Pith. sign in

REVIEW 4 major objections 5 minor 23 references

Vision Meets WiFi: Physics-Grounded Estimation of Volumetric Mechanical Properties

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read ViWi combines object-centric material slots with a simulated WiFi-band RF descriptor to estimate per-voxel Young's modulus, Poisson's ratio, and density, cutting Young's modulus error by about 54 percent over the prior voxel-wise baseline.

desk verdict The material-slot reformulation is solid and the vision-only ABO gains are credible, but the RF descriptor is simulated from the ground-truth material labels, so the headline GVM improvement is leakage, not a deployable WiFi gain. read the letter →

arxiv 2608.07726 v1 pith:LR7DRODB submitted 2026-08-07 cs.CV

Ali Bahri , Hongliang Li , Soufiane Lamghari , Jie Chuai , Zhitang Chen This is my paper · ORCID
classification cs.CV
keywords object-centriclearningmaterialpropertyestimationvolumetricmechanicalpropertiesvision-RFfusionWiFisensingslotattentionYoung'smodulusmultimodalperception
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

This paper tries to establish that volumetric mechanical-property estimation, meaning per-voxel Young's modulus, Poisson's ratio, and density, is improved by modeling an object as a small set of shared material slots and by conditioning those slots on a physics-based WiFi-band radio-frequency descriptor. The authors argue that vision alone cannot disambiguate visually similar objects with different material compositions, and that independent per-voxel regression ignores the piecewise-constant material structure of real objects. ViWi groups voxels into slots by visual and material-latent compatibility, predicts coherent slot-level property prototypes, and reconstructs voxel properties by soft assignment. On the GVM benchmark it reports the best performance on four of six per-voxel metrics, with about 54 percent lower Young's modulus errors than the prior method, while its vision-only variant improves all four mass-estimation metrics on real objects from ABO-500. If correct, the work suggests that combining object-centric structure with complementary RF evidence yields more physically coherent property fields than vision alone.

What carries the argument

The load-bearing machinery is the material-slot attention module together with RF-conditioned slot initialization. Slot Attention's iterative assignment and aggregation is adapted so each slot carries a feature embedding and a two-dimensional material latent in the frozen MatVAE space; the assignment score combines a vision-based attention term with a material-latent distance term, and normalization across slots forces competition among material hypotheses. The RF descriptor enters only through FiLM-style scale-and-shift modulation of the initial slot states, bounded by a constant and initialized at zero so training starts from the vision-only solution, plus a bounded post-iteration correction to the final material latent. This design keeps RF at the slot and object level while voxel-level visual features carry spatial localization, and the frozen MatVAE decoder guarantees that predicted triplets lie in the mechanically valid range.

What would settle it

A direct test would replace the simulated RF descriptor with real WiFi-band channel measurements taken around objects whose per-voxel material ground truth is known, then compare ViWi's error to its vision-only variant; the central claim fails if real-RF ViWi does not outperform the RF-disabled variant by a margin comparable to the simulated-RF gain, roughly 55 percent on Young's modulus and 1.63 percentage points on the hardest quintile. A cheaper partial check is to compare the simulated descriptor's ability to separate visually similar but materially different objects against a real measured descriptor on the same object set.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that object-centric material decomposition plus an RF descriptor makes dense mechanical-property prediction both more accurate and more physically coherent. Each voxel is softly assigned to one of K material slots using a score that combines visual compatibility with proximity in a pretrained material-latent space; slots aggregate evidence and emit shared property prototypes that the frozen material decoder maps to (E, nu, rho). A 64-dimensional RF descriptor, produced by WiFi-band electromagnetic simulation from ground-truth permittivity and conductivity, conditions the initial slot states through bounded, zero-initialized feature-wise modulation and a final RF-dependent correction, so the radio evidence says which materials are present while vision says where they are. This yields the reported reductions of roughly 54.6 percent and 54.3 percent in Young's modulus ALDE and ALRE relative to the voxel-wise baseline, along with Poisson's-ratio improvements of about 39 percent, at the cost of slightly worse density errors.

