acceptodds
Under review as a conference paper at ICLR 2027

PhysVGGT: Feed-Forward Dense Physical Property Estimation from A Single Image

Abstract

Physical properties, such as friction, hardness, stiffness, and density, govern how robots should grasp, manipulate and interact with objects, yet estimating these properties from RGB images remains challenging. Existing methods typically employ per-object reconstruction augmented with physical properties or directly query vision-language models at test time, which results in substantial computational overhead that limits their applicability. In this work, we present PhysVGGT, a feed-forward model that predicts dense maps of friction coefficient, Shore hardness, Young's modulus, and density, together with object-level mass, from a single RGB image in one forward pass. The key idea of PhysVGGT is to formulate physical property estimation as a dense per-pixel prediction problem and employ a visual geometry transformer to extract geometry-aware tokens from the input image followed by a dense prediction branch for estimating local physical properties and a global prediction branch for estimating object-level mass. In addition, we introduce a scalable pseudo-label generation pipeline that enables large-scale weakly supervised training for dense physical property prediction, substantially reducing the need for expensive direct physical measurements. Extensive experiments show that PhysVGGT achieves state-of-the-art performance on the ABO-500 dataset and generalizes effectively to the out-of-distribution NeRF2Physics dataset. Moreover, PhysVGGT eliminates the need for per-object reconstruction and test-time optimization, achieving an inference latency of only 0.13 s per image, making it faster than the previous state of the art.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Physical Property Estimation, Dense Prediction, Single-Image Physical Property Estimation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aKCAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents PhysVGGT, a feedforward method to predict dense maps of object material properties from a single image. The author uses a frozen StreamVGGT backbone, a dense prediction branch with cross-property attention and shared decoder, and a mass head. The authors introduce a pseudo-labeling pipeline combining part-segmentation, VLM reasoning and cross-view refinement to obtain the training data. Experiments on ABO-500 and NeRF2Physics demonstrate the accuracy of the method.

Critical strengths
  • Studying feedforward object material property estimating is useful.
  • The use of pretrained VGGT model and overall pipeline is sound.
  • The authors compared different backbones, frozen strategies, decoder, losses in the experiments.
Critical weaknesses
  • How is the methods results applied to real-world inputs?
  • WIth the development and effectiveness of VLM models, how does this work compared to directly using VLMs for the task?
  • Tables 5 and 11 report substantial mass improvements from cross-property coupling, however the mass head is seperate from DPT, how could that happen?
  • The authors didn't explain how multiview training transfers to single-view inference.The backbone comparison also does not isolate whether StreamVGGT’s advantage comes from its causal design or something else.
Questions for authors

Why do the reproduced NeRF2Physics and PhysGS results differ substantially from their original reports? Is the evaluation protocal different?

3 comments

We thank the reviewer for the valuable comment. We have revised the manuscript to address the concerns (marked in blue in the revised manuscript). Our responses are detailed below.

Q3-1: How are the method's results applied to real-world inputs?

Response: Thanks for the question. The predicted physical-property maps and object-level mass estimates from PhysVGGT can serve as inputs to various real-world downstream tasks. Specifically, PhysVGGT takes real-world RGB images as input and produces dense physical-property maps and object-level mass estimates through feed-forward inference, without requiring VLM prompting or material-property lookup. For robotic manipulation, the predicted friction maps can be further converted into dense grip-force maps, as illustrated in Figures 8 and 1(c). These maps provide spatial guidance for selecting grasp locations and adjusting grip forces according to local surface friction, favoring high-friction regions while applying larger forces to slippery regions. For outdoor scenes, the predicted friction and Young's modulus maps (Figure 5) can be used as additional environmental perception cues for terrain-aware navigation, traversability estimation, and slip-risk assessment. Thus, PhysVGGT provides dense physical information that can be incorporated into downstream robotic manipulation and autonomous systems.

Q3-2: How does this work compared to directly using VLMs for the task?

Response: Unlike directly prompting VLMs for physical-property estimation, PhysVGGT uses VLMs only during pseudo-label generation and does not require VLM inference at test time. PhysVGGT learns to map image-specific visual features from the geometry encoder backbone to dense physical-property maps through lightweight trainable prediction heads. To resolve this concern, we additionally provide quantitative comparisons against direct VLM inference, as shown below. These results demonstrate that PhysVGGT provides more accurate physical-property estimation than direct VLM prompting, while offering dense feed-forward prediction without requiring VLM inference at deployment.

