Pith. sign in

REVIEW 3 major objections 4 minor 81 references

VolumetricSMPL: A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions

T0 review · 3 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 VolumetricSMPL represents the human body as a neural signed distance field, claiming 10x faster inference, 6x lower GPU memory, and higher accuracy than the prior COAP occupancy model.

desk verdict Real speedups and a new weight-blending architecture, but the min-composed SDF is not smooth or globally consistent and the paper should say so. read the letter →

arxiv 2506.23236 v1 pith:YBX5TCEC submitted 2025-06-29 cs.CV cs.AI

Marko Mihajlovic , Siwei Zhang , Gen Li , Kaifeng Zhao , Lea Müller , Siyu Tang This is my paper
classification cs.CVcs.AI
keywords volumetricbodymodelsigneddistancefieldneuralblendweightshuman-objectinteractioncollisionavoidancemotionsynthesisparametrichumanself-intersections
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 is trying to establish that a parametric human body model can be represented as a neural signed distance field (SDF) without losing expressiveness, and that this representation is cheap enough for real optimization loops. It claims that by dynamically blending small learned weight matrices through a Neural Blend Weights generator, an SDF decoder with only 64 neurons per layer can match or beat the accuracy of the prior COAP occupancy model while using roughly 10x less inference time and 6x less GPU memory. If correct, collision, contact, and penetration losses can be computed directly and differentiably against a body model, which would matter for reconstructing human-object interactions, recovering human meshes in 3D scenes, synthesizing scene-constrained motion, and resolving self-intersections. The deeper point is that the bottleneck in volumetric body modeling is decoder cost, not representation capacity.

What carries the argument

The central object is the Neural Blend Weights (NBW) generator. For each body part $k$ and layer $l$, the decoder weight matrix is $W^l_k = \bar{W}^l + \sum_{r=1}^R v^l_k[r] \hat{W}^l_k[r]$, where $\bar{W}^l$ is a shared base weight matrix, $\hat{W}^l_k[r]$ are learned shape-weight matrices, and $v^l_k$ are blending coefficients regressed from the part's latent code by a linear layer. The work it does is to let a 7-layer 64-neuron MLP behave as if it had many more parameters, because the weights are reconfigured per pose and shape instead of being fixed. The second load-bearing mechanism is the fusion rule $\tilde{d}(x) = \min_k \tilde{d}_k(x)$ over part-wise SDF predictions, with an analytic box-SDF for points outside all bounding boxes, which is what makes the global field usable for gradient-based collision optimization.

What would settle it

Sample query points in the overlap region of two body-part bounding boxes from the PosePrior test set, compute the model's SDF value, and compare it with the true Euclidean distance to the union surface mesh; systematic overestimation or large discontinuities at part boundaries would falsify the claim that the min-aggregated field is a globally consistent signed distance field.

Watch

Extended reading notes

Core claim

On the paper's own terms, VolumetricSMPL is a feed-forward volumetric body model conditioned on SMPL shape and pose parameters. It partitions the body into kinematic parts, canonicalizes each part by its bone transformation, encodes each part with a shared PointNet, and uses a Neural Blend Weights generator to predict layer weights for a compact 64-neuron SDF decoder per part. Local SDF predictions are aggregated by taking the minimum, and far-field queries are replaced by an analytic bounding-box SDF. The paper reports that this yields higher IoU and lower SDF error than COAP on the PosePrior evaluation set, with about 10x faster inference and 6x lower GPU memory, and that downstream tasks inherit the gains, including a 500x speedup in human-object reconstruction, 3.4x faster egocentric mesh recovery, 7x faster scene-constrained motion synthesis with 20x lower memory, and faster self-intersection resolution.

Load-bearing premise

The load-bearing premise is that taking the smallest distance predicted by the body parts yields a single smooth distance field whose gradients can guide collision optimization, even though at seams between parts that field is not truly smooth and may not measure the true distance to the body surface.

