Pith. sign in

REVIEW 3 major objections 4 minor 43 references

MotionMap: Representing Multimodality in Human Pose Forecasting

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read MotionMap represents all likely human pose futures as peaks on a heatmap, found in a single forward pass.

desk verdict A genuinely new heatmap-based representation for multimodal pose forecasting, undercut by codebook averaging and an evaluation protocol the method is trained on. read the letter →

arxiv 2412.18883 v2 pith:CCLSF67D submitted 2024-12-25 cs.CV eess.IV

Reyhaneh Hosseininejad , Megh Shukla , Saeed Saadatnejad , Mathieu Salzmann , Alexandre Alahi This is my paper · ORCID
classification cs.CVeess.IV
keywords humanposeforecastingmultimodalmotionpredictionheatmaprepresentationsampleefficiencyuncertaintyestimationcodebooktransfer3Ddatasets
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

The paper argues that human pose forecasting can be made well-posed by reframing it as learning motion transitions that already exist in the training data and translating them to unseen test sequences. On top of that reframing, it introduces MotionMap, a heatmap representation in which each local maximum is a distinct possible future motion for a given observation, with a confidence attached. The central claim is that one forward pass over this heatmap, followed by a codebook lookup, replaces the large-scale random sampling of diffusion and VAE baselines while covering more modes, including rare ones. A secondary claim is that the representation naturally supports uncertainty estimation, ranking of forecasts, and user-controllable mode selection.

What carries the argument

The central object is the MotionMap: a 2D heatmap built by t-SNE projection and quantization of autoencoder latents over all future pose sequences in the dataset. Each heatmap cell stores, through a codebook, the mean latent of all futures that fell there; a Gaussian is plotted at each cell for each multimodal ground truth. The heatmap predictor H, a GRU encoder followed by 1x1 convolutions, is trained with weighted binary cross-entropy to reproduce this map from the last three observed frames. At inference, local maxima are found deterministically, and the codebook supplies the missing latent zy that the decoder needs, so no sampling is required.

What would settle it

Construct a test set in which the true future contains a motion transition that never appears in the training data (for instance, a direction change or action combination that was deliberately held out). Running the full pipeline on this test set and observing that the true future falls outside all predicted heatmap peaks, while a diffusion baseline with sufficient samples covers it, would falsify the central claim that training-set transitions suffice for well-posed forecasting.

Watch

Extended reading notes

Core claim

The paper's central discovery is that multimodality in human pose forecasting can be encoded as a spatial distribution over a learned two-dimensional latent space of future motions, rather than by sampling many stochastic predictions. The authors first redefine multimodal ground truth using motion transfer between skeletons and the last three observed frames, so that similar motions from different individuals count as alternatives. They then project latent codes of all futures onto a 2D map, quantize it into a heatmap, and train a GRU-based heatmap predictor per input. At test time, local maxima of the predicted heatmap index a codebook of average latents, each decoded into a full future pose. Because the number of peaks is not fixed, the method outputs a variable number of modes, ranks them by confidence, and recovers rare transitions that averaging would erase. The authors show on Human3.6M and AMASS that MotionMap achieves strong multimodal accuracy with a small fixed number of forecasts.

Load-bearing premise

The method assumes every test-time future is a translation of a motion transition that already appears in the training set, so a genuinely novel transition at test time cannot be predicted or evaluated under its protocol.

Editorial extensions

If this is right

  • The number of forecasts is decided by the input itself through the number of heatmap peaks, so both common and rare modes are represented instead of being suppressed by averaging.
  • Mode confidences turn the forecast set into a ranking, allowing practitioners to pick the most likely future or deliberately select a rare alternative.
  • Because the heatmap encodes transitions seen in training, sample efficiency at a fixed budget (e.g., 5–7 forecasts) improves over diffusion and anchor-based baselines on both datasets.
  • The same representation decomposes uncertainty into mode confidence and per-joint heteroscedastic uncertainty, and enables controllable generation using action labels or nearby latents.

