{"id":"15ac6065-9428-4bcf-b062-c2885cbe652b","arxiv_id":"2505.21335","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A new model, SfC-NeRF, estimates hidden internal cavities of deformable objects from collision videos, using physics constraints and known mass, and demonstrates the idea on 115 simulated objects.","lead":"This paper introduces Structure from Collision (SfC), a method that estimates the invisible internal structure of an object, such as the location and size of a hidden cavity, from multiview videos of the object colliding with a surface. It matters because ordinary 3D reconstruction sees only the exterior, while SfC extracts information about the interior from the way the object deforms during impact.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline improvement over Static is largely purchased by the ground-truth mass prior: removing L_mass (Eq.","rationale":"The paper is honest about its assumptions: Section 3.1 states that physical properties including mass are known, and Appendix A.4 acknowledges real-world difficulty in obtaining them. The ablations in Tables 1 and 6 are exactly the right experiments and they reveal the sensitivity. Even without the mass loss, the method beats Static, so the video does carry some internal-structure signal; the central claim is not false, but its quantitative strength is conditional on an external measurement. This matches the reader's weakest-assumption identification. Other concerns—shared simulator physics, lack of released code/data, no error bars—are real but secondary; they affect generalizability and reproducibility rather than the internal logic of the method. The verdict should remain conditional: accept the contribution as a proof-of-concept under explicitly stated priors, but do not generalize beyond the known-physics, known-mass setting until the mass prior is relaxed or real-data validation appears.","tokens_in":1003,"tokens_out":852,"duration_ms":111458,"concrete_test":"Run Experiments I and II with the ground-truth mass loss (Eq. 7) replaced by an estimated mass/density parameter optimized jointly with sigma, keeping all other losses fixed. If the resulting CD for default cavities rises above about 0.5 (near the SfC-NeRF−mass value), the known-mass prior is load-bearing and the claim must be scoped to localization of a known-volume cavity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3 (Eqs. 7–12) makes the object's total mass m-hat a hard constraint, while the appearance-preserving losses (Eqs. 9–10) fix the external volume during Step (ii). Together these pin the total cavity volume before any video frame is used; the collision video can then only shift the cavity's location and shape within that fixed-volume envelope. Table 1 quantifies the effect: removing L_mass (and its learning-rate adaptation, Appendix C.3) degrades CD from 0.165 to 0.503, and Table 6 shows a −30% density error degrades CD to 0.798, near the Static baseline of 0.920. The central claim that videos reveal invisible internal structure is not refuted—SfC-NeRF−mass still beats Static (0.503 vs 0.720)—but the large reported margin depends on a measurement (total mass/density) that is neither observable in the video nor part of the task's raw input. This is a disclosed assumption (Section 3.1, Appendix A.4), not a hidden flaw, but it is load-bearing: it converts the ill-posed problem into 'locate a known-volume cavity,' which is substantially easier than 'discover the cavity's existence, size, and position from deformation alone.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new problem, Structure from Collision (SfC): given multi-view videos of an object colliding with a surface, estimate the object's 3D structure, including invisible internal cavities. The proposed model, SfC-NeRF, builds on PAC-NeRF and optimizes the initial volume-density field under physical constraints (known material properties and total mass), appearance-preserving losses, a keyframe loss, and a volume-annealing schedule. Experiments on a synthetic dataset of 115 objects across five external shapes, multiple cavity sizes/locations, and various materials report chamfer-distance improvements over a static reconstruction baseline and over ablated variants, together with a future-prediction application. The paper includes extensive appendices with per-shape results, component ablations, robustness to erroneous physical parameters, and collision-angle variations.","tokens_in":33221,"tokens_out":4548,"duration_ms":50122,"significance":"If the claims hold, the paper opens a useful new task and provides a systematic first solution with a thorough empirical study. It is honest about its assumptions, reports detailed ablations and per-shape numbers, and demonstrates that the full method improves over baselines even when the mass loss is removed (SfC-NeRF−mass 0.503 vs Static 0.720 in Table 1). The known-mass prior, however, is load-bearing for the headline margin, and the evaluation is entirely synthetic with the same MPM formulation used by the model, so the external validity of the central claim is not yet established. The paper's strength lies in its clearly framed problem, reproducible experimental protocol (detailed in Appendices), and the range of ablations isolating each component.","major_comments":[{"comment":"The known total mass m-hat is a hard constraint via L_mass, and