NeRF2PhysicsHardness ADE ↓\downarrowALDE ↓\downarrowAPE ↓\downarrowMnRE ↑\uparrow
GPT-4V32.750.3300.3040.758
CLIP32.860.2940.2660.774
Qwen3 (Direct Prompting)17.920.3020.2770.776
PhysVGGT7.050.1030.1110.908
PhysXNet-100Density MAE (kg/m$^3$) ↓\downarrowEE MAE (GPa) ↓\downarrow
Qwen3 (Direct Prompting)159464
PhysVGGT90033

Q3-3: In Tables 5 and 11 (Tables 6 and 12 in the revised manuscript), why cross-property coupling and dense objective losses improve mass prediction despite the mass head being separate?

Response: Thanks for pointing out this issue. We apologize for the confusion caused by the presentation of the mass results in Tables 5 and 11.

  1. The mass values of variants in Tables 5 and 11 (Tables 6 and 12 in the revised manuscript) are derived from dense density predictions and estimated volume, m^=ρˉVκ\hat{m}=\bar{\rho}V\kappa, rather than directly from the independent global mass prediction head. Our original purpose was to investigate how cross-property coupling and dense objective losses affect density predictions and, consequently, density-derived mass estimation. Since CPC and dense prediction objectives influence the predicted density maps, their effects can propagate to the derived mass estimates. This does not imply any direct influence on the independent global mass head.
  2. To ensure consistent comparisons, we have corrected the Full-model derived mass results in Tables 5 and 11 to MnRE = 0.517 and ADE = 10.30, respectively, and have updated the corresponding column headers to "Derived Mass (m^=ρˉVκ\hat{m}=\bar{\rho}V\kappa) MnRE ↑\uparrow" and "Derived Mass (m^=ρˉVκ\hat{m}=\bar{\rho}V\kappa) ADE (kg) ↓\downarrow". In our default architecture, the dense prediction branch and global prediction branch share only the frozen backbone features, while their trainable prediction modules remain independent.

Q3-4: How does multi-view training transfer to single-view inference, and does StreamVGGT's advantage come from its causal design?

Response:

  1. Multi-view training and single-view inference. StreamVGGT supports variable-length image sequences, and the dense prediction head shares parameters across frames. During causal training, the first frame has no preceding context and is supervised by the same per-frame prediction objective. Therefore, single-view inference is naturally supported by setting the sequence length to one, without architectural modifications. The dense predictions follow the same prediction pipeline described in Section~3.1. We further evaluate PhysVGGT with multiple views for dense prediction. The results show that PhysVGGT maintains comparable performance across different input view counts, but additional views do not necessarily bring performance gains.
Input ViewsMass ADE (kg) ↓\downarrowDensity MAE (kg/m$^3$) ↓\downarrowEE MAE (GPa) ↓\downarrow
14.7390033
24.6590133
44.5390834
  1. Backbone comparison and causal design. We have compared PhysVGGT with various backbones in Table 8, but this comparison does not isolate the effect of causal design from other architectural and pretraining differences. Isolating the contribution of causal design would require a controlled comparison with a non-causal StreamVGGT variant using the same backbone architecture and training configurations. In the revised manuscript, we have clarified that the results demonstrate the effectiveness of StreamVGGT as the backbone in our framework, rather than attributing its advantage exclusively to its causal design.

Q3-5: Why do the reproduced NeRF2Physics and PhysGS results differ substantially from the original reports?

Response: Our evaluation protocol is consistent with that of existing works and benchmarks. We follow the original paper descriptions to reproduce NeRF2Physics and PhysGS independently. Both NeRF2Physics and PhysGS rely on per-scene 3D reconstruction and vision-language-based material/property estimation, making their predictions sensitive to implementation details, model configurations, and reconstruction results. We have clarified this distinction in the revised manuscript. Nevertheless, despite the discrepancies between our independent reproductions and the originally reported results, PhysVGGT still outperforms the originally reported NeRF2Physics and PhysGS results.

Dear Reviewer,

Thank you for your time and consideration. We hope our response has addressed your concerns, particularly regarding the comparisons with direct VLMs. If any questions or concerns remain, we would be grateful for the opportunity to provide further clarification during the discussion period.

Thanks.