Load-bearing premise

The load-bearing premise is that the simulated WiFi-band RF descriptor, which is built from the object's ground-truth per-voxel electromagnetic properties, faithfully represents a real WiFi measurement that a deployed system could obtain at inference; if real RF readings differ from the simulation, or if the descriptor cannot be produced without already knowing the materials being predicted, the reported RF gains will not transfer.

Editorial extensions

If this is right

  • Young's modulus ALDE and ALRE drop by roughly 54.6 percent and 54.3 percent relative to the voxel-wise baseline, from 0.3952 to 0.1793 and from 0.0427 to 0.0195.
  • Poisson's ratio ADE and ARE improve by about 38.8 percent and 39.3 percent, from 0.0245 to 0.0150 and from 0.0842 to 0.0511.
  • The vision-only variant of ViWi improves mass estimation on ABO-500 on all four reported metrics, with ADE falling from 37.37 to 20.18 kilograms and MnRE rising from 0.395 to 0.415.
  • RF gains concentrate on visually difficult objects: the mean gain rises from negative 0.04 percentage points in the easiest quintile to positive 1.63 percentage points in the hardest quintile, where 67 percent of objects improve.
  • ViWi degrades gradually under RF descriptor noise, with mean error rising from 4.99 percent with clean RF to 5.74 percent at sigma equal to 1.0, close to the 5.56 percent of the RF-disabled variant.

Reading between the lines

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

  • If the simulated RF descriptor can be replaced by real WiFi-band measurements that approximate it, the method offers a path to material-composition inference under occlusion or darkness where vision is unreliable; the paper itself flags real RF observations as future work.
  • The material-slot formulation is not tied to mechanical properties, so the same grouping machinery could apply to other dense volumetric fields with piecewise-constant structure, such as thermal, electrical, or acoustic property maps.
  • The composition-supervision loss, which trains the RF embedding to predict the object-level material distribution, suggests the RF embedding could serve as a standalone material-classification signal, a claim that is testable on GVM without requiring full volumetric labels.
  • As stated, the RF descriptor is computed from the object's ground-truth per-voxel electromagnetic properties, meaning inference still requires the very material map the model aims to predict; a deployed system must either measure RF directly or estimate the descriptor from its own visual predictions, and the paper's experiments do not close that loop.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces ViWi, a framework for volumetric estimation of Young's modulus, Poisson's ratio, and density from multi-view images augmented with a 64-dimensional RF descriptor. The RF descriptor is produced by WiFi-band electromagnetic simulation from per-voxel permittivity and conductivity values that are mapped from the object's ground-truth mechanical material categories. On the synthetic GVM benchmark, ViWi reports improvements over VoMP on four of six per-voxel metrics, including roughly 54% reductions in Young's modulus ALDE/ALRE; on the real-object ABO-500 benchmark, a vision-only variant of ViWi improves all four mass-estimation metrics. The paper argues that object-centric material slots plus complementary RF evidence yield more accurate and physically coherent property estimates.

Significance. The slot-attention reformulation of volumetric property estimation is a reasonable and potentially useful idea, and the vision-only ABO-500 mass-estimation improvements are a credible positive result that should be credited. However, the paper's headline claim that 'Vision Meets WiFi' improves mechanical property estimation rests entirely on the GVM experiments, and those experiments are invalidated by test-time label leakage: the RF descriptor is a deterministic function of the ground-truth material map that also defines the target properties. Because the privileged descriptor is available only when the target is already known, the reported four-of-six GVM gains do not support the central claim. The paper is clearly written and includes an explicit limitation statement, but the main empirical evidence for RF fusion is not sound; the remaining contribution is a vision-only object-centric grouping result that does not match the paper's title and framing.

