{"id":"22c96520-c062-4339-9809-5a6b50542f53","arxiv_id":"2505.10205","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"VolE estimates food volume from phone video using ARKit/ARCore camera poses, COLMAP reconstruction, automatic food segmentation, and mesh volume integration.","lead":"VolE is a smartphone-based pipeline that reconstructs a 3D model of food from video and uses that model to estimate food volume without a reference object or depth sensor. It reports state-of-the-art volume accuracy on the MTF food benchmark, but the paper has unresolved questions about how the phone's AR scale is merged with the 3D reconstruction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The point-cloud masking defined in Sec. III-E (P = intersection over all views) would discard every surface point not visible in all frames, so a 360° free-motion capture cannot produce the dense object mesh the volume numbers require.","rationale":"The reader's CONDITIONAL verdict is right but for a weaker reason. The AR scale alignment is certainly unspecified: SfM is metric-less, and Sec. IV-C-1 only says ARKit coordinates 'can be used' to scale, without giving the similarity transform or alignment objective used after COLMAP refinement. That alone justifies a conditional. But the masking definition is more load-bearing because it makes the pipeline impossible as written regardless of scale. The abstract's 2.22% MAPE is also contradicted by Table III's 3.08%; this is a reporting error that must be corrected. I do not change the verdict: the paper should not be accepted without (i) a corrected, visibility-aware masking description, (ii) code/data release showing retained point clouds, (iii) the exact AR-to-COLMAP scale procedure, and (iv) reconciliation of the headline MAPE. The masking flaw could be a typographical simplification of the OpenMVS masking step, which is why I request the computational check rather than outright rejection.","tokens_in":18641,"tokens_out":10804,"duration_ms":107687,"concrete_test":"Implement the masking step exactly as written on one Foodkit-style 360° sequence: take the COLMAP point cloud and all FoodMem masks, project every point into every frame, keep only P = ∩ M_j, and count retained points. If the retained cloud is empty or has no connected surface for reconstructing a closed mesh (expected, since a point on one side is invisible in roughly half the frames), the published method cannot account for the reported volumes. The authors should instead release the actual masking code or define the visibility-aware rule.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. III-E defines M_j as points projecting into mask S_j, then takes P = ∩_{j=1}^{N_I} M_j. For any point on a food surface in a 360° acquisition, there are frames in which the point is out of view or occluded; its projection either falls outside S_j or is undefined, so it is excluded. The retained set is therefore empty or far too sparse for the Delaunay/graph-cut/marching-cubes mesh pipeline in Sec. III-F. This is not an AR-scale or metric issue: it is an internal contradiction in the published pipeline. The Foodkit captures are explicitly '360-degree view' (Sec. IV-C-1), and MTF scenes are multi-view, so the reported volume errors in Tables II and III cannot follow from the stated algorithm. A visibility-aware masking rule ('point must be inside the mask in every view where it is visible') would repair the text, but it is not what is written, and no such visibility test is specified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents VolE, a mobile-device-based framework for food volume estimation. The pipeline captures images and camera poses with ARKit/ARCore, refines poses with COLMAP, generates food masks with FoodMem, extracts a segmented point cloud, reconstructs and refines a mesh with OpenMVS, and computes volume via the divergence theorem. The authors introduce a new Foodkit dataset of 21 food objects with water-displacement ground truth and report experiments on Foodkit, MTF, and DTU. The headline claim is that VolE achieves 2.22% MAPE on MTF and 1.22% mean absolute error on Foodkit, outperforming prior volume-estimation methods without requiring reference objects or depth sensors.","tokens_in":18857,"tokens_out":8351,"duration_ms":78241,"significance":"If the technical issues are resolved, the idea of using AR-based pose priors to obtain metric scale without reference objects is practically valuable for mobile dietary assessment, and the Foodkit dataset addresses a real gap in food volume benchmarks. The paper is well-structured and includes useful robustness checks, such as repeated runs in the ablation tables. However, the central comparative claim is currently weakened by several load-bearing inconsistencies: the point-cloud masking rule as written is incompatible with 360-degree capture, the headline MAPE does not match the reported table, and the reference-free scaling procedure is not specified. These issues need to be addressed before the results can be considered reproducible.","major_comments":[{"comment":"The masking rule is defined as P = ∩_{j=1}^{N_I} M_j, meaning a scene point is retained only if its projection lies inside the food mask in every frame. For a 360-degree free-motion capture (Sec. IV-C-1) and for the multi-view MTF scenes, most surface points are occluded or out of view in a large fraction of frames, so this intersection will be empty or far too sparse to support the Delaunay/graph-cut/marching-cubes