, 28(2): 129--137

    work page 1982

  20. [20]

    Lu, W.; Xu, Y.; Zhang, J.; Wang, C.; and Tao, D. 2024. HandRefiner: Refining Malformed Hands in Generated Images by Diffusion-Based Conditional Inpainting. In Proceedings of the 32nd ACM International Conference on Multimedia (ACM MM)

    work page 2024

  21. [21]

    Luo, H.; Feng, Y.; Zhang, W.; Zheng, S.; Wang, Y.; Yuan, H.; Liu, J.; Xu, C.; Jin, Q.; and Lu, Z. 2025. Being-H0: Vision-Language-Action Pretraining from Large-Scale Human Videos. arXiv preprint arXiv:2507.15597

    Pith/arXiv arXiv 2025

  22. [22]

    Luo, S.; Yu, B.; and Wang, Z. 2023. DASKEL: An Interactive Choreographic System with Labanotation-Skeleton Translation. In Pacific Graphics Short Papers. The Eurographics Association

    work page 2023

  23. [23]

    Mentzer, F.; Minnen, D.; Agustsson, E.; and Tschannen, M. 2024. Finite Scalar Quantization: VQ-VAE Made Simple. In International Conference on Learning Representations (ICLR)

    work page 2024

  24. [24]

    Moon, G.; Yu, S.-I.; Wen, H.; Shiratori, T.; and Lee, K. M. 2020. InterHand2.6M: A Dataset and Baseline for 3D Interacting Hand Pose Estimation from a Single RGB Image. In European Conference on Computer Vision (ECCV)

    work page 2020

  25. [25]

    Pavlakos, G.; Shan, D.; Radosavovic, I.; Kanazawa, A.; Fouhey, D.; and Malik, J. 2024. Reconstructing Hands in 3D with Transformers. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2024

  26. [26]

    Pinyoanuntapong, E.; Wang, P.; Lee, M.; and Chen, C. 2024. MMM: Generative Masked Motion Model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2024

  27. [27]

    Prillwitz, S.; Leven, R.; Zienert, H.; Hanke, T.; and Henning, J. 1989. HamNoSys Version 2.0: Hamburg Notation System for Sign Languages: An Introductory Guide . Hamburg: Signum Press

    work page 1989

  28. [28]

    Qian, L.; Gu, H.; Zhao, J.; and Wang, Z. 2026. BEAT: Tokenizing and Generating Symbolic Music by Uniform Temporal Steps. In International Conference on Machine Learning (ICML)

    work page 2026

  29. [29]

    Qin, Y.; Wu, Y.-H.; Liu, S.; Jiang, H.; Yang, R.; Fu, Y.; and Wang, X. 2022. DexMV: Imitation Learning for Dexterous Manipulation from Human Videos. In European Conference on Computer Vision (ECCV)

    work page 2022

  30. [30]

    Qin, Y.; Yang, W.; Huang, B.; Van Wyk, K.; Su, H.; Wang, X.; Chao, Y.-W.; and Fox, D. 2023. AnyTeleop: A General Vision-Based Dexterous Robot Arm-Hand Teleoperation System. In Robotics: Science and Systems (RSS)

    work page 2023

  31. [31]

    Romero, J.; Tzionas, D.; and Black, M. J. 2017. Embodied Hands: Modeling and Capturing Hands and Bodies Together. ACM Transactions on Graphics (Proc. SIGGRAPH Asia), 36(6)

    work page 2017

  32. [32]

    C.; and Bowden, R

    Saunders, B.; Camgoz, N. C.; and Bowden, R. 2020. Progressive Transformers for End-to-End Sign Language Production. In European Conference on Computer Vision (ECCV)

    work page 2020

  33. [33]

    Shaw, K.; Bahl, S.; and Pathak, D. 2023. VideoDex: Learning Dexterity from Internet Videos. In Conference on Robot Learning (CoRL)

    work page 2023

  34. [34]