Reviewer_3yWoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addressed estimating physical properties from visual observations. The proposed PhysVGGT predicts dense per-pixel physical-property maps and object-level mass from a single RGB image, using trainable heads with a StreamVGGT geometry backbone. The paper introduces a pseudo-labeling pipeline that combines SAM 2 part segmentation, VLM-based material prediction, and cross-view refinement, providing dense supervision over ABO objects. The method is evaluated on ABO-500 and the NeRF2Physics datasets, achieving state-of-the-art accuracy.

Critical strengths

The proposed PhysVGGT directly predicts dense physical properties and object mass from a single RGB image, achieving state-of-the-art performance. The scalable pseudo-label generation pipeline enables training of dense physical property prediction models.

Critical weaknesses
  1. Although the paper claims dense prediction, dense property evaluation is conducted only on NeRF2Physics sparse measurements for friction and hardness (Table 2). The claims need to be appropriately scoped.
  2. Hardness supervision comes from a VLM, and Young's modulus is derived from predicted Shore hardness via an empirical relation. Further discussion is required on how the reported hardness accuracy is derived from VLM priors vs. learned from visual features.
  3. It appears that multiple views are used in the backbone, while inference uses a single image. How does the model perform when given more views?
  4. There is qualitative terrain generalization on RUGD/RELLIS-3D. Can the proposed method be generalized and applied to robotics/autonomous driving?
Questions for authors

See weakness. This work presents a cheaply trainable model on a frozen backbone. The pseudo-label pipeline is effective on a real dataset. However, more discussions and descriptions are required regarding the claim scope, the VLM-derived hardness labels, and reported hardness accuracy. The OOD NeRF2Physics evaluation is provided. Further descriptions on VLM-prompting baseline on NeRF2Physics and an ablation of the multi-view-training/single-view-inference gap would be helpful.

2 comments

We thank the reviewer for the valuable comment. We have revised the manuscript to address the concerns (marked in blue in the revised manuscript). Our responses are detailed below.

Q2-1: Dense prediction claims need to be appropriately scoped?

Response: Thanks for the reminder.

  1. Dense prediction evaluation. We evaluate on NeRF2Physics because it is a widely used dataset for image-based physical-property prediction, even though it provides only sparse measurements. To resolve this concern, we additionally evaluate on PhysXNet-100, which provides dense, human-reviewed annotations of Young's modulus and density, thereby demonstrating the dense prediction capability of PhysVGGT.
PhysXNet-100Density MAE (kg/m$^3$) ↓\downarrowEE MAE (GPa) ↓\downarrow
NeRF2Physics204468
PUGS129768
SiPhy131552
PhysVGGT90033
  1. Clarification of claim scope. We clarify that PhysVGGT is capable of dense prediction, but quantitative evaluation is limited to datasets with available annotations. NeRF2Physics provides sparse measurements, while PhysXNet-100 provides dense, human-reviewed annotations. This distinction is now clearly stated in the revised manuscript.

Q2-2: How much does hardness accuracy come from VLM priors versus learned visual features?

Response: A VLM provides material and Shore hardness pseudo-labels during training, introducing material-property priors into the supervision. However, PhysVGGT does not use VLM predictions at inference. Instead, it learns to map image-specific visual features to dense hardness predictions through trainable prediction heads. To investigate whether the hardness accuracy merely reflects the VLM's prior knowledge, we additionally compare PhysVGGT against Qwen3 direct predictions on the NeRF2Physics dataset. PhysVGGT's substantially lower errors indicate that the learned feed-forward predictor provides benefits beyond direct VLM prompting. Nevertheless, we acknowledge that its supervision inherits VLM-derived priors and clarify this limitation in the revised manuscript.

NeRF2PhysicsHardness ADE ↓\downarrowALDE ↓\downarrowAPE ↓\downarrowMnRE ↑\uparrow
Qwen3 direct predictions17.920.3020.2770.776
PhysVGGT7.050.1030.1110.908

Q2-3: Multi-view performance?

Response:

  1. The geometry encoder (StreamVGGT) supports variable-length image sequences, while the dense prediction head shares its parameters across frames. During causal training, the first frame is supervised without preceding context. Therefore, single-view inference is naturally supported by setting the input sequence length to one, without architectural modifications.
  2. We further evaluate PhysVGGT with multiple views for dense prediction. The results show that PhysVGGT maintains comparable performance across different input view counts, but additional views do not necessarily bring performance gains.
Input ViewsMass ADE (kg) ↓\downarrowDensity MAE (kg/m$^3$) ↓\downarrowEE MAE (GPa) ↓\downarrow
14.7390033
24.6590133
44.5390834

Q2-4: Can the proposed method be generalized and applied to robotics/autonomous driving?