major comments (4)
  1. [Section 3.3, Section 4.1, Table 1] The RF descriptor is computed by mapping each ground-truth mechanical material category to relative permittivity and conductivity, running a WiFi-band simulation, and aggregating the results into a fixed-size vector. The descriptor is therefore a deterministic function of the per-voxel material labels that also determine the ground-truth E, nu, and rho. At inference on the GVM benchmark, the model receives this privileged descriptor, so the comparison against the vision-only VoMP baseline measures label leakage rather than a deployable sensing gain. The paper itself states in Section 4.1 that RF input is unavailable for ABO-500 because descriptors are simulated from per-voxel electromagnetic properties and the real objects lack such annotations; this confirms that the RF input is not a measurement that can be obtained at inference without knowing the target material map. Consequently, the four-of-six metric improvements and the roughly 54% Young's modulus reduction in Table 1 cannot be attributed to 'Vision Meets WiFi'.
  2. [Eq. (9), Figure 4] The composition loss L_comp explicitly trains the RF embedding to predict the ground-truth material-category histogram h_b. Combined with the frozen MatVAE decoder, this creates a direct pathway from the privileged RF embedding to the mechanical properties of the materials present, so the model can bypass visual evidence entirely for composition. The ablation in Figure 4, which shows that RF gains increase with vision-only difficulty, is exactly what would be expected under leakage: the descriptor is most informative precisely when it reveals material composition that vision cannot disambiguate. This experiment therefore does not demonstrate complementary sensing; it demonstrates that the leaked composition signal is useful when vision is weak.
  3. [Section 4.2, Table 3] The robustness experiment in Table 3 adds Gaussian noise to the standardized RF descriptor at inference, but it does not address the validity of the descriptor as a sensor. A noisy version of a privileged descriptor remains privileged: the descriptor still originates from the ground-truth material map. To support the central claim, the authors would need to (i) acquire RF measurements from real objects or from a forward model that is not conditioned on the target material labels, and (ii) demonstrate that the simulated descriptor is a faithful proxy for such measurements. The gradual degradation under noise does not provide this evidence.
  4. [Section 5, Table 2] The paper's only uncontaminated empirical support is the vision-only mass-estimation result on ABO-500 (Table 2). That result is a useful contribution, but it does not support the 'Vision Meets WiFi' headline or the claim that RF evidence improves volumetric mechanical-property estimation. Since every GVM comparison involving the RF pathway is confounded, the paper overclaims relative to what its experiments establish.
minor comments (5)
  1. [Title and Abstract] The terms 'WiFi' and 'Vision Meets WiFi' are used throughout, but no WiFi hardware or measured WiFi signal is involved; the RF descriptor is a simulated fingerprint. The framing should be softened or the simulation status stated prominently in the abstract.
  2. [Table 1] The density rows show ViWi underperforming VoMP, so the 'best on four of six metrics' framing should be accompanied by an explicit discussion of the density regression; the current narrative emphasizes only the improved metrics.
  3. [Table 3] The 'mean error' column averages three relative errors with different scales and units. Report per-metric errors and define the aggregate normalization explicitly.
  4. [Figure 2] The distance-percentile comparison is based on a single object pair; state how many pairs were evaluated and show a distribution rather than a single percentile to support the claim of visual ambiguity and RF separability.
  5. [Section 4.2] The quintile analysis uses 'the combined validation and test sets' to rank objects by vision-only error. This post hoc grouping should be described as exploratory, not as a held-out evaluation, and the paper should clarify whether any hyperparameters were selected using this analysis.

Circularity Check

2 steps flagged · score 8.0 of 10

The GVM RF descriptor is simulated from the ground-truth material map that defines the target mechanical properties, so the headline RF gains reflect target leakage rather than a deployable complementary sensing modality.

  1. self definitional [Section 3.3 (Physics-Based RF Representation) and Section 4.1 (Main Results)]
    "Each mechanical material category is mapped to representative electromagnetic parameters, namely relative permittivity ϵr and conductivity σ, which determine how RF waves interact with the object. ... RF input is unavailable for ABO-500 because our descriptors are simulated from per-voxel electromagnetic properties, whereas these real objects provide only a single ground-truth mass and no per-voxel material annotations from which to simulate RF responses."