Editorial extensions

If this is right

  • A volumetric body model can replace mesh-triangle collision checks and precomputed scene SDFs, making human-object penetration loss cheap enough to run inside large optimization loops.
  • Because the representation is an SDF, collision guidance in diffusion-based pose and motion samplers can use smooth gradients instead of step-function occupancy signals.
  • The efficiency gains scale to bigger batches: egocentric human mesh recovery runs with a 10x larger batch on the same GPU, and full 80-frame scene-constrained motion sequences fit in 24 GB of memory.
  • Self-intersections can be resolved with volumetric constraints at roughly 3x lower memory and fewer remaining triangle collisions.
  • The same NBW decoder recipe can be attached to SMPL-based models with a single line of code, giving existing pipelines an SDF query, a self-penetration loss, and a collision loss.

Reading between the lines

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

  • If the min-aggregated part SDFs are as accurate near the iso-surface as reported, the same weight-blending trick could be applied to hands or to non-human articulated objects, where decoder memory is often the limiting factor; the paper itself notes that hand articulation is not modeled.
  • The analytic far-field approximation means accuracy claims are tied to bounding-box padding, so a reader should expect larger boxes to reduce SDF fidelity away from the surface, consistent with the paper's own supplementary ablation.
  • The 500x speedup in human-object reconstruction comes from replacing repeated object-mesh SDF computation with a human-as-SDF representation, and the same inversion could be applied to rigid objects in human-scene interaction tasks.
  • The comparison with imGHUM is only approximate, since the released imGHUM model was trained on private data and the paper re-implements its architecture under its own training setup, so claims about that specific model should be read with that caveat.
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

3 major / 4 minor

Summary. The paper introduces VolumetricSMPL, a neural volumetric body model that outputs signed distances for SMPL-family body meshes. The model partitions the body into kinematic parts, encodes each part with a PointNet, and uses a Neural Blend Weights (NBW) generator to produce compact per-part MLP weights. The global SDF is obtained by taking the minimum over part-wise SDF predictions, with an analytic bounding-box SDF used outside all boxes. The authors report roughly 10x faster inference and 6x lower GPU memory than COAP, with comparable or slightly higher IoU, and demonstrate the model in four downstream tasks: image-based human-object reconstruction, egocentric human mesh recovery, scene-constrained motion synthesis, and self-intersection resolution.

Significance. If the efficiency and accuracy claims hold, VolumetricSMPL is a practically useful replacement for volumetric occupancy models like COAP in optimization-heavy applications: the reported 10x/6x resource reductions are concrete, the ablations in Tables 2 and C.1-C.3 are thorough, and the stated MIT-licensed code release would make the add-on easy to adopt. The paper also makes a useful architectural contribution in the NBW weight-generation scheme, which is shown to improve accuracy over a base MLP while keeping the decoder small. However, the central representational claim—that the min-composed field is a smooth, globally consistent signed distance field—is not analyzed for correctness in overlapping joint regions, and the downstream comparisons are not fully controlled. The accuracy gains over COAP are small (mean IoU 94.67 vs. 94.31 in Table 1), so the paper's value rests mainly on the efficiency improvements and on the validity of the SDF for differentiable contact modeling, which is exactly the part that needs closer scrutiny.