mesh pipeline in Sec. III-F. The volume errors in Tables II and III therefore cannot follow from the stated algorithm. Please replace the intersection rule with a visibility-aware rule (e.g., retain a point if it projects inside the mask in all frames where the point is actually visible, or in a sufficient subset) and re-run or re-report the affected experiments.","section":"Section III-E (Point Cloud Masking)"},{"comment":"The abstract and Sec. IV-D.2 state that VolE achieves 2.22% MAPE on the MTF dataset, but Table III reports a MAPE of 3.08%, and averaging the 13 listed per-scene error percentages for VolE gives 3.08%. This is a direct numerical inconsistency in the headline claim. Please correct the abstract/text or explain which subset of scenes the 2.22% figure refers to.","section":"Abstract and Section IV-D.2 vs. Table III"},{"comment":"The reference-free scale claim is not supported as written. Section III-B states that ARKit/ARCore positions provide metric scale, and Sec. IV-C-1 says ARKit coordinates are used to scale the reconstruction, but no alignment procedure between the AR poses and the COLMAP SfM reconstruction is specified (e.g., similarity-transform estimation, which frames are used, and how AR drift is handled). Furthermore, Sec. IV-C-2 states that on MTF 'we used the reference board to scale the reconstructed scenes to their original size,' which, if applied to VolE, would contradict the reference-free claim. Please specify the scale calibration for both datasets and clarify whether the MTF results are reference-free.","section":"Sections III-B, IV-C-1, and IV-C-2"},{"comment":"The Foodkit ground-truth volumes were measured by water displacement with a stated error margin of ±5 mL. Table II reports per-item mean absolute errors as low as 0.30–0.40 mL (e.g., Durum 0.40 mL, Yellow Cane 0.30 mL), far below the precision of the ground-truth reference. These errors should be interpreted against the ±5 mL GT uncertainty, and the paper should avoid implying sub-mL accuracy for volumes whose reference is only known to within ±5 mL.","section":"Section IV-C-1 and Table II"},{"comment":"The claim of outperforming existing volume estimators 'across multiple datasets' is only demonstrated for volume on MTF (Table III). Foodkit (Table II) contains no baseline volume comparisons, and DTU (Table IV) reports only Chamfer distance, not volume. Please state how the baseline numbers in Table III were obtained (official code, reimplementation, same masks and scaling) and add at least one volume-estimation baseline on Foodkit to support the comparative claim.","section":"Section IV-D and Tables II–IV"}],"minor_comments":[{"comment":"The phrase 'mononuclear data' should be 'monocular data'.","section":"Section I (Introduction)"},{"comment":"Two consecutive paragraphs both begin 'Table III presents...' and refer to different content (volume error and Chamfer distance) in the same table; please renumber or rename the table or headings for clarity.","section":"Section IV-D.2"},{"comment":"Reference [13] is corrupted: the citation string contains raw BibTeX fragments and a malformed author list. Please fix this entry.","section":"References"},{"comment":"The ID column in Table III jumps from 11 to 13 with no row 12; please renumber the scenes consistently.","section":"Table III"},{"comment":"The symbol P is used for both the complete scene point cloud and the final segmented object point cloud; please use distinct symbols to avoid confusion.","section":"Section III-E"},{"comment":"The sentence 'the error margin of ±5 mL comes from the scaler' presumably refers to the measuring cylinder or scale used in water displacement; please clarify the wording.","section":"Section IV-C-1"}],"recommendation":"major_revision","confidential_remarks":"The pipeline relies heavily on the authors' own prior works (VolETA, MVS-Boost, FoodMem), which are used as components or baselines without independent validation inside this manuscript, and the Foodkit dataset is not yet public. These issues increase the difficulty of independent verification. In addition, if the MTF evaluation used the reference board for scaling, the central comparison may not exercise the reference-free mechanism at all; the authors should be asked to clarify this explicitly in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe thing to know: this paper's headline number (2.22% MAPE on MTF) doesn't match its own Table III (3.08%), and the point-cloud masking step as written cannot work for 360° captures. Both are fixable, but they need to be fixed before anyone trusts the pipeline.\n\nWhat's genuinely new: the end-to-end combination of ARKit/ARCore poses with COLMAP refinement, FoodMem auto-masking, OpenMVS meshing, and tetrahedral volume integration. That integration is not in the prior work they cite, and the new Foodkit dataset (21 objects, water-displacement ground truth) is a useful resource. The experimental effort is real: multiple datasets, ablations on frame skipping and Hamming-distance selection, reasonable baselines. If the pipeline works, it addresses a genuine bottleneck: reference-free, phone-based food volume estimation for dietary monitoring.