the paper's own ablations show that this prior is load-bearing: removing it degrades CD from 0.165 to 0.503 (Table 1), and a -30% density error degrades CD to 0.798 (Table 6). Because the appearance-preserving losses fix the external volume, the mass prior effectively pins down the total cavity volume before any video frame is used. To support the claim that the collision video itself reveals the internal structure, the paper should include a control experiment that uses the mass prior and appearance-preserving losses but no temporal frames (i.e., a 'mass-only' optimization on t0). This control would quantify how much of the reported improvement is attributable to the prior versus the dynamics, and would strengthen the interpretation of the ablation results.","section":"§3.3, Eq. (7); Tables 1 and 6"},{"comment":"All headline quantitative claims are based on averages over five external shapes without any measure of variance or statistical significance. Given the wide per-shape spread (e.g., SfC-NeRF CD 0.067 for sphere vs 0.264 for cylinder in Table 16), the robustness of the reported improvements is unclear. The paper should report error bars (e.g., standard deviation over shapes, or over multiple random seeds) and, where the sample size allows, a paired significance test between the full model and its key ablations/baselines. At minimum, the range of per-shape scores should be stated in the main text alongside the averages.","section":"Tables 1–4 and 16–17"},{"comment":"The evaluation is conducted entirely on synthetic data generated with the MLS-MPM simulator, which shares the same MPM formulation (DiffMPM) used inside SfC-NeRF. This simulator-to-simulator agreement likely overstates the method's performance, since the physical prior matches the data-generation process exactly. The paper acknowledges real-data validation as future work (Section 5 and Appendix A.4), but this remains a central limitation: the claim that appearance changes during collision carry usable information about invisible internal structure would be substantially strengthened by a cross-simulator experiment (e.g., generating data with a different MPM or FEM solver) or by any real-world test. The authors should either add such a test or explicitly discuss the extent to which the current evaluation can support the general claim.","section":"§4.1 and Appendix C.1"}],"minor_comments":[{"comment":"The text says 'We also examined Stacic'—this should be 'Static'.","section":"§4.1"},{"comment":"The table caption reads 'Young's moduls'—this should be 'Young's modulus'.","section":"Table 3"},{"comment":"The related work contains the typo 'assumeed' in 'considereded non-transparent objects but assumeed that the interior is filled'—should be 'assumed'.","section":"§2"},{"comment":"The text uses 'SFC-NeRF' in several places (e.g., 'SFC-NeRF with only Lpixel0')—should be 'SfC-NeRF' for consistency.","section":"Appendix A.1.1"},{"comment":"In the definition of the depth-preserving loss, the integral for Z(r, t0) is written with a time variable t in the integrand (Tr(s,t) and σ(r(s),t)) rather than t0; please clarify whether this is a typo or an intentional notation.","section":"§3.3, Eq. (10)"},{"comment":"The full objective omits the background loss L_bg, which is described in Section 4.1 as being included in the pixel-related losses; please state this explicitly in the equation or its surrounding text to avoid confusion.","section":"§3.3, Eq. (12)"}],"recommendation":"major_revision","confidential_remarks":"The paper is by an author who also developed the LPO baseline; while the comparison itself is technically fair, the novelty assessment should weigh that the primary baselines are from the author's own prior line of work. The manuscript is detailed and the appendices are thorough, but the central claim's dependence on the known-mass prior and the lack of statistical support are likely to be the main points of contention in review. The paper may be a better fit for a journal that accommodates the extensive supplementary material, as the main text already appears dense."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things worth knowing. First, the SfC task is genuinely new: recovering hidden internal cavities of opaque deformable objects from collision video, where earlier reverse-engineering either assumed transparency or filled interiors. Second, the paper is upfront about its main assumption, and that assumption does more work than the abstract suggests.\n\nWhat's good: The problem framing is complementary to PAC-NeRF, and the method is a sensible assembly: a differentiable MPM simulator, appearance-preserving losses, a keyframe loss, and volume annealing. The experiments cover 115 synthetic objects across shapes, cavity sizes, locations, and materials, and the ablations isolate each component. Removing the mass loss still beats the static baseline (0.503 vs 0.720), so the deformation video does carry information about the cavity. The paper is also honest about its limitations: it states the known-physics assumption, notes the synthetic-only validation, and lists real-world challenges.