major comments (3)
  1. [Sec. 3.2, Eq. (3); Sec. 4.2.4] The global SDF is formed by taking the minimum over part-wise SDF predictions. For overlapping body parts, this min operation does not generally produce the signed distance function of the union volume: inside a point that lies in multiple parts, min_k d_k(x) equals the negative of the maximum of the distances to the individual part boundaries, whereas the true distance to the union boundary can be larger when the closest part-boundary point is interior to another part. The min is also non-differentiable at argmin switches. Training in Eq. (7) supervises only the min-composed output, so the individual part-wise SDFs are not constrained to be valid signed distance functions, yet Sec. 4.2.4 relies on part-wise SDF signs to select intersection points. The paper provides no analysis of the magnitude or effect of this approximation error in overlap regions. Please quantify this error, or change the aggregation and/or supervision so that part-wise predictions are semantically meaningful and gradients are well-behaved.
  2. [Sec. 3.2, Eq. (3)] The hybrid fusion between the analytic bounding-box SDF and the neural implicit SDF is discontinuous at the bounding-box boundaries: outside all boxes the returned value is the distance to the closest box surface, while inside a box the value is the neural SDF prediction. Because the bounding box over-approximates the body part, the analytic value is generally smaller than the true distance to the body, and the switch between the two regimes can create nonsmooth transitions that affect gradient-based optimization. The paper does not report how frequently queries fall in this transition region or whether the discontinuity influences the downstream collision losses. Please provide an analysis of this transition or smooth the fusion.
  3. [Sec. 4.2.2 and Appendix D.2] The EgoHMR comparison is not fully controlled: in Eq. (D.8) the collision-guidance scale is set to a=0.4 for the COAP baseline and a=30 for VolumetricSMPL. Since the downstream claim of improved collision handling depends on this guidance strength, the observed reduction in collision rate from 19.1% to 18.4% could be due to hyperparameter tuning rather than to the SDF representation itself. Please either use identical hyperparameters for both baselines or justify the different values with a sensitivity analysis.
minor comments (4)
  1. [Table 1] The mean IoU improvement over COAP is 0.36 percentage points (94.67 vs. 94.31); the abstract's phrase "enhanced accuracy" should be softened or accompanied by a statistical significance test.
  2. [Sec. 4.1] There is a typo in "striks a good balance" that should read "strikes a good balance".
  3. [Sec. 3.3, Eq. (4)-(5)] The text says each layer "applies affine transformation over the the layer input"; the duplicated "the" should be removed.
  4. [Fig. 3 caption] The caption begins "Neural Blend Weights (NBW) Generatortakes"; a space is missing between "Generator" and "takes".

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: VolumetricSMPL's efficiency and accuracy claims are measured benchmarks, not predictions derived from fitted inputs or self-citations.

full rationale

VolumetricSMPL's central claims (10x faster inference, 6x lower GPU memory, improved IoU/MSE, and downstream speedups) are empirical measurements reported in Tables 1-6 under fixed evaluation protocols, using publicly released baselines and identical hardware. The training loss in Eq. 7 supervises the actual SDF output against mesh-derived ground truth, so there is no fitted parameter that is later renamed as a prediction. The NBW mechanism is described as a repurposing of the authors' ResFields architecture, but that citation supplies an architectural prior, not the result being claimed, and Table 2 directly measures NBW's contribution via ablations. COAP and LEAP are used as baselines and as building blocks, but the comparisons are measured rather than derived from those papers' own numbers. The min-composition of part-wise SDFs in Eq. 3 is a modeling approximation with potential correctness and robustness risks, not a circular step, because it does not assume the downstream claims. No equation in the paper reduces a claimed prediction to its own input by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces an architectural mechanism (NBW) rather than a new physical entity. The main assumptions center on the correctness of the min-composed SDF and the adequacy of the analytic far-field approximation, along with the representativeness of the training data.

free parameters (5)
  • NBW rank R = 80
    Number of learned shape weight matrices blended to produce decoder weights; chosen from ablation in Table 2 as a saturation point.
  • Bounding box padding = 12.5%
    Padding of each body-part bounding box, selected in Appendix C.2.
  • Point cloud samples per part = 1000
    Resampled points per body part for the PointNet encoder, chosen in Appendix C.1.
  • MLP decoder width = 64
    Neurons per layer of the SDF decoder, selected in Appendix C.3 as an accuracy/efficiency trade-off.
  • Positional encoding levels = 2
    Only two frequency levels for local query points, chosen in Section 4.1 ablation.
assumptions (4)
  • standard math Rigid bone transformations preserve signed distances, so canonical-space SDF can be evaluated after inverse transformation (Eq. 2).
    Used to transform query points into part-local canonical space.
  • domain assumption The minimum of part-wise neural SDF predictions forms a valid global SDF.
    Eq. 3 and Section 3.2 use min aggregation without justifying smoothness or distance correctness in overlap regions.
  • domain assumption Analytic SDF outside all bounding boxes is a sufficient approximation for far points.
    Section 3.2 adopts bounding-box distance to avoid neural queries far from the body; accuracy depends on box padding and shape.
  • domain assumption Training on MoVi and DFaust subsets generalizes to challenging poses such as PosePrior.
    Section 3.4 uses these AMASS subsets to train the model; no domain adaptation is applied.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VolumetricSMPL: A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions." pith.science (2026). https://pith.science/paper/YBX5TCEC