    The target outputs (E, ν, ρ) are defined per mechanical material category, and the RF descriptor is simulated from those same categories via permittivity and conductivity. At inference on GVM, the model therefore receives a descriptor that is a deterministic function of the ground-truth material map it is asked to reconstruct. The Table 1 gains — including the 54.6% ALDE reduction for Young's modulus — do not demonstrate that WiFi evidence complements vision; they demonstrate that the simulated RF channel carries the target material labels into the model.

  2. fitted input called prediction [Equation (9), Section 3.5 (Learning Objectives)]
    "In addition, the RF embedding is supervised to predict the object-level material composition. Let hb ∈ ∆C denote the ground-truth material-category distribution of object b ... The composition loss, defined over objects B with valid RF composition targets, encourages the RF embedding to encode which materials are present."

    L_comp explicitly trains the RF embedding to output the ground-truth material-category histogram h_b. Because the per-voxel mechanical properties E, ν, and ρ are generated from these same material categories, the RF pathway is fitted to carry the target itself. The subsequent improvement in mechanical-property prediction is then reported as evidence that 'combining object-centric material structure with complementary RF evidence' works, but the RF embedding has been optimized to encode exactly the composition that determines the prediction target.

full rationale

The central GVM claim — that RF evidence beyond vision improves volumetric mechanical-property estimation — is circular by construction. The RF descriptor is not a measurement obtained from a deployable sensor; it is simulated from per-voxel electromagnetic properties mapped from the ground-truth mechanical material categories, which are the same categories that determine the ground-truth E, ν, and ρ fields. Additionally, the composition loss in Eq. (9) trains the RF embedding to predict the ground-truth material-category distribution, making the leakage explicit. The paper's own Section 4.1 confirms that RF descriptors require per-voxel material annotations and are therefore absent on ABO-500. Consequently, the four-of-six GVM improvements, especially the roughly 54% Young's modulus error reductions, cannot be attributed to 'Vision Meets WiFi' in any real deployment; they reflect target information injected through the simulated RF channel. The vision-only ABO-500 mass-estimation results and the material-slot grouping mechanism remain independent and credible contributions, but they do not support the headline multimodal RF claim. The paper acknowledges only that RF is simulated rather than measured, not that the simulation leaks the target, so the limitation statement does not mitigate the circularity.

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

The central RF contribution rests on the untested assumption that simulated descriptors derived from ground-truth material maps behave like real measurements, and on several hand-set hyperparameters. The vision-only slot-attention path inherits the frozen VoMP backbone and MatVAE without revalidation.

free parameters (5)
  • Number of material slots K = 16
    Model capacity choice; K=16 shown in Fig. 3; not derived from data or theory.
  • RF descriptor dimension D = 64
    D = T × F with F = 4 statistics per transmitter; D=64 stated in Fig. 3; transmitter/receiver layout not specified in main text.
  • Auxiliary loss weights lambda_mat, lambda_comp = not reported
    Weights in Eq. (10) controlling material-latent and composition losses; values omitted from the main text.
  • FiLM bound c = not reported
    Bounds the ctanh scale/shift and the final RF correction Delta m_k in Eqs. (6) and (7); value not given.
  • Material-to-electromagnetic mapping (epsilon_r, sigma) = mapping table not shown
    Each mechanical material category is mapped to representative permittivity and conductivity for simulation in Section 3.3; the mapping table is not in the main text and determines the RF descriptor.
assumptions (5)
  • domain assumption Simulated electromagnetic propagation at WiFi bands is a faithful proxy for real RF measurements of the object
    The entire RF pathway rests on this; the paper itself notes the descriptors are simulated and do not capture all practical sources of RF variation (Section 5).
  • domain assumption The RF descriptor is computable at inference without access to ground-truth material composition
    In GVM the descriptor is simulated from ground-truth per-voxel electromagnetic properties; the paper does not describe how a deployed system would obtain it (Section 3.3, Fig. 3(c)).
  • domain assumption MatVAE frozen from VoMP maps a 2D latent to physically valid (E, nu, rho) triplets
    ViWi inherits the pretrained decoder without revalidating it on GVM ground truth (Section 3.1).
  • domain assumption Voxel-level features from the pretrained DINOv2 geometry backbone capture local material structure
    The backbone is from VoMP and frozen; its feature quality bounds the spatial localization of slots (Section 3.1).
  • standard math Standard electrodynamics simulation is accurate
    The paper relies on unstated EM simulation software to compute scattered power and delay spread (Section 3.3, Eq. (5)).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vision Meets WiFi: Physics-Grounded Estimation of Volumetric Mechanical Properties." pith.science (2026). https://pith.science/paper/LR7DRODB