Reading between the lines

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

  • A natural extension is to treat the heatmap itself as a shared interface, letting text descriptions or action labels query modes directly, which could unify language-conditioned and pose-conditioned motion generation.
  • The well-posedness reframing suggests a testable protocol: measure how performance degrades when test actions are drawn from a distribution disjoint from training; if the transition-learning assumption holds, retraining on a superset should help more than additional sampling at test time.
  • Since the codebook stores mean latents per cell, the method could be adapted to output per-cell variance or multiple latents per cell, enabling fine-grained intra-mode variations without losing the heatmap's mode structure.
  • The t-SNE and quantization pipeline is data-dependent; a learned, amortized discrete embedding might preserve rare modes better than the fixed projection, a claim the paper does not test.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes MotionMap, a heatmap-based representation for multimodal human pose forecasting. Future pose sequences in the training set are encoded by a GRU encoder, embedded into two dimensions via t-SNE, and quantized onto a 2D grid; a codebook stores, for each occupied cell, the mean latent of all sequences falling into that cell. A heatmap predictor H is trained to map an observed pose sequence to this distribution, and at inference local maxima are selected and decoded by the autoencoder's decoder, giving a variable number of ranked, deterministic forecasts. The paper also proposes a motion-transfer procedure for building multimodal ground truths from the last three frames with skeleton-size normalization, and introduces an evaluation protocol in which test ground truths are the training-set transitions closest to each test sample. Experiments are reported on Human3.6M and AMASS, including the new protocol (Table 1) and the standard test-split protocol (Tables 2 and 3), together with qualitative results on controllability, uncertainty, and rare-mode coverage.

Significance. The paper addresses a real limitation of sampling-based multimodal predictors: mode coverage usually requires many stochastic samples, while MotionMap obtains modes in one feed-forward pass and additionally supplies per-mode confidence. The representation also enables controllability and heteroscedastic uncertainty, and the authors release code and a project page, which supports reproducibility. If the central claims are substantiated, MotionMap would be a useful alternative to diffusion- or flow-based multimodal forecasting. However, the paper's quantitative case currently rests on a nonstandard evaluation protocol that is partly aligned with the method's own training target, and the codebook-averaging mechanism raises an unresolved concern about whether distinct modes are preserved after quantization. The qualitative results are suggestive but do not, by themselves, establish the sample-efficiency and recall claims at the level needed for a journal publication.

major comments (3)
  1. [Sec. 4.2.1, Sec. 4.3, Sec. 4.4] The codebook construction collapses all latents that quantize to the same heatmap cell into a single mean latent, B = {hy : zy}. Because the decoder is fine-tuned to consume this mean latent and inference decodes each local maximum through the codebook, two semantically distinct futures for the same observation that land in the same cell cannot both be represented; the decoded forecast will be an average that may match neither. This directly affects the claims of variable mode count, transition recall, and diversity. The manuscript provides no intra-cell statistics, such as the number of distinct futures per occupied cell, the dispersion of the zy vectors within a cell, the decoded-pose variance for a cell, or the action-label composition of cells. With a 128x128 grid and hundreds of thousands of sequences in AMASS, such collisions are plausible and should be quantified. Please report these statistics and, if collisions are substantial, consider a codebook with multiple prototypes per cell or a higher effective resolution.
  2. [Sec. 5.5, Table 1, Tables 2 and 3] The headline quantitative evaluation is partly circular. The test multimodal ground truths in Table 1 are obtained by selecting, for each test sample, the closest training-label transitions, using the same motion-transfer and skeleton-scaling procedure that is used to construct the training heatmaps. The model is therefore evaluated on how well it recalls targets that were generated from the training set through the same pipeline that defines its supervision. This favors MotionMap by construction. The standard-split results (Tables 2 and 3) do not consistently support the claimed advantage: on Human3.6M, MotionMap has ADE 0.491 and MMADE 0.505, while BeLFusion has 0.472 and 0.497 and CoMusion has 0.460 and 0.505; on AMASS, MotionMap has ADE 0.624 and MMADE 0.643, while CoMusion has 0.601 and 0.629. No error bars or significance tests are reported. Please provide multi-seed error bars, report both protocols with the same number of samples, and state explicitly which conclusions hold under the strict test-split protocol.
  3. [Sec. 1, Sec. 5.5, Fig. 14] The paper's well-posedness argument relies on the assumption that possible futures for any input are bounded by the transitions present in the training set, and that unseen test futures are translatable training transitions. This premise is load-bearing: the codebook contains only training latents, so a genuinely novel test transition cannot be produced by construction. The paper notes in Fig. 14 that the AMASS train/test splits can be highly imbalanced, but it does not test how often the premise fails or how performance degrades when it does. A concrete test would be to split test sequences by their nearest-neighbor distance to the training distribution, or to evaluate on held-out action categories, and report ADE/MMADE or recall as a function of that distance. Without such an analysis, the 'well-posed' framing is an assumption rather than a validated property.
minor comments (4)
  1. [Sec. 4.3] The sentence 'we use zy instead of zy' is ambiguous; it should read 'we use the averaged codebook latent z̄y in place of the encoder latent zy'.
  2. [Table 1 vs Tables 2 and 3] The Diversity metric is reported with an upward arrow in Table 1 but with a downward arrow in Tables 2 and 3; the desired direction should be stated once and used consistently.
  3. [Sec. 4.2.1] The choice of grid resolution m and the t-SNE perplexity/random seed are not reported; these choices influence cell collisions and the codebook size, so they should be documented for reproducibility.
  4. [Fig. 2] Figure 2 is dense and the role of the codebook during fine-tuning versus inference is hard to follow; a simplified schematic with explicit tensor shapes would improve readability.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity in the transition-recall evaluation; the MotionMap representation itself is not circular, and standard test-split results provide independent grounding.