@misc{pith2026250623236,
  author       = {Pith},
  title        = {Pith review of: VolumetricSMPL: A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YBX5TCEC}},
  note         = {Machine review of arXiv:2506.23236}
}
read the original abstract

Parametric human body models play a crucial role in computer graphics and vision, enabling applications ranging from human motion analysis to understanding human-environment interactions. Traditionally, these models use surface meshes, which pose challenges in efficiently handling interactions with other geometric entities, such as objects and scenes, typically represented as meshes or point clouds. To address this limitation, recent research has explored volumetric neural implicit body models. However, existing works are either insufficiently robust for complex human articulations or impose high computational and memory costs, limiting their widespread use. To this end, we introduce VolumetricSMPL, a neural volumetric body model that leverages Neural Blend Weights (NBW) to generate compact, yet efficient MLP decoders. Unlike prior approaches that rely on large MLPs, NBW dynamically blends a small set of learned weight matrices using predicted shape- and pose-dependent coefficients, significantly improving computational efficiency while preserving expressiveness. VolumetricSMPL outperforms prior volumetric occupancy model COAP with 10x faster inference, 6x lower GPU memory usage, enhanced accuracy, and a Signed Distance Function (SDF) for efficient and differentiable contact modeling. We demonstrate VolumetricSMPL's strengths across four challenging tasks: (1) reconstructing human-object interactions from in-the-wild images, (2) recovering human meshes in 3D scenes from egocentric views, (3) scene-constrained motion synthesis, and (4) resolving self-intersections. Our results highlight its broad applicability and significant performance and efficiency gains.

Figures

Figures reproduced from arXiv: 2506.23236 by the authors.

Figure 1
Figure 1. We introduce VolumetricSMPL, a volumetric body model that represents human shape as a Signed Distance Field (SDF), en [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview. VolumetricSMPL models the human body volume as a Signed Distance Field (SDF) based on a parametric mesh body model [38, 49], controlled via shape and pose parameters (β, θ). First, the human mesh is sampled into a point cloud, where each point is assigned to one of K body parts (colored regions). Each part is then canonicalized based on its respective bone transformation Gk and over-approximated by a bound… view at source ↗
Figure 3
Figure 3. Neural Blend Weights (NBW) Generator takes a local latent code zk as input and outputs layer-specific weights Wk for the kth SDF decoder. First, zk is transformed into a set of blend￾ing coefficients {v l k} L l=1 using L independent learnable linear lay￾ers. These coefficients modulate the shape weights Wl k, which are then added to the base network weights Wl . The resulting neural blend weights Wk = {Wl k} L l=1 … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

81 extracted references · 74 canonical work pages

  1. [1]

    Ijaz Akhter and Michael J. Black. Pose-conditioned joint angle limits for 3D human pose reconstruction. In CVPR,

  2. [2]

    imGHUM: Implicit generative models of 3d human shape and articulated pose

    Thiemo Alldieck, Hongyi Xu, and Cristian Sminchisescu. imGHUM: Implicit generative models of 3d human shape and articulated pose. In ICCV, 2021. 2, 3, 5, 6, 1

    work page 2021

  3. [3]

    Ultra inertial poser: Scalable motion capture and track- ing from sparse inertial sensors and ultra-wideband ranging

    Rayan Armani, Changlin Qian, Jiaxi Jiang, and Christian Holz. Ultra inertial poser: Scalable motion capture and track- ing from sparse inertial sensors and ultra-wideband ranging. In ACM SIGGRAPH, 2024. 2, 3

    work page 2024

  4. [4]

    Behave: Dataset and method for tracking human object in- teractions

    Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object in- teractions. In CVPR, 2022. 3

    work page 2022

  5. [5]

    Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In ECCV, 2016. 3

    2016

  6. [6]

    Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bod- ies in motion. In CVPR, 2017. 5, 6

    work page 2017

  7. [7]

    pi-gan: Periodic implicit genera- tive adversarial networks for 3d-aware image synthesis

    Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit genera- tive adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 3

    work page 2021

  8. [8]

    Parametric modeling of 3d human body shape—a survey

    Zhi-Quan Cheng, Yin Chen, Ralph R Martin, Tong Wu, and Zhan Song. Parametric modeling of 3d human body shape—a survey. Computers & Graphics, 2018. 3

    work page 2018

Show all 81 references
  1. [9]

    Openmmlab pose estimation tool- box and benchmark

    MMPose Contributors. Openmmlab pose estimation tool- box and benchmark. https://github.com/open- mmlab/mmpose, 2020. 3

    2020

  2. [10]

    In- terfusion: Text-driven generation of 3d human-object inter- action

    Sisi Dai, Wenhao Li, Haowen Sun, Haibin Huang, Chongyang Ma, Hui Huang, Kai Xu, and Ruizhen Hu. In- terfusion: Text-driven generation of 3d human-object inter- action. In ECCV, 2025. 2, 3

    2025

  3. [11]

    NASA: Neural Articulated Shape Approxima- tion

    Boyang Deng, JP Lewis, Timothy Jeruzalski, Gerard Pons- Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. NASA: Neural Articulated Shape Approxima- tion. In ECCV, 2020. 3, 5, 1

    2020

  4. [12]

    Remips: Phys- ically consistent 3d reconstruction of multiple interacting people under weak supervision

    Mihai Fieraru, Mihai Zanfir, Teodor Szente, Eduard Baza- van, Vlad Olaru, and Cristian Sminchisescu. Remips: Phys- ically consistent 3d reconstruction of multiple interacting people under weak supervision. NeurIPS, 2021. 2, 3

    2021

  5. [13]

    MoVi: A large multipurpose motion and video dataset

    Saeed Ghorbani, Kimia Mahdaviani, Anne Thaler, Konrad Kording, Douglas James Cook, Gunnar Blohm, and Niko- laus F Troje. MoVi: A large multipurpose motion and video dataset. arXiv preprint arXiv:2003.01888, 2020. 5, 6

    2003 arXiv

  6. [14]

    Fast R-CNN

    Ross Girshick. Fast R-CNN. In ICCV, 2015. 6

    2015

  7. [15]

    Smpl-a: Modeling person-specific deformable anatomy

    Hengtao Guo, Benjamin Planche, Meng Zheng, Srikrishna Karanam, Terrence Chen, and Ziyan Wu. Smpl-a: Modeling person-specific deformable anatomy. In CVPR, 2022. 2

    2022

  8. [16]

    Dai, and Quoc V

    David Ha, Andrew M. Dai, and Quoc V . Le. Hypernetworks. In ICLR, 2017. 3

    2017

  9. [17]

    Tetgen, a delaunay-based quality tetrahedral mesh generator

    Si Hang. Tetgen, a delaunay-based quality tetrahedral mesh generator. TOMS, 2015. 3

    2015

  10. [18]

    Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambigu- ities with 3D scene constraints. In ICCV, 2019. 2, 3, 8

    2019

  11. [19]

    Populating 3d scenes by learning human-scene interaction

    Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3d scenes by learning human-scene interaction. In CVPR, 2021. 2

    2021

  12. [20]

    Learning and tracking the 3d body shape of freely moving infants from rgb-d sequences

    Nikolas Hesse, Sergi Pujades, Michael J Black, Michael Arens, Ulrich G Hofmann, and A Sebastian Schroeder. Learning and tracking the 3d body shape of freely moving infants from rgb-d sequences. PAMI, 2019. 2

    2019

  13. [21]

    Denoising diffu- sion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models. NeurIPS, 2020. 5

    2020

  14. [22]

    Robust inside-outside segmentation using generalized wind- ing numbers

    Alec Jacobson, Ladislav Kavan, and Olga Sorkine-Hornung. Robust inside-outside segmentation using generalized wind- ing numbers. TOG, 2013. 2, 3

    2013

  15. [23]

    Coherent reconstruction of multiple humans from a single image

    Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In CVPR, 2020. 3

    2020

  16. [24]