\n\nThe soft spots are real and proportionate. The mass loss is the single biggest contributor: adding it improves CD from 0.503 to 0.165, and a -30% error in density degrades the result to 0.798, close to the static baseline. The paper discloses this in Section 3.1 and Appendix A.4, but the takeaway is that the headline margin depends on a measurement not present in the video. With mass known, the problem becomes 'locate a cavity of known volume,' which is still hard but is not the same as discovering the cavity from deformation alone. I also flag the absence of error bars (per-shape CD ranges 0.067 to 0.264 in Table 16), the shared MPM family between data generator and differentiable simulator, and no release of code or data.\n\nNone of this kills the paper. The scoped claim - internal structure can be recovered from collision video in a known-physics simulation setting - is supported. The ablation shows the video itself contributes beyond the static prior, and the limitations are stated rather than buried. The stress-test note slightly overstates when it says the improvement is 'largely purchased' by the mass prior; the no-mass version still beats the static baseline by a solid margin.\n\nWho this is for: people working on physics-informed neural 3D representations, system identification of deformable objects, and possibly non-destructive testing. It deserves a serious referee. I'd want to see error bars, code/data release, and a clearer decomposition of how much the mass prior contributes to the final result.\n\nRecommendation: yes to peer review, with requests for reproducibility and a more careful claim about what the video alone contributes.","headline":"A genuinely new task with a method that works in simulation; the headline numbers lean on the known-mass prior, but the core claim survives ablation.","tokens_in":33769,"tokens_out":4090,"would_cite":true,"duration_ms":40331,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Collision video can reveal the hidden internal structure of a deforming object, and SfC-NeRF recovers it by matching a physics simulation to the observed deformation.","keywords":["Structure from Collision","internal structure estimation","neural radiance fields","physics-informed learning","collision video","volume annealing","cavity estimation","differentiable physics simulation"],"falsifier":"Run SfC-NeRF on a real hollow object whose cavity is known from a CT scan, using only collision video and independently measured material properties, without providing the mass. If the recovered cavity no longer matches the scan, or if supplying an intentionally wrong mass by 30% changes the recovered cavity as sharply as the paper's density-error experiments show, then the known-mass constraint, not the video itself, is doing the decisive work.","tokens_in":32736,"feed_emoji":"💥","tokens_out":6661,"duration_ms":72166,"temperature":0.7,"pith_summary":"The paper introduces Structure from Collision (SfC), a task: recover an object's internal structure, including cavities hidden behind an opaque surface, from video of the object deforming as it collides. The proposed model, SfC-NeRF, starts from a static reconstruction of the object's outer surface and then adjusts the interior's volume density field so that a physics simulation of the collision matches the observed video, while preserving the learned exterior and using the object's total mass as a known constraint. On 115 simulated objects with varied cavity sizes, locations, and materials, the method locates the cavity better than models that treat the interior as filled, and it improves prediction of future collision frames. The paper's point is that deformation on impact carries usable information about what is inside, not just what is on the surface.","feed_headline":"Collision videos reveal what's inside a falling object","feed_subtitle":"A neural model reads how an object deforms on impact to estimate hidden internal cavities.","key_machinery":"The load-bearing object is the voxel-grid volume density field $\\sigma^{G'}(t_0)$ of the object at the first frame, which encodes where material is and where the cavity is. SfC-NeRF optimizes this field by coupling it to a differentiable material point method: particle samples of the field are simulated through the collision, rendered back to pixels, and compared with the video. Around that core sit the appearance-preserving losses (pixel and depth differences against the first frame), the keyframe pixel loss at the frame right after impact, the mass loss $L_{\\text{mass}}$ that pins total mass to the known value $\\hat{m}$, and volume annealing, which periodically expands the volume during optimization to escape local minima.","core_discovery":"The central claim is that the shape and location of an invisible internal cavity can be estimated from appearance changes during collision, even when static images of the object give no clue that the cavity exists. SfC-NeRF does this by taking the voxel-based volume density field learned from the first frame and optimizing the internal part of it so that a differentiable continuum-mechanics simulation, run forward through the video, reproduces the observed deformation. Three additional constraints keep the search well-posed: appearance-preserving losses hold the outer surface to the first frame, a keyframe loss emphasizes the frame just after impact, and a known-mass loss fixes the total amount of material, turning the problem from finding any cavity into placing a cavity of the right size. Volume annealing, which repeatedly shrinks and regrows the volume, helps avoid local optima. The supporting