Response: Yes.

  1. As illustrated in Figures 8 and 1(c), PhysVGGT can directly derive a dense grip-force map that favors high-friction surfaces for grasping while assigning substantially larger forces to slippery regions. Thus, the predicted physical properties provide spatially grounded cues for both where to grasp and how hard to squeeze, illustrating their potential utility for downstream robotic manipulation.
  2. As for autonomous driving, Figure 5 illustrates how PhysVGGT predicts dense friction and Young's modulus maps for outdoor scenes, demonstrating its potential to generalize to real-world outdoor environments. In driving scenarios, dense physical-property maps can serve as additional perception cues for terrain-aware navigation, traversability estimation, and slip-risk assessment.
  3. In the revised manuscript (Figure 9), we have further provided zero-shot qualitative results on robotic manipulation (DROID) and autonomous driving (KITTI) scenarios.

Dear Reviewer,

Thank you for your time and consideration. We hope our response has addressed your concerns, particularly regarding the evauation of dense prediction and comparisons with VLM priors. If any questions or concerns remain, we would be grateful for the opportunity to provide further clarification during the discussion period.

Thanks.

Reviewer_VYCMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PhysVGGT, a feed-forward model for estimating physical properties from a single RGB image. The model uses a frozen StreamVGGT backbone with two prediction branches: a dense branch for estimating the coefficient of friction, Shore hardness, Young’s modulus, and density, and a global branch for directly regressing object mass. The authors also propose a pseudo-label generation pipeline based on SAM 2, a vision-language model, a material-property database, and cross-view voting. On ABO-500, PhysVGGT achieves better mass estimation performance than the compared methods. On the real-world NeRF2Physics dataset, it also reports improved friction and hardness estimation. The paper further reports an inference latency of 0.13 seconds per image. The main strengths of the method are its efficiency and system-level integration. However, Young’s modulus and density are not independently evaluated quantitatively, several important details of the pseudo-label generation pipeline are missing, and the comparisons with related methods and evaluation datasets are limited. These issues make the method difficult to reproduce and its overall effectiveness difficult to assess comprehensively.

Critical strengths
  1. The paper is clearly written and easy to follow.
  2. It introduces an efficient single-image inference framework. PhysVGGT combines a frozen geometry-aware backbone with lightweight physical-property prediction heads, avoiding test-time NeRF/3DGS optimization and VLM queries. The reported 0.13-second latency is potentially valuable for real-time robotic perception.
  3. The results on friction, hardness, and mass are reasonably competitive. On ABO-500 and NeRF2Physics, the paper reports improvements over the compared baselines and provides a reasonable set of ablation studies. These results suggest that the proposed framework has practical potential.
Critical weaknesses
  1. My major concern is that Young’s modulus and density are not independently validated quantitatively. These properties are derived indirectly through the pseudo-labeling pipeline, while the paper only provides qualitative visualizations. The model may therefore be learning priors from the VLM and material-property tables rather than recovering the true physical quantities from images. This issue directly affects the paper’s central claim.
  2. The comparison with closely related work is incomplete. In particular, the paper does not compare with SLAT-Phys [1] and SiPhy [2], which also study single-image physical-property estimation. Both works are highly relevant to the paper’s claims regarding single-image inference, dense material-property prediction, and fast feed-forward estimation.
  3. The experimental evaluation is limited in terms of datasets. The real measurements in NeRF2Physics are sparse, while RUGD and RELLIS-3D do not provide physical-property ground truth. The paper does not evaluate on relevant datasets such as PixieVerse [3], MVImgNet-100 [4], or PhysXNet-100 [5]. In particular, PhysXNet-100 provides dense annotations for density and Young’s modulus and would be highly relevant for validating the paper’s main claims.