    Op- timizing diffusion noise can serve as universal motion priors

    Korrawe Karunratanakul, Konpat Preechakul, Emre Aksan, Thabo Beeler, Supasorn Suwajanakorn, and Siyu Tang. Op- timizing diffusion noise can serve as universal motion priors. In CVPR, 2024. 3

    2024

  17. [25]

    Neu- ral 3d mesh renderer

    Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neu- ral 3d mesh renderer. In CVPR, 2018. 5

    2018

  18. [26]

    Oc- cluded human mesh recovery

    Rawal Khirodkar, Shashank Tripathi, and Kris Kitani. Oc- cluded human mesh recovery. In CVPR, 2022. 3

    2022

  19. [27]

    Beyond the contact: Discovering comprehensive affor- dance for 3d objects from pre-trained 2d diffusion models

    Hyeonwoo Kim, Sookwan Han, Patrick Kwon, and Hanbyul Joo. Beyond the contact: Discovering comprehensive affor- dance for 3d objects from pre-trained 2d diffusion models. In ECCV, 2024. 2, 3

    2024

  20. [28]

    Data-driven physics for human soft tissue animation

    Meekyoung Kim, Gerard Pons-Moll, Sergi Pujades, Seung- bae Bang, Jinwook Kim, Michael J Black, and Sung-Hee Lee. Data-driven physics for human soft tissue animation. TOG, 2017. 3

    2017

  21. [29]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6

    2015

  22. [30]

    Pointrend: Image segmentation as rendering

    Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Gir- shick. Pointrend: Image segmentation as rendering. In CVPR, 2020. 5

    2020

  23. [31]

    Huang, Otmar Hilliges, and Michael J

    Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In ICCV, 2021. 3

    2021

  24. [32]

    Dreamhuman: Animatable 3d avatars from text

    Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Ed- uard Gabriel Bazavan, Mihai Fieraru, and Cristian Sminchis- escu. Dreamhuman: Animatable 3d avatars from text. In NeurIPS, 2023. 2

    2023

  25. [33]

    Egogen: An egocentric synthetic data generator

    Gen Li, Kaifeng Zhao, Siwei Zhang, Xiaozhong Lyu, Mi- hai Dusmanu, Yan Zhang, Marc Pollefeys, and Siyu Tang. Egogen: An egocentric synthetic data generator. In CVPR, pages 14497–14509, June 2024. 2, 3

    2024

  26. [34]

    Lodge++: High-quality and long dance gen- eration with vivid choreography patterns

    Ronghui Li, Hongwen Zhang, Yachao Zhang, Yuxiang Zhang, Youliang Zhang, Jie Guo, Yan Zhang, Xiu Li, and Yebin Liu. Lodge++: High-quality and long dance gen- eration with vivid choreography patterns. arXiv preprint arXiv:2410.20389, 2024. 3

    arXiv

  27. [35]

    Interdance: Reactive 3d dance generation with re- alistic 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 re- alistic duet interactions. arXiv preprint arXiv:2412.16982 ,

    arXiv

  28. [36]

    Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and ex- pression from 4D scans. SIGGRAPH Asia, 2017. 3

    2017

  29. [37]

    Nimble: a non-rigid hand model with bones and muscles

    Yuwei Li, Longwen Zhang, Zesong Qiu, Yingwenqi Jiang, Nianyi Li, Yuexin Ma, Yuyao Zhang, Lan Xu, and Jingyi Yu. Nimble: a non-rigid hand model with bones and muscles. TOG, 2022. 3

    2022

  30. [38]

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi- person linear model. TOG, 2015. 2, 3, 4, 5, 1

    2015

  31. [39]

    Dynamics-regulated kinematic policy for egocentric pose es- timation

    Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose es- timation. NeurIPS, 2021. 3

    2021

  32. [40]

    Troje, Ger- ard Pons-Moll, and Michael J

    Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Ger- ard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In ICCV, 2019. 5

    2019

  33. [41]