    Shoemake, K. 1985. Animating Rotation with Quaternion Curves. In Proceedings of SIGGRAPH (ACM Computer Graphics), volume 19, 245--254

    work page 1985

  35. [35]

    Symeonidis-Herzig, A.; Low, J.; Mercanoglu Sincan, O.; and Bowden, R. 2026. M3T: Discrete Multi-Modal Motion Tokens for Sign Language Production. arXiv preprint arXiv:2603.23617

    work page arXiv 2026

  36. [36]

    Touvron, H.; Lavril, T.; Izacard, G.; et al. 2023. LLaMA: Open and Efficient Foundation Language Models. arXiv preprint arXiv:2302.13971

    Pith/arXiv arXiv 2023

  37. [37]

    van den Oord, A.; Vinyals, O.; and Kavukcuoglu, K. 2017. Neural Discrete Representation Learning. In Advances in Neural Information Processing Systems (NeurIPS)

    work page 2017

  38. [38]

    Wang, B.; et al. 2025. RealisHuman: A Two-Stage Approach for Refining Malformed Human Parts in Generated Images. In Proceedings of the AAAI Conference on Artificial Intelligence

    work page 2025

  39. [39]

    Wang, T.-K.; Tan, C.-P.; and Yang, Y.-H. 2025. Time-Shifted Token Scheduling for Symbolic Music Generation. arXiv preprint arXiv:2509.23749

    work page arXiv 2025

  40. [40]

    Xie, P.; Zhang, Q.; Peng, T.; Tang, H.; Du, Y.; and Li, Z. 2024. G2P-DDM: Generating Sign Pose Sequence from Gloss Sequence with Discrete Diffusion Model. In Proceedings of the AAAI Conference on Artificial Intelligence

    work page 2024

  41. [41]

    Yuan, W.; Yu, Y.; He, J.; Zhang, Y.; et al. 2024. MoGenTS: Motion Generation based on Spatial-Temporal Joint Modeling. In Advances in Neural Information Processing Systems (NeurIPS)

    work page 2024

  42. [42]

    Zeghidour, N.; Luebs, A.; Omran, A.; Skoglund, J.; and Tagliasacchi, M. 2021. SoundStream: An End-to-End Neural Audio Codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing

    work page 2021

  43. [43]

    Zhang, H.; Li, Y.; Fan, K.; Xu, T.; and Tang, H. 2026. Automatic Generation of Labanotation Based on a Hybrid Transformer--LSTM Network with Multi-Scale Spatio-Temporal Features. Scientific Reports, 16: 18244

    work page 2026

  44. [44]

    Zhang, J.; Zhang, Y.; Cun, X.; Huang, S.; Zhang, Y.; Zhao, H.; Lu, H.; and Shen, X. 2023. T2M-GPT: Generating Human Motion from Textual Descriptions with Discrete Representations. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2023

  45. [45]

    a henb \

    Zhao, Y.; Xiong, Y.; and Kr \"a henb \"u hl, P. 2024. Image and Video Tokenization with Binary Spherical Quantization. arXiv preprint arXiv:2406.07548

    Pith/arXiv arXiv 2024

  46. [46]

    Zimmermann, C.; Argus, M.; and Brox, T. 2021. Contrastive Representation Learning for Hand Shape Estimation. In German Conference on Pattern Recognition (GCPR)

    work page 2021

  47. [47]

    Zimmermann, C.; Ceylan, D.; Yang, J.; Russell, B.; Argus, M.; and Brox, T. 2019. FreiHAND: A Dataset for Markerless Capture of Hand Pose and Shape from Single RGB Images. In IEEE/CVF International Conference on Computer Vision (ICCV)

    work page 2019

  48. [48]

    A.; Ververas, E.; Deng, J.; and Zafeiriou, S

    Zuo, R.; Potamias, R. A.; Ververas, E.; Deng, J.; and Zafeiriou, S. 2025. Signs as Tokens: A Retrieval-Enhanced Multilingual Sign Language Generator. In IEEE/CVF International Conference on Computer Vision (ICCV)

    work page 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.