  1. fitted input called prediction [Section 5.5 (Table 1) vs. Section 4.2.2]
    "We quantitatively assess different baselines on their ability to translate multimodality from observed data to any unseen test sample. To achieve this, we identify multimodal ground truths in the training labels that are closest to each test sample."

    The heatmap predictor H is trained (Sec. 4.2.2) to minimize pixel-wise binary cross-entropy on heatmaps 'constructed using ground-truth multimodal samples,' where those multimodal ground truths are training-set transitions. Table 1 then constructs test ground truths from 'the training labels that are closest to each test sample' and reports the result as recall of transitions. The headline metric's target is therefore the same class of training-label heatmaps that H was fitted to; scoring high on it largely restates the training objective under the paper's well-posed paradigm rather than evaluating an independent held-out future. This partially reduces the rare-mode/transition-recall claim to fitting the training target, although Tables 2/3 use the standard test split and are not circular.

full rationale

No load-bearing self-citation is present: architecture reuse from BeLFusion [1] is external, and the self-citations ([16], [27], [34], [35]) are baselines or recommendations, not premises forcing the result. The codebook averaging of latents per quantized cell (Sec. 4.2.1) is a real fidelity risk but is not circularity; it is an approximation acknowledged by the fine-tuning step. The main circularity concern is the Section 5.5 protocol, which defines the test targets out of the same training labels used to build MotionMap's training heatmaps, making the headline 'transition recall' metric partially a measure of fit to the model's own training target. Because the core representation still generalizes under the standard test-split evaluation in Tables 2 and 3, the overall circularity is partial rather than total.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's central claim rests on several domain assumptions rather than new physical entities. The main free parameters are evaluation thresholds and heatmap resolution, which are chosen and not fitted to an external benchmark. No new particles, forces, or dimensions are introduced.

free parameters (3)
  • multimodal ground truth threshold = 0.5 (Human3.6M), 0.4 (AMASS)
    Distance threshold on the last three frames used to define when two observed sequences yield multimodal ground truths; adopted from prior work but retained as a hyperparameter.
  • heatmap resolution m = 128
    Resolution of the quantized 2D heatmap and codebook; controls granularity of mode separation and codebook size (about 64MB for m=n=128).
  • weighted binary cross-entropy false-negative penalty = not reported
    Weighting that penalizes false negatives more than false positives to encourage extra modes; the exact weight is not stated.
assumptions (4)
  • domain assumption Future motions for unseen test samples are present as transitions in the training set
    Underlies the well-posed paradigm and the Section 5.5 evaluation in which training-set transitions are used as multimodal ground truths for test samples.
  • domain assumption Scaling a skeleton to a reference body size while keeping spherical joint angles preserves the motion identity
    Motion transfer in Section 3 defines which futures count as multimodal ground truths; if it changes motion semantics, the ground truths and evaluation are affected.
  • domain assumption t-SNE embedding of future latents preserves the multimodal structure so that heatmap peaks correspond to distinct modes
    The whole MotionMap representation and codebook decoding in Section 4.2.1 rely on the 2D geometry of the embedding matching latent similarities.
  • domain assumption The BeLFusion backbone (GRU encoders, GRUCell decoder) is a suitable forecasting model
    The autoencoder and decoder architectures are reused from [1]; if this backbone is inadequate, the MotionMap gains would not transfer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MotionMap: Representing Multimodality in Human Pose Forecasting." pith.science (2026). https://pith.science/paper/CCLSF67D