[1] Rocktim Jyoti Das and Dinesh Manocha. SLAT-Phys: Fast Material Property Field Prediction from Structured 3D Latents. arXiv, 2026. [2] Hoang Le, Joonwoo Kwon, Elkhan Ismayilzada, Yufei Zhang, and Zijun Cui. SiPhy: Single-Image Physical Property Reasoning. ECCV 2026. [3] 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. 2025. [4] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, Guanying Chen, Shuguang Cui, and Xiaoguang Han. MVImgNet: A Large-Scale Dataset of Multi-View Images. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9150–9161. [5] Ziang Cao, Zhaoxi Chen, Liang Pan, and Ziwei Liu. PhysX-3D: Physical-Grounded 3D Asset Generation. NeurIPS 2025.

Questions for authors
  1. Please provide the complete 12-material vocabulary, the material-property table, and the exact formula or empirical relationship used to convert Shore hardness into Young’s modulus.
  2. Please specify the VLM version, prompts, confidence threshold, number of views used per object, and the details of the cross-view refinement procedure.
3 comments

We thank the reviewer for the valuable comment. We have revised the manuscript to address the concerns (marked in blue in the revised manuscript). Our responses are detailed below.

Q1-1: a) Young's modulus and density are not validated quantitatively? b) The model may therefore be learning priors from VLM and material property tables rather than recovering physical quantities?

Response: Thanks for the reminder.

  1. Quantitative evaluation. In the original manuscript, we only provide qualitative results for Young's modulus (EE) and density (ρ\rho) because the datasets used in our evaluation do not provide independently measured and spatially dense annotations for these properties. To address this concern, we further conduct quantitative evaluation on the PhysXNet-100 dataset, which provides human-reviewed physical-property annotations. The results demonstrate that PhysVGGT achieves lower errors in predicting both EE and ρ\rho compared to existing methods, such as SiPhy (ECCV26).
PhysXNet-100Density MAE (kg/m$^3$) ↓\downarrowEE MAE (GPa) ↓\downarrow
NeRF2Physics204468
PUGS129768
SiPhy131552
PhysVGGT90033
  1. Comparison with VLM priors and material-property lookup. To further investigate whether PhysVGGT only learns priors from VLM and material property tables, we compare PhysVGGT with Qwen3 using direct prompting and material-property lookup. The results show that PhysVGGT outperforms both methods, suggesting that the learned feed-forward framework provides benefits beyond direct VLM prompting and material-property lookup by learning representations for dense property prediction.
PhysXNet-100Density MAE (kg/m$^3$) ↓\downarrowEE MAE (GPa) ↓\downarrow
Qwen3 (direct prompting)159464
Qwen3 + Material Lookup124864
PhysVGGT90033

Q1-2: Incomplete comparison with closely related works, particularly SLAT-Phys and SiPhy?

Response: Thanks for bringing these works to our attention. We have cited these works in the revised manuscript.

  1. Discussion. SLAT-Phys employs TRELLIS-based 3D latent representations for voxel-level physical-property prediction, while SiPhy leverages VLMs and CLIP for single-image physical-property reasoning. In contrast, PhysVGGT uses a visual geometry transformer backbone with lightweight prediction heads for direct feed-forward dense physical-property estimation. Note that PhysVGGT can achieve an inference latency of only 0.13s per image.
  2. Comparison. To resolve the concern, we compare with SiPhy on the ABO-500 and PhysXNet-100 datasets in the tables below. We can see that PhysVGGT consistently outperforms SiPhy on both datasets, demonstrating its effectiveness. As for SLAT-Phys, we note that it predicts 3D voxel-level physical-property fields, whereas PhysVGGT directly predicts dense 2D image-space maps. Therefore, the original evaluation results of SLAT-Phys are not directly comparable to those obtained by PhysVGGT. Reproducing SLAT-Phys would require substantial effort and access to their 3D voxel-level predictions, which is beyond the scope of this work.
ABO-500 (Mass prediction)ADE (kg) ↓\downarrowALDE ↓\downarrowMnRE ↑\uparrow
NeRF2Physics12.730.7360.564
PUGS9.460.6610.576
SiPhy7.780.740.58
PhysVGGT4.730.5740.655
PhysXNet-100Density MAE (kg/m$^3$) ↓\downarrowEE MAE (GPa) ↓\downarrow
NeRF2Physics204468
PUGS129768
SiPhy131552
PhysVGGT90033

Q1-3: Limited dataset evaluation, particularly PhysXNet-100?

Response: Thanks for the reminder. We have cited these works in the revised manuscript.