experiments argue that the method captures the positional bias of cavities across sizes and materials and outperforms baselines that optimize particles or grids without these constraints.","pith_inferences":["If the known-mass assumption is removed, the problem becomes appreciably harder; a natural extension would be to jointly estimate mass and structure, perhaps by exploiting the material-property ambiguity the paper acknowledges.","Because a hollow soft object and a filled stiff object can deform similarly, the method is most convincing when material properties are independently known; in real use, material identification and structure estimation would need to be solved together.","A direct real-data test, recovering the interior of a hollow object from a collision video and comparing with a CT or X-ray scan, would be the natural next validation step beyond simulation."],"forward_implications":["If the central claim holds, collision video is a usable observation modality for internal structure, not just for surface shape.","Optimizing the internal structure, rather than assuming the interior is filled, improves prediction of how the object will deform in future frames.","The method works across elastic materials, fluids, plasticine, and sand, with the best results for moderately deformable objects; very stiff or very soft objects give weaker cues.","Larger cavities are harder to recover because optimization starts from a filled volume, but volume annealing and the mass constraint reduce that gap.","The same optimization approach could in principle be attached to other neural 3D representations, since the constraints act on the density field rather than on a specific renderer."],"supporting_citations":[{"why":"Supplies the physics-augmented continuum neural radiance field backbone and dataset protocol that the proposed method extends.","marker":"[36]"},{"why":"Supplies the differentiable material point method used as the physics simulator for particles through the collision.","marker":"[26]"},{"why":"Provides the volume rendering formulation that turns density fields into pixel colors.","marker":"[47]"},{"why":"Provides the voxel-grid density representation that SfC-NeRF optimizes.","marker":"[63]"},{"why":"Defines the Lagrangian particle optimization baseline that the paper compares against.","marker":"[31]"},{"why":"Generates the ground-truth collision simulations used to build the dataset.","marker":"[25]"}],"fun_headline_variants":["Collision videos expose internal cavities","Impact footage maps hidden voids","SfC-NeRF: seeing inside by collision","Deformation reveals unseen inner structure"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The object's total mass is known in advance and is enforced as a hard loss; if the mass or density is misestimated, the recovered cavity volume is wrong even when the video is perfectly explained.","fun_headline_variants_meta":{"raw":{"variants":["Collision videos expose internal cavities","Impact footage maps hidden voids","SfC-NeRF: seeing inside by collision","Deformation reveals unseen inner structure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1477,"prompt_tokens":947,"completion_tokens":530,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":479}},"tokens_in":563,"tokens_out":530,"duration_ms":6607,"temperature":1.0,"reasoning_tokens":479,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:30:01.738856+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SfC-NeRF on a real hollow object whose cavity is known from a CT scan, using only collision video and independently measured material properties, without providing the mass. If the recovered cavity no longer matches the scan, or if supplying an intentionally wrong mass by 30% changes the recovered cavity as sharply as the paper's density-error experiments show, then the known-mass constraint, not the video itself, is doing the decisive work.","supporting_citations":[{"cited_title":"PAC-NeRF: Physics augmented continuum neural ra- diance fields for geometry-agnostic system identification","cited_arxiv_id":null,"evidence_quote":"Supplies the physics-augmented continuum neural radiance field backbone and dataset protocol that the proposed method extends."},{"cited_title":"Diff- Taichi: Differentiable programming for physical simulation","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable material point method used as the physics simulator for particles through the collision."},{"cited_title":"Srinivasan, Matthew Tancik, Jonathan T","cited_arxiv_id":null,"evidence_quote":"Provides the volume rendering formulation that turns density fields into pixel colors."},{"cited_title":"Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction","cited_arxiv_id":null,"evidence_quote":"Provides the voxel-grid density representation that SfC-NeRF optimizes."},{"cited_title":"Improving physics-augmented continuum neural radiance field-based geometry-agnostic system iden- tification with Lagrangian particle optimization","cited_arxiv_id":null,"evidence_quote":"Defines the Lagrangian particle optimization baseline that the paper compares against."},{"cited_title":"A moving least squares material point method with displacement discontinu- ity and two-way rigid body coupling","cited_arxiv_id":null,"evidence_quote":"Generates the ground-truth collision simulations used to build the dataset."}],"review_version":1}