@misc{pith2026241218883,
  author       = {Pith},
  title        = {Pith review of: MotionMap: Representing Multimodality in Human Pose Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CCLSF67D}},
  note         = {Machine review of arXiv:2412.18883}
}
read the original abstract

Human pose forecasting is inherently multimodal since multiple futures exist for an observed pose sequence. However, evaluating multimodality is challenging since the task is ill-posed. Therefore, we first propose an alternative paradigm to make the task well-posed. Next, while state-of-the-art methods predict multimodality, this requires oversampling a large volume of predictions. This raises key questions: (1) Can we capture multimodality by efficiently sampling a smaller number of predictions? (2) Subsequently, which of the predicted futures is more likely for an observed pose sequence? We address these questions with MotionMap, a simple yet effective heatmap based representation for multimodality. We extend heatmaps to represent a spatial distribution over the space of all possible motions, where different local maxima correspond to different forecasts for a given observation. MotionMap can capture a variable number of modes per observation and provide confidence measures for different modes. Further, MotionMap allows us to introduce the notion of uncertainty and controllability over the forecasted pose sequence. Finally, MotionMap captures rare modes that are non-trivial to evaluate yet critical for safety. We support our claims through multiple qualitative and quantitative experiments using popular 3D human pose datasets: Human3.6M and AMASS, highlighting the strengths and limitations of our proposed method. Project Page: https://vita-epfl.github.io/MotionMap

Figures

Figures reproduced from arXiv: 2412.18883 by the authors.

Figure 1
Figure 1. MotionMap uses heatmaps to depict a spatial distribution over the space of motions. Local maxima imply that the correspond￾ing motions have a higher likelihood of being a future motion for an observed pose sequence. MotionMap not only predicts a vari￾able number of modes with the corresponding confidence, but it explicitly encodes rare modes that could otherwise be averaged out. ing methods make a finite number of p… view at source ↗
Figure 2
Figure 2. We define a two stage pipeline for human pose forecasting. At first, we train a framework similar to an autoencoder to predict the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The current approach to finding multimodal ground [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Controllability. MotionMap can also be used with auxiliary data such as action labels for controllable pose forecasting. Since each pose sequence is associated with an embedding and action label, a spatial distribution over the space of motions is the same as that over…
Figure 5
Figure 5. Figure 5: Ranking. Since MotionMap can predict variable number of modes with their associated confidences, our method also allows us to rank predictions. For instance, the highest ranked prediction (top row among Yˆ ) closely matches the ground truth motion. However, rare modes …
Figure 6
Figure 6. Figure 6: Uncertainty. Previous work [16] assumed homoscedasticity, learning generic uncertainty trends independent of the input. In contrast, we show that heteroscedastic modeling can result in semantically richer uncertainty estimates (Sec 5.3). We do this by decomposing uncer…
Figure 7
Figure 7. Figure 7: Diversity. By virtue of using the same decoder as BeLFusion [1], our method learns realistic yet diverse motions. This is because MotionMap can be decoded to select modes that are sufficiently different yet likely. 5.4. Ranking Predictions and Diversity Unlike state-of…
Figure 8
Figure 8. Figure 8: Sampling comparison. We compare the pose forecasts of different methods using the MotionMap representation for three different observations (a, b, c). The multimodal ground truth is represented through MotionMap, and the red points are two-dimensional projections of th…
Figure 9
Figure 9. Figure 9: We visualize the modes (in red crosses) predicted by MotionMap. By hovering over the demo tool, we can view the decoded future [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: We show different strategies for controlled selection of [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Qualitative comparison between MotionMap and the ground truth multimodal heatmap. Our observations indicate that MotionMap [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 12
Figure 12. Figure 12: We overlay predictions for each baseline on the ground truth heatmap, for each of the three input pose sequences. The encoding [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: We show additional forecasts along with the predicted uncertainty per joint and time frame. [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: We plot the density map of ground truth sequences [PITH_FULL_IMAGE:figures/full_fig_p016_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages

  1. [1]