@misc{pith2026260807726,
  author       = {Pith},
  title        = {Pith review of: Vision Meets WiFi: Physics-Grounded Estimation of Volumetric Mechanical Properties},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LR7DRODB}},
  note         = {Machine review of arXiv:2608.07726}
}
read the original abstract

Estimating volumetric mechanical properties, including Young's modulus, Poisson's ratio, and density at each voxel, is intrinsically ambiguous from vision alone, as visually similar objects may have substantially different material compositions and physical behavior. Existing approaches predict these properties independently across voxels, overlooking the piecewise-constant material structure of real objects and producing noisy or inconsistent estimates for voxels that share the same material, while lacking an explicit mechanism to resolve visual ambiguity. We introduce ViWi (Vision Meets WiFi), an object-centric framework for volumetric mechanical-property estimation. ViWi represents each object using a compact set of material slots that aggregate evidence from voxels with a shared material identity and produce coherent slot-level property predictions. To complement visual appearance, ViWi incorporates a compact RF descriptor generated through WiFi-band electromagnetic simulation using permittivity and conductivity. The RF descriptor conditions the material slots with global composition cues that may be unavailable from images, while visual features preserve voxel-level spatial localization. Across volumetric mechanical-property and mass-estimation benchmarks, ViWi improves over the prior state of the art on four of six per-voxel metrics, while its vision-only variant improves all mass-estimation metrics. These results demonstrate that combining object-centric material structure with complementary RF evidence enables more accurate and physically coherent volumetric property estimation beyond what is possible from visual appearance alone.

Figures

Figures reproduced from arXiv: 2608.07726 by the authors.

Figure 1
Figure 1. From volumetric mechanical-property estimation to physics simulation. ViWi predicts voxel-level Young’s modulus, Poisson’s ratio, and density from the input 3D object. The predicted material fields are then used to simulate the object’s physical response to a ball impact. ABSTRACT Estimating volumetric mechanical properties, including Young’s modulus, Poisson’s ratio, and density at each voxel, is intrinsically ambi… view at source ↗
Figure 2
Figure 2. Visual ambiguity and RF separabil￾ity. (a) Visually similar objects with different material compositions. (b) The pair overlaps in the visual embedding but is clearly separated in the RF embedding. Understanding how objects deform, resist external forces, and dis￾tribute mass is essential for physically grounded perception. Ap￾plications such as robotic manipulation, physics-based simulation, digital twins, augmente… view at source ↗
Figure 3
Figure 3. Overview of ViWi. Multi-view RGB features are aggregated on a sparse voxel grid and grouped into material slots using visual and material-latent compatibility. Each slot material latent is decoded by the frozen MatVAE into E, ν, and ρ, and voxel-level properties are reconstructed through soft assignments. A simulated 64-D RF descriptor conditions the initial material-slot states through bounded feature-wise modulati… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: RF gains increase with visual difficulty. Ob￾jects are grouped into equal-count quintiles by ViWi† (Vision Only) error. Bars show mean RF gain; positive values indicate improvement. The gain is small for easy objects and reaches +1.63 pp on the hardest quintile. We fur…
Figure 5
Figure 5. Figure 5: Qualitative comparison of volumetric mechanical-property estimation. For four representative GVM objects, we compare the ground-truth fields with predictions from VoMP [4] and ViWi for Young’s modulus E, Poisson’s ratio ν, and density ρ. A shared color scale is used wi…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 11 canonical work pages

  1. [1]