  1. Our original dataset selection serves complementary purposes. ABO-500 test objects enable quantitative evaluation of object-level mass estimation, NeRF2Physics and image2mass further provide real-world physical-property and mass evaluation, respectively, while RUGD/RELLIS-3D assess generalization to outdoor scenes.
  2. To address the concern, we evaluate PhysVGGT on PhysXNet-100. PhysVGGT achieves lower errors than existing methods on both properties, demonstrating its effectiveness.
  3. PixieVerse provides 3D voxel-level physical-property fields, which differ from PhysVGGT's image-space dense prediction protocol and require additional 3D-to-2D alignment for a fair comparison. MVImgNet-100 provides multi-view images and geometric annotations but lacks physical-property GT for quantitative evaluation of EE and density.
PhysXNet-100Density MAE (kg/m$^3$) ↓\downarrowEE MAE (GPa) ↓\downarrow
NeRF2Physics204468
PUGS129768
SiPhy131552
PhysVGGT90033

Q1-4: Complete 12-material vocabulary, material-property table, and Shore-hardness-to-Young's-modulus formula/relationship??

Response:

  1. 12-material vocabulary and material-property table. The implementation uses 12 material categories, including fabric, foam, wood, metal, plastic, glass, ceramic, paper, leather, rubber, food, and stone.
IDMaterialDensity (kg/m³)Friction coefficient μ\muPoisson's ratio ν\nuYoung's modulus range (GPa)
1fabric2000.600.301e-05–0.001
2foam500.400.451e-06–5e-05
3wood6000.500.351–15
4metal78000.400.3050–200
5plastic12000.350.380.05–4
6glass25000.250.2250–80
7ceramic22000.300.2250–400
8paper7000.400.300.1–2
9leather8600.550.400.01–0.5
10rubber11000.800.490.0001–0.05
11food8000.300.451e-06–0.005
12stone27000.450.2510–100
  1. Shore hardness to Young's modulus conversion. We use a material-constrained heuristic to convert the predicted Shore hardness to approximate Young's modulus. Given a predicted hardness interval [Hlow,Hhigh][H_{\mathrm{low}},H_{\mathrm{high}}], we first calculate its arithmetic midpoint:H=Hlow+Hhigh2.H=\frac{H_{\mathrm{low}}+H_{\mathrm{high}}}{2}. The empirical conversion is:
\begin{cases} \operatorname{clip}_c(10^6 \cdot 10^{(H-50)/40}), & s=\mathrm{Shore\\ A},\\\\ \operatorname{clip}_c(10^6 \cdot 10^{(H-30)/20}), & s=\mathrm{Shore\\ D},\\ H\leq85,\\\\ E_{\max,c}, & s=\mathrm{Shore\\ D},\\ H>85. \end{cases}$$ where $s$ denotes the Shore hardness scale, and $$\operatorname{clip}_c(x)=\min(E_{\max,c},\max(E_{\min,c},x)).$$ Here, $E_{\min,c}$ and $E_{\max,c}$ are the material-specific Young's modulus bounds converted from GPa to Pa. The factor $10^6$ converts the empirical output from MPa to Pa. ### Q1-5: **VLM version, prompts, confidence threshold, number of views used per object, and the details of the cross-view refinement procedure**? **Response:** 1. VLM version: Qwen3-VL-32B-Instruct. 2. Prompts: The prompt asks the VLM to identify the intrinsic material from the 12-material vocabulary and predict its Shore hardness interval, scale, confidence, and functional description. The output format is below. The complete prompts are provided in the Appendix (see Section A.7 in pages 22-24 in the revised manuscript). `material | hardness_low-hardness_high | Shore A/D | confidence | part_description` 3. Confidence threshold: 0.1. 4. Number of views used per object: 15 calibrated RGB views per object. 5. Details of the cross-view refinement procedure: We first independently generate part-level predictions for each view. All labeled views are tiled into a composite image, with each segmented part annotated by its predicted material and functional description. The VLM jointly examines the composite image and per-view predictions, and outputs corrections in the format of (view ID, part ID, corrected material). The corrections are aggregated through majority voting for each (view, part) pair and applied to obtain refined material labels.

Dear Reviewer,

Thank you for your time and consideration. We hope our response has addressed your concerns, particularly regarding comparisons with state-of-the-art methods (and on more datasets), the evaluation of Young's modulus and density, and some missing details. If any questions or concerns remain, we would be grateful for the opportunity to provide further clarification during the discussion period.

Thank you!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.