    Belfusion: Latent diffusion for behavior-driven human motion prediction,

    G. Barquero, S. Escalera, and C. Palmero, “Belfusion: Latent diffusion for behavior-driven human motion prediction,” in Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), 2023, pp. 2317–2327. 2, 3, 4, 7, 8,

    work page 2023

  2. [2]

    Comusion: Towards con- sistent stochastic human motion prediction via motion diffusion,

    J. Sun and G. Chowdhary, “Comusion: Towards con- sistent stochastic human motion prediction via motion diffusion,” Proceedings of the European conference on computer vision (ECCV), 2024. 2, 3, 8,

    work page 2024

  3. [3]

    Dlow: Diversifying latent flows for diverse human motion prediction,

    Y . Yuan and K. Kitani, “Dlow: Diversifying latent flows for diverse human motion prediction,” in Pro- ceedings of the European conference on computer vi- sion (ECCV). Springer, 2020, pp. 346–364. 2, 6, 8,

    work page 2020

  4. [4]

    Di- verse human motion prediction via gumbel-softmax sampling from an auxiliary space,

    L. Dang, Y . Nie, C. Long, Q. Zhang, and G. Li, “Di- verse human motion prediction via gumbel-softmax sampling from an auxiliary space,” in Proceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 5162–5171. 2, 3, 8,

    work page 2022

  5. [5]

    Convo- lutional sequence to sequence model for human dy- namics,

    C. Li, Z. Zhang, W. S. Lee, and G. H. Lee, “Convo- lutional sequence to sequence model for human dy- namics,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 5226–5234. 2

    work page 2018

  6. [6]

    Back to mlp: A sim- ple baseline for human motion prediction,

    W. Guo, Y . Du, X. Shen, V . Lepetit, X. Alameda- Pineda, and F. Moreno-Noguer, “Back to mlp: A sim- ple baseline for human motion prediction,” in Proceed- ings of the IEEE/CVF winter conference on applica- tions of computer vision, 2023, pp. 4809–4819. 2

    work page 2023

  7. [7]

    Re- current network models for human dynamics,

    K. Fragkiadaki, S. Levine, P. Felsen, and J. Malik, “Re- current network models for human dynamics,” in Pro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2015, pp. 4346–4354. 2

    work page 2015

  8. [8]

    Structural-rnn: Deep learning on spatio-temporal graphs,

    A. Jain, A. R. Zamir, S. Savarese, and A. Saxena, “Structural-rnn: Deep learning on spatio-temporal graphs,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 5308–5317

    work page 2016

Show all 43 references
  1. [9]

    On human motion prediction using recurrent neural networks,

    J. Martinez, M. J. Black, and J. Romero, “On human motion prediction using recurrent neural networks,” in Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2017, pp. 2891–2900. 2

    2017

  2. [10]

    History repeats itself: Human motion prediction via motion attention,

    W. Mao, M. Liu, and M. Salzmann, “History repeats itself: Human motion prediction via motion attention,” in Proceedings of the European conference on com- puter vision (ECCV). Springer, 2020, pp. 474–489. 2

    2020

  3. [11]

    Learning trajectory dependencies for human motion prediction,

    W. Mao, M. Liu, M. Salzmann, and H. Li, “Learning trajectory dependencies for human motion prediction,” in Proceedings of the IEEE/CVF International Confer- ence on Computer Vision (ICCV), 2019, pp. 9489–9497. 2

    2019

  4. [12]

    Learning dynamic re- lationships for 3d human motion prediction,

    Q. Cui, H. Sun, and F. Yang, “Learning dynamic re- lationships for 3d human motion prediction,” in Pro- ceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2020, pp. 6519– 6527

    2020

  5. [13]

    Motion prediction using trajectory cues,

    Z. Liu, P. Su, S. Wu, X. Shen, H. Chen, Y . Hao, and M. Wang, “Motion prediction using trajectory cues,” in Proceedings of the IEEE/CVF International Confer- ence on Computer Vision (ICCV), 2021, pp. 13 299– 13 308

    2021

  6. [14]

    Space-time-separable graph convolutional network for pose forecasting,

    T. Sofianos, A. Sampieri, L. Franco, and F. Galasso, “Space-time-separable graph convolutional network for pose forecasting,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 11 209–11 218

    2021

  7. [15]

    Progres- sively generating better initial guesses towards next stages for high-quality human motion prediction,