\n\nThe soft spots, in order of severity. First, the masking rule in Sec. III-E is an intersection over all views: P = ∩ M_j. For a 360° free-motion capture, most surface points are out of view or occluded in some frames, so the retained point cloud would be empty or far too sparse for the Delaunay/graph-cut/marching-cubes pipeline that follows. This is an internal contradiction with the stated acquisition protocol (Sec. IV-C-1 explicitly says 360-degree view). A visibility-aware rule would repair it, but that is not what is written. Second, the abstract and Sec. IV-D-2 claim 2.22% MAPE on MTF while Table III reports 3.08%; the discrepancy is not explained. Third, the scale alignment between ARKit poses and COLMAP is asserted but no algorithm or drift analysis is given; if AR scale is off, every volume is wrong. Fourth, code and dataset are not released, and the pipeline leans heavily on the authors' own prior components without independent validation.\n\nNone of these are fatal to the underlying idea. The masking issue is likely a typo (union instead of intersection, or a missing visibility test), the MAPE inconsistency is a reporting error, and the scale alignment can be specified. But as submitted, the paper cannot be verified from the text alone.\n\nWho this is for: people working on food volume estimation, mobile 3D reconstruction, and dietary assessment. It deserves a serious referee, but that referee should ask for the artifacts and a corrected method description. I'd engage with it after those fixes.","headline":"Useful integration paper with a load-bearing description bug: the all-views intersection masking rule contradicts the 360° capture protocol, the headline MAPE disagrees with Table III, and the AR-to-COLMAP scale alignment is underspecified.","tokens_in":19436,"tokens_out":2849,"would_cite":false,"duration_ms":27051,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"VolE claims that a reference-free mobile phone scan can estimate food volume with a mean absolute percentage error of 2.22% on the MTF benchmark, by using AR-tracked camera positions to give the 3D reconstruction real-world scale.","keywords":["food volume estimation","3D reconstruction","point cloud","augmented reality","structure from motion","video segmentation","dietary assessment","mobile phone scanning"],"falsifier":"Record a rigid object of known volume, such as a 200 ml measuring cup, with VolE under slow steady motion and under fast jerky motion, then compare estimated volumes to the true volume; if the fast capture or a long camera path pushes the error well above the reported 2-3%, the AR scale and drift assumption fails.","tokens_in":18423,"feed_emoji":"🍽️","tokens_out":10255,"duration_ms":97527,"temperature":0.7,"pith_summary":"VolE aims to establish that an ordinary AR-capable smartphone, moved freely around a plate of food, can estimate food volume precisely without a reference object, a depth sensor, or training on food shapes. The paper's load-bearing idea is that the phone's augmented-reality pose tracking gives each captured frame a 3D camera location in real-world units, and structure-from-motion refinement keeps that scale when the frames are reconstructed into a point cloud and mesh. On the MTF benchmark the paper claims a 2.22% mean absolute percentage error, and on its new 21-item Foodkit dataset it reports a 1.22% mean absolute error. The framework also outputs a refined 3D mesh, so testable claims include both volume accuracy and geometric fidelity, measured by Chamfer distance. If the scale assumption holds, this would remove the main practical barrier to everyday dietary volume logging: you just film the food.","feed_headline":"Phone AR scan estimates food volume to 2.22% error","feed_subtitle":"No reference card or depth camera needed: the phone's own AR coordinates scale the 3D model to real-world units.","key_machinery":"The load-bearing mathematics is the closed-mesh volume identity from the divergence theorem, $V = \\frac{1}{6}\\sum_{k=1}^{N} v_k^1 \\cdot (v_k^2 \\times v_k^3)$, which sums the signed volumes of tetrahedra formed by each triangular face and the origin. What makes that identity usable in real-world units is the metric scale supplied by the phone's AR pose tracker at capture time and preserved through structure-from-motion refinement; every downstream object—the dense point cloud, the masked food point set, and the final mesh—inherits that scale, so no reference object or depth sensor is needed. The mask-generation stage supplies per-frame food regions, and point-cloud masking retains only 3D points whose projection lies inside the mask in every view, so the tetrahedron sum runs over food geometry rather than background.","core_discovery":"The central discovery, on the paper's own terms, is that mobile augmented-reality camera poses can serve as the metric scaffold that replaces reference objects and depth information in food volume estimation. The pipeline captures a video of the food while the phone logs a 3D pose per frame, refines those poses with structure from motion, generates per-frame food masks with a video segmentation model, projects and intersects those masks into the scene point cloud to keep only food points, builds and refines a closed mesh, and computes volume as the sum of signed tetrahedra from the divergence theorem. The paper's evidence is the reported error rates and Chamfer distances across three datasets, including lower mean absolute percentage error than the compared baselines on the MTF benchmark and a lower mean Chamfer distance than several established reconstruction methods on the DTU multi-view benchmark.","pith_inferences":["Because nothing in the pipeline is food-specific, the same reference-free AR-scaffolded reconstruction could be transferred