    ResFields: Residual neural fields for spatiotem- poral signals

    Marko Mihajlovic, Sergey Prokudin, Marc Pollefeys, and Siyu Tang. ResFields: Residual neural fields for spatiotem- poral signals. In ICLR, 2024. 3, 5

    2024

  34. [42]

    Coap: Compositional articulated occupancy of people

    Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. Coap: Compositional articulated occupancy of people. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 8

    2022

  35. [43]

    LEAP: Learning articulated occupancy of people

    Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. LEAP: Learning articulated occupancy of people. In CVPR, 2021. 2, 3, 5, 6, 1

    2021

  36. [44]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. In ECCV, 2020. 4, 5, 6, 1

    2020

  37. [45]

    Generative proxemics: A prior for 3d social interaction from images

    Lea M ¨uller, Vickie Ye, Georgios Pavlakos, Michael Black, and Angjoo Kanazawa. Generative proxemics: A prior for 3d social interaction from images. In CVPR, 2024. 2, 3

    2024

  38. [46]

    Star: Sparse trained articulated human body regressor

    Ahmed AA Osman, Timo Bolkart, and Michael J Black. Star: Sparse trained articulated human body regressor. In ECCV, 2020. 2

    2020

  39. [47]

    Neural parametric models for 3d de- formable shapes

    Pablo Palafox, Aljaz Bozic, Justus Thies, Matthias Nießner, and Angela Dai. Neural parametric models for 3d de- formable shapes. In ICCV, 2021. 3

    2021

  40. [48]

    Spams: Structured implicit parametric models

    Pablo Palafox, Nikolaos Sarafianos, Tony Tung, and Angela Dai. Spams: Structured implicit parametric models. In CVPR, 2022. 3

    2022

  41. [49]

    Expressive body capture: 3d hands, face, and body from a single image

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 2, 3, 4, 8, 1

    2019

  42. [50]

    Building statistical shape spaces for 3d human modeling

    Leonid Pishchulin, Stefanie Wuhrer, Thomas Helten, Chris- tian Theobalt, and Bernt Schiele. Building statistical shape spaces for 3d human modeling. Pattern Recognition, 2017. 3

    2017

  43. [51]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2, 4, 1

    2017

  44. [52]

    Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In ICCV, 2021. 2

    2021

  45. [53]

    Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bod- ies together. SIGGRAPH Asia, 2017. 3

    2017

  46. [54]

    Implicit neural representa- tions with periodic activation functions

    Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representa- tions with periodic activation functions. NeurIPS, 2020. 3

    2020

  47. [55]

    Human pose estimation from silhouettes

    Cristian Sminchisescu and Alexandru C Telea. Human pose estimation from silhouettes. a consistent approach using dis- tance level sets. In WSCG, 2002. 2

    2002

  48. [56]

    Recovering 3d human mesh from monocular images: A sur- vey

    Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A sur- vey. PAMI, 2023. 2, 3, 1

    2023

  49. [57]

    Bodynet: V olu- metric inference of 3d human body shapes

    Gul Varol, Duygu Ceylan, Bryan Russell, Jimei Yang, Ersin Yumer, Ivan Laptev, and Cordelia Schmid. Bodynet: V olu- metric inference of 3d human body shapes. In ECCV, 2018. 3

    2018

  50. [58]

    Metaavatar: Learning animatable clothed human models from few depth images

    Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Metaavatar: Learning animatable clothed human models from few depth images. In NeurIPS,

  51. [59]

    Reconstructing action- conditioned human-object interactions using commonsense knowledge priors

    Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action- conditioned human-object interactions using commonsense knowledge priors. In 3DV, 2022. 1, 2, 3, 6, 7, 4, 5

    2022

  52. [60]

    Envposer: Environment-aware realistic human motion estimation from sparse observations with uncertainty modeling

    Songpengcheng Xia, Yu Zhang, Zhuo Su, Xiaozheng Zheng, Zheng Lv, Guidong Wang, Yongjie Zhang, Qi Wu, Lei Chu, and Ling Pei. Envposer: Environment-aware realistic human motion estimation from sparse observations with uncertainty modeling. arXiv preprint arXiv:2412.10235, 2024. 2, 3

    2024 arXiv

  53. [61]