    T. Ma, Y . Nie, C. Long, Q. Zhang, and G. Li, “Progres- sively generating better initial guesses towards next stages for high-quality human motion prediction,” in Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2022, pp. 6437–6446. 2

    2022

  8. [16]

    Toward reliable hu- man pose forecasting with uncertainty,

    S. Saadatnejad, M. Mirmohammadi, M. Daghyani, P. Saremi, Y . Z. Benisi, A. Alimohammadi, Z. Tehrani- nasab, T. Mordan, and A. Alahi, “Toward reliable hu- man pose forecasting with uncertainty,”IEEE Robotics and Automation Letters (RA-L), 2024. 2, 6

    2024

  9. [17]

    A spatio-temporal transformer for 3d human motion pre- diction,

    E. Aksan, M. Kaufmann, P. Cao, and O. Hilliges, “A spatio-temporal transformer for 3d human motion pre- diction,” in International Conference on 3D Vision (3DV). IEEE, 2021, pp. 565–574. 2

    2021

  10. [18]

    Motionbert: A unified perspective on learning hu- man motion representations,

    W. Zhu, X. Ma, Z. Liu, L. Liu, W. Wu, and Y . Wang, “Motionbert: A unified perspective on learning hu- man motion representations,” in Proceedings of the IEEE/CVF International Conference on Computer Vi- sion (ICCV), 2023. 2

    2023

  11. [19]

    Hp-gan: Proba- bilistic 3d human motion prediction via gan,

    E. Barsoum, J. Kender, and Z. Liu, “Hp-gan: Proba- bilistic 3d human motion prediction via gan,” in Pro- ceedings of the IEEE conference on computer vision and pattern recognition workshops, 2018, pp. 1418–

    2018

  12. [20]

    Bihmp-gan: Bidirectional 3d human motion prediction gan,

    J. N. Kundu, M. Gor, and R. V . Babu, “Bihmp-gan: Bidirectional 3d human motion prediction gan,” inPro- ceedings of the AAAI conference on artificial intelli- gence, vol. 33, no. 01, 2019, pp. 8553–8560. 2

    2019

  13. [21]

    The pose knows: Video forecasting by generating pose futures,

    J. Walker, K. Marino, A. Gupta, and M. Hebert, “The pose knows: Video forecasting by generating pose futures,” in Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), 2017, pp. 3332–3341. 2

    2017

  14. [22]

    Mt- vae: Learning motion transformations to generate mul- timodal human dynamics,

    X. Yan, A. Rastogi, R. Villegas, K. Sunkavalli, E. Shechtman, S. Hadap, E. Yumer, and H. Lee, “Mt- vae: Learning motion transformations to generate mul- timodal human dynamics,” in Proceedings of the Euro- pean conference on computer vision (ECCV), 2018, pp. 265–281

    2018

  15. [23]

    A unified 3d human motion synthesis model via conditional varia- tional auto-encoder,

    Y . Cai, Y . Wang, Y . Zhu, T.-J. Cham, J. Cai, J. Yuan, J. Liu, C. Zheng, S. Yan, H. Ding et al., “A unified 3d human motion synthesis model via conditional varia- tional auto-encoder,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, p...

    2021

  16. [24]

    Generating smooth pose sequences for diverse human motion pre- diction,

    W. Mao, M. Liu, and M. Salzmann, “Generating smooth pose sequences for diverse human motion pre- diction,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2021, pp. 13 309–13 318. 2

    2021

  17. [25]

    Motron: Mul- timodal probabilistic human motion forecasting,

    T. Salzmann, M. Pavone, and M. Ryll, “Motron: Mul- timodal probabilistic human motion forecasting,” Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2

    2022

  18. [26]

    Humanmac: Masked motion completion for human motion prediction,

    L.-H. Chen, J. Zhang, Y . Li, Y . Pang, X. Xia, and T. Liu, “Humanmac: Masked motion completion for human motion prediction,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 9544–9555. 2

    2023

  19. [27]

    A generic diffusion-based approach for 3d human pose prediction in the wild,

    S. Saadatnejad, A. Rasekh, M. Mofayezi, Y . Medghalchi, S. Rajabzadeh, T. Mordan, and A. Alahi, “A generic diffusion-based approach for 3d human pose prediction in the wild,” in International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 8246–8253. 2, 3, 8

    2023

  20. [28]

    Diverse human mo- tion prediction guided by multi-level spatial-temporal anchors,