    Physical property understanding from language-embedded feature fields

    Albert J Zhai, Yuan Shen, Emily Y Chen, Gloria X Wang, Xinlei Wang, Sheng Wang, Kaiyu Guan, and Shenlong Wang. Physical property understanding from language-embedded feature fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28296–28305, 2024

    work page 2024

  2. [2]

    Pugs: Zero-shot physical understanding with gaussian splatting

    Yinghao Shuai, Ran Yu, Yuantao Chen, Zijian Jiang, Xiaowei Song, Nan Wang, Jv Zheng, Jianzhu Ma, Meng Yang, Zhicheng Wang, et al. Pugs: Zero-shot physical understanding with gaussian splatting. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 4478–4485. IEEE, 2025

    work page 2025

  3. [3]

    Phys4dgen: Physics-compliant 4d generation with multi-material composition perception

    Jiajing Lin, Zhenzhong Wang, Dejun Xu, Shu Jiang, Yunpeng Gong, and Min Jiang. Phys4dgen: Physics-compliant 4d generation with multi-material composition perception. InProceedings of the 33rd ACM International Conference on Multimedia, pages 10398–10407, 2025

    work page 2025

  4. [4]

    V omp: Predicting volumetric mechanical property fields.arXiv preprint arXiv:2510.22975, 2025

    Rishit Dagli, Donglai Xiang, Vismay Modi, Charles Loop, Clement Fuji Tsang, Anka He Chen, Anita Hu, Gavriel State, David IW Levin, and Maria Shugrina. V omp: Predicting volumetric mechanical property fields.arXiv preprint arXiv:2510.22975, 2025

    arXiv

  5. [5]

    Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T

    Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y . Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T. Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation.arXiv preprint arXiv:2404.13026, 2024

    arXiv 2024

  6. [6]

    Tianyu Huang, Haoze Zhang, Yihan Zeng, Zhilu Zhang, Hui Li, Wangmeng Zuo, and Rynson W. H. Lau. Dreamphysics: Learning physics-based 3d dynamics with video diffusion priors.arXiv preprint arXiv:2406.01476, 2024

    arXiv 2024

  7. [7]

    Physics3d: Learning physical properties of 3d gaussians via video diffusion.arXiv preprint arXiv:2406.04338, 2024

    Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. Physics3d: Learning physical properties of 3d gaussians via video diffusion.arXiv preprint arXiv:2406.04338, 2024

    arXiv 2024

  8. [8]

    Physgaussian: Physics-integrated 3d gaussians for generative dynamics.arXiv preprint arXiv:2311.12198, 2023

    Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics.arXiv preprint arXiv:2311.12198, 2023

    arXiv 2023

Show all 23 references
  1. [9]

    Physgen: Rigid-body physics-grounded image-to-video generation

    Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded image-to-video generation. InEuropean Conference on Computer Vision, pages 360–378. Springer, 2024

    2024

  2. [10]

    Phys- gen3d: Crafting a miniature interactive world from a single image

    Boyuan Chen, Hanxiao Jiang, Shaowei Liu, Saurabh Gupta, Yunzhu Li, Hao Zhao, and Shenlong Wang. Phys- gen3d: Crafting a miniature interactive world from a single image. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6178–6189, 2025

    2025

  3. [11]

    Phys4dgen: A physics-driven framework for controllable and efficient 4d content generation from a single image.arXiv preprint arXiv:2411.16800, 2024

    Jiajing Lin, Zhenzhong Wang, Shu Jiang, Yongjie Hou, and Min Jiang. Phys4dgen: A physics-driven framework for controllable and efficient 4d content generation from a single image.arXiv preprint arXiv:2411.16800, 2024

    2024

  4. [12]

    Pixie: Fast and generalizable supervised learning of 3d physics from pixels.arXiv preprint arXiv:2508.17437, 2025

    Long Le, Ryan Lucas, Chen Wang, Chuhao Chen, Dinesh Jayaraman, Eric Eaton, and Lingjie Liu. Pixie: Fast and generalizable supervised learning of 3d physics from pixels.arXiv preprint arXiv:2508.17437, 2025