to other free-motion volume tasks, such as parcel sizing or wound measurement; the paper does not claim this.","A testable consequence of the AR-scale premise is that volume accuracy should degrade with path length or fast motion as AR tracking drifts; the paper reports neither a drift analysis nor an alignment check against a known object size.","Substituting manually drawn masks for the video segmentation output would isolate how much of the reported error comes from segmentation versus geometry, an ablation the paper does not run."],"forward_implications":["A user can log food volume on an ordinary phone by filming the meal, since the AR tracker supplies scale and the pipeline adds masks automatically.","The same captured video yields both a volume estimate and a refined 3D mesh, so dietary logging can be combined with shape-level analysis of the meal.","Selecting frames by visual dissimilarity can cut reported processing time by more than 90% while keeping accuracy above 97%, which points toward near-real-time use.","Low-textured foods and unbounded scenes, which the paper identifies as hard cases for NeRF-style reconstruction, are the cases the framework is designed to handle.","The new 21-object Foodkit dataset, once released, gives the field food items with water-displacement ground-truth volumes and masses for benchmarking future reference-free methods."],"supporting_citations":[{"why":"It supplies the structure-from-motion pose refinement and point-cloud generation that turn frames plus initial AR locations into a 3D scene.","marker":"[20]"},{"why":"It defines the prior winning method on the MTF benchmark whose volume and reconstruction results VolE compares against.","marker":"[1]"},{"why":"It generates the per-frame food masks used before point-cloud masking.","marker":"[6]"},{"why":"It is the phone's AR location service that records initial 3D camera coordinates in real-world units.","marker":"[27]"},{"why":"It is the alternative phone AR location service providing the same metric camera coordinates on the other platform.","marker":"[28]"},{"why":"It is the source of the MTF benchmark and the comparison baseline results reported for it.","marker":"[13]"},{"why":"It provides the mesh construction and refinement routines that create the closed mesh from the masked point cloud.","marker":"[39]"},{"why":"It is a neural implicit surface baseline whose DTU Chamfer distances VolE competes with.","marker":"[42]"},{"why":"It supplies the DTU benchmark scenes and ground-truth point clouds used to evaluate reconstruction accuracy.","marker":"[32]"}],"fun_headline_variants":["AR phone scan estimates food volume with 2.22% error","No depth camera or reference: AR poses scale food volume models","Food volume from phone AR: 2.22% mean absolute error","VolE: AR-based food volume estimation without reference objects","Mobile AR reconstructs food, estimates volume to 2.22%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method assumes the phone's augmented-reality tracker reports real-world distances accurately enough that no reference object is needed, and that these distances do not drift while the user moves the camera around the food.","fun_headline_variants_meta":{"raw":{"variants":["AR phone scan estimates food volume with 2.22% error","No depth camera or reference: AR poses scale food volume models","Food volume from phone AR: 2.22% mean absolute error","VolE: AR-based food volume estimation without reference objects","Mobile AR reconstructs food, estimates volume to 2.22%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1340,"prompt_tokens":878,"completion_tokens":462,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":372}},"tokens_in":494,"tokens_out":462,"duration_ms":3925,"temperature":1.0,"reasoning_tokens":372,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:14:10.477287+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a rigid object of known volume, such as a 200 ml measuring cup, with VolE under slow steady motion and under fast jerky motion, then compare estimated volumes to the true volume; if the fast capture or a long camera path pushes the error well above the reported 2-3%, the AR scale and drift assumption fails.","supporting_citations":[{"cited_title":"FoodMem: Near Real-time and Precise Food Video Segmentation","cited_arxiv_id":"2407.12121","evidence_quote":"It generates the per-frame food masks used before point-cloud masking."},{"cited_title":"Arcore overview,","cited_arxiv_id":null,"evidence_quote":"It is the phone's AR location service that records initial 3D camera coordinates in real-world units."},{"cited_title":"Arkit - augmented reality,","cited_arxiv_id":null,"evidence_quote":"It is the alternative phone AR location service providing the same metric camera coordinates on the other platform."},{"cited_title":"Openmvs: Open multiple view stereovision,","cited_arxiv_id":null,"evidence_quote":"It provides the mesh construction and refinement routines that create the closed mesh from the masked point cloud."},{"cited_title":"Neus2: Fast learning of neural implicit surfaces for multi- view reconstruction,","cited_arxiv_id":null,"evidence_quote":"It is a neural implicit surface baseline whose DTU Chamfer distances VolE competes with."},{"cited_title":"Large scale multi-view stereopsis evaluation,","cited_arxiv_id":null,"evidence_quote":"It supplies the DTU benchmark scenes and ground-truth point clouds used to evaluate reconstruction accuracy."}],"review_version":1}