    S. Xu, Y .-X. Wang, and L.-Y . Gui, “Diverse human mo- tion prediction guided by multi-level spatial-temporal anchors,” in European Conference on Computer Vision (ECCV). Springer, 2022, pp. 251–269. 3, 8

    2022

  21. [29]

    Learning semantic latent directions for accurate and controllable human motion prediction,

    G. Xu, J. Tao, W. Li, and L. Duan, “Learning semantic latent directions for accurate and controllable human motion prediction,” arXiv preprint arXiv:2407.11494,

    arXiv

  22. [30]

    Motion diversification net- works,

    H. J. Kim and E. Ohn-Bar, “Motion diversification net- works,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 1650–1660. 3

    2024

  23. [31]

    Faithful heteroscedastic regres- sion with neural networks,

    A. Stirn, H. Wessels, M. Schertzer, L. Pereira, N. San- jana, and D. Knowles, “Faithful heteroscedastic regres- sion with neural networks,” in International Confer- ence on Artificial Intelligence and Statistics. PMLR, 2023, pp. 5593–5613. 4

    2023

  24. [32]

    On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks,

    M. Seitzer, A. Tavakoli, D. Antic, and G. Martius, “On the pitfalls of heteroscedastic uncertainty estimation with probabilistic neural networks,” in International Conference on Learning Representations , 2022. [Online]. Available: https://openreview.net/forum?id= aPOpXlnV1T 4

    2022

  25. [33]

    Effective bayesian heteroscedastic regression with deep neural networks,

    A. Immer, E. Palumbo, A. Marx, and J. E. V ogt, “Effective bayesian heteroscedastic regression with deep neural networks,” in Thirty-seventh Conference on Neural Information Processing Systems , 2023. [Online]. Available: https://openreview.net/forum?id= A6EquH0enk

    2023

  26. [34]

    Tic-tac: A framework for improved covariance estimation in deep heteroscedastic regression,

    M. Shukla, M. Salzmann, and A. Alahi, “Tic-tac: A framework for improved covariance estimation in deep heteroscedastic regression,” in Proceedings of the 41th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research. PMLR, 21–27 Jul 2024

    2024

  27. [35]

    Towards self-supervised covariance estimation in deep heteroscedastic regression,

    M. Shukla, A. Shameem, M. Salzmann, and A. Alahi, “Towards self-supervised covariance estimation in deep heteroscedastic regression,” in The Thirteenth International Conference on Learning Representations,

  28. [36]

    Visualizing data using t-sne,

    L. van der Maaten and G. Hinton, “Visualizing data using t-sne,” Journal of Machine Learning Research,

  29. [37]

    Available: https://opentsne.readthedocs.io/ en/stable/ 5,

    [Online]. Available: https://opentsne.readthedocs.io/ en/stable/ 5,

  30. [38]

    Amass: Archive of motion capture as surface shapes,

    N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. J. Black, “Amass: Archive of motion capture as surface shapes,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 5442–5451. 6,

    2019

  31. [39]

    Human3. 6m: Large scale datasets and predictive meth- ods for 3d human sensing in natural environments,

    C. Ionescu, D. Papava, V . Olaru, and C. Sminchisescu, “Human3. 6m: Large scale datasets and predictive meth- ods for 3d human sensing in natural environments,” IEEE transactions on pattern analysis and machine intelligence, vol. 36, no. 7, pp. 1325–1339, 2013. 6,

    2013

  32. [40]

    The pose knows: Video forecasting by generating pose futures,

    J. Walker, K. Marino, A. Gupta, and M. Hebert, “The pose knows: Video forecasting by generating pose futures,” Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2017. 8,

    2017

  33. [41]

    Generating smooth pose sequences for diverse human motion pre- diction,

    W. Mao, M. Liu, and M. Salzmann, “Generating smooth pose sequences for diverse human motion pre- diction,” Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 8, Appendix We make the code available on https://github. com/vita-epfl/MotionMap. A...

    2021

  34. [2008]

    Available: http://jmlr.org/papers/v9/ vandermaaten08a.html 5

    [Online]. Available: http://jmlr.org/papers/v9/ vandermaaten08a.html 5

  35. [2025]

    Available: https://openreview.net/ forum?id=Q1kPHLUbhi 4

    [Online]. Available: https://openreview.net/ forum?id=Q1kPHLUbhi 4

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.