    2025 arXiv

  5. [13]

    Physsplat: Efficient physics simulation for 3d scenes via mllm-guided gaussian splatting

    Haoyu Zhao, Hao Wang, Xingyue Zhao, Hao Fei, Hongqiu Wang, Chengjiang Long, and Hua Zou. Physsplat: Efficient physics simulation for 3d scenes via mllm-guided gaussian splatting. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 5242–5252, 2025

    2025

  6. [14]

    Sophy: Generating simulation-ready objects with physical materials.arXiv preprint arXiv:2504.12684, 2025

    Junyi Cao and Evangelos Kalogerakis. Sophy: Generating simulation-ready objects with physical materials.arXiv preprint arXiv:2504.12684, 2025

    2025 arXiv

  7. [15]

    Physx-3d: Physical-grounded 3d asset generation.arXiv preprint arXiv:2507.12465, 2025

    Ziang Cao, Zhaoxi Chen, Liang Pan, and Ziwei Liu. Physx-3d: Physical-grounded 3d asset generation.arXiv preprint arXiv:2507.12465, 2025

    2025

  8. [16]

    Object-centric learning with slot attention.arXiv preprint arXiv:2006.15055, 2020

    Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention.arXiv preprint arXiv:2006.15055, 2020

    2006 arXiv

  9. [17]

    The field-based model: A new perspective on rf-based material sensing.arXiv preprint arXiv:2412.05640, 2024

    Fei Shang, Haocheng Jiang, Panlong Yang, Dawei Yan, Haohua Du, and Xiang-Yang Li. The field-based model: A new perspective on rf-based material sensing.arXiv preprint arXiv:2412.05640, 2024

    2024 arXiv

  10. [18]

    Hupr: A benchmark for human pose estimation using millimeter wave radar.arXiv preprint arXiv:2210.12564, 2022

    Shih-Po Lee, Niraj Prakash Kini, Wen-Hsiao Peng, Ching-Wen Ma, and Jenq-Neng Hwang. Hupr: A benchmark for human pose estimation using millimeter wave radar.arXiv preprint arXiv:2210.12564, 2022

    2022 arXiv

  11. [19]

    Rfpose-ot: Rf-based 3d human pose estimation via optimal transport theory.arXiv preprint arXiv:2301.13013, 2023

    Cong Yu, Dongheng Zhang, Zhi Wu, Zhi Lu, Chunyang Xie, Yang Hu, and Yan Chen. Rfpose-ot: Rf-based 3d human pose estimation via optimal transport theory.arXiv preprint arXiv:2301.13013, 2023

    2023 arXiv

  12. [20]

    Diffusion model is a good pose estimator from 3d rf-vision.arXiv preprint arXiv:2403.16198, 2024

    Junqiao Fan, Jianfei Yang, Yuecong Xu, and Lihua Xie. Diffusion model is a good pose estimator from 3d rf-vision.arXiv preprint arXiv:2403.16198, 2024. 10 arXivTemplateA PREPRINT

    2024 arXiv

  13. [21]

    Improving real-time omnidirectional 3d multi-person human pose estimation with people matching and unsupervised 2d–3d lifting.arXiv preprint arXiv:2403.09437, 2024

    Pawel Knap, Peter Hardy, Alberto Tamajo, Hwasup Lim, and Hansung Kim. Improving real-time omnidirectional 3d multi-person human pose estimation with people matching and unsupervised 2d–3d lifting.arXiv preprint arXiv:2403.09437, 2024

    2024 arXiv

  14. [22]

    Film: Visual reasoning with a general conditioning layer

    Ethan Perez, Florian Strub, Harm De Vries, Vincent Dumoulin, and Aaron Courville. Film: Visual reasoning with a general conditioning layer. InProceedings of the AAAI conference on artificial intelligence, volume 32, 2018

    2018

  15. [23]

    Abo: Dataset and benchmarks for real-world 3d object understanding

    Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In2022 IEEE/CVF Conference on Computer Vision an...

    2022

Pith tools

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