    Neural fields in visual computing and beyond

    Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tomp- kin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. InComputer Graphics Forum,

  54. [62]

    H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion

    Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. NeurIPS, 2021. 2

    2021

  55. [63]

    Ghum & ghuml: Generative 3d human shape and articulated pose models

    Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Smin- chisescu. Ghum & ghuml: Generative 3d human shape and articulated pose models. In CVPR, 2020. 2, 3, 4

    2020

  56. [64]

    Gavatar: Animatable 3d gaussian avatars with implicit mesh learning

    Ye Yuan, Xueting Li, Yangyi Huang, Shalini De Mello, Koki Nagano, Jan Kautz, and Umar Iqbal. Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. CVPR, 2024. 2

    2024

  57. [65]

    Physdiff: Physics-guided human motion diffusion model

    Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In CVPR, 2023. 2, 3

    2023

  58. [66]

    Simpoe: Simulated character control for 3d human pose estimation

    Ye Yuan, Shih-En Wei, Tomas Simon, Kris Kitani, and Jason Saragih. Simpoe: Simulated character control for 3d human pose estimation. In CVPR, 2021. 3

    2021

  59. [67]

    Twenty years of mixture of experts

    Seniha Esen Yuksel, Joseph N Wilson, and Paul D Gader. Twenty years of mixture of experts. IEEE TNNLS, 2012. 2

    2012

  60. [68]

    Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa

    Jason Y . Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In ECCV, 2020. 1, 2, 3, 6, 4, 5

    2020

  61. [69]

    Probabilistic human mesh recovery in 3d scenes from egocentric views

    Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. InICCV, 2023. 1, 2, 3, 7, 5, 6

    2023

  62. [70]

    Ego- body: Human body shape and motion of interacting people from head-mounted devices

    Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego- body: Human body shape and motion of interacting people from head-mounted devices. In ECCV, 2022. 3

    2022

  63. [71]

    Ego- body: Human body shape and motion of interacting people from head-mounted devices

    Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego- body: Human body shape and motion of interacting people from head-mounted devices. In ECCV, 2022. 7, 6

    2022

  64. [72]

    Learning motion priors for 4d human body capture in 3d scenes

    Siwei Zhang, Yan Zhang, Federica Bogo, Pollefeys Marc, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In ICCV, 2021. 2

    2021

  65. [73]

    Black, and Siyu Tang

    Siwei Zhang, Yan Zhang, Qianli Ma, Michael J. Black, and Siyu Tang. PLACE: Proximity learning of articulation and contact in 3D environments. In 3DV, 2020. 2

    2020

  66. [74]

    Couch: Towards controllable human-chair interactions

    Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In ECCV, 2022. 3

    2022

  67. [75]

    Black, and Siyu Tang

    Yan Zhang, Michael J. Black, and Siyu Tang. We are more than our joints: Predicting how 3D bodies move. In CVPR,

  68. [76]

    Generating 3d people in scenes with- out people

    Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3d people in scenes with- out people. In CVPR, 2020. 2, 3

    2020

  69. [77]

    The wanderings of odysseus in 3d scenes

    Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In CVPR, 2022. 3

    2022

  70. [78]

    A diffusion-based au- toregressive motion model for real-time text-driven motion control

    Kaifeng Zhao, Gen Li, and Siyu Tang. A diffusion-based au- toregressive motion model for real-time text-driven motion control. In ICLR, 2025. 1, 2, 7, 8, 6

    2025

  71. [79]

    Compositional human-scene interaction synthe- sis with semantic control

    Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthe- sis with semantic control. In ECCV, 2022. 3

    2022

  72. [80]

    Synthesizing diverse human motions in 3d indoor scenes

    Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In ICCV, 2023. 3 VolumetricSMPL: VolumetricSMPL: A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions – Supplementary Mate...

    2023

  73. [81]

    Instead, they learn volumetric representations directly from scans, body meshes, or a combination of both

    do not rely on explicit SMPL priors. Instead, they learn volumetric representations directly from scans, body meshes, or a combination of both. While this increases flex- ibility, it tends to be computationally expensive—NASA re- quires per-subject training, while imGHUM relie...

Pith tools

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