{"id":"43f83982-925e-444f-9534-bbbec74f5b0f","arxiv_id":"2507.11037","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"FootGait3D provides graded partial-to-complete real-world point clouds of the foot-ankle complex during gait, enabling benchmarking of occlusion-robust 3D completion.","lead":"The paper introduces FootGait3D, a multi-view point cloud dataset of the foot and ankle captured from 46 people walking barefoot, totaling 8,403 frames. It pairs complete 5-camera reconstructions with partial views from fewer cameras, creating a benchmark for 3D shape completion on deformable biological tissue.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth completeness of the 5-view reconstructions is asserted, not demonstrated; if systematic missing regions exist, the benchmark's completion target is biased.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the 5-view reconstruction is treated as ground-truth complete geometry, but this is only visually inspected and the synchronization limitations are acknowledged. I agree that this is the most fundamental risk. If the ground-truth complete cloud has systematic holes, the paired partial-complete structure loses its meaning and all reported completion metrics become biased. I considered two other concerns. First, the train/test split is described by instance count, not by subject, so same-subject frames could appear in both training and test and inflate benchmark scores; this is important but secondary, and it can be checked after the split metadata is released. Second, for multi-modal baselines it is ambiguous whether the \"five-view depth maps\" are restricted to the available camera subset in each partial condition; this affects only the reported benchmark numbers, not the dataset itself, and is a clarity issue rather than a fundamental flaw. The proposed static validation directly tests whether the 5-view reconstruction is complete enough to serve as ground truth, using equipment already available in the acquisition setup. The verdict remains conditional because the dataset is promising, but the ground-truth completeness claim should be quantitatively validated before the benchmark is relied upon.","tokens_in":14396,"tokens_out":7561,"duration_ms":105090,"concrete_test":"Perform a static validation study with about 10 subjects: while a foot is stationary in the capture zone, record it with the 5-camera PFA rig and independently with a high-resolution structured-light scanner (e.g., Artec or EinScan). Register the two surfaces and compute (a) two-sided Chamfer distance, (b) the fraction of scanner surface with no PFA point within 2 mm, and (c) a region-wise missing-area map for the arch, heel, medial/lateral sides, and toes. If any region shows a consistent uncovered fraction above a small threshold (e.g., >5%) or if the 2 mm coverage criterion fails in the same regions that the bottom camera is expected to cover, the ground-truth completeness assumption is falsified and the benchmark's ground truth needs revision or explicit caveats.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central product is paired partial and complete point clouds, so the 5-view reconstruction must actually be a complete surface for the dataset's ground-truth claim to hold. In Section 3.3.3, completeness is only established by visual inspection, and the only quantitative fidelity statement, \"surface reconstruction accuracy below 2 mm RMSE,\" is inherited from the PFA system description rather than measured on these dynamic foot captures. The paper itself acknowledges synchronization offsets that are most visible at heel-strike and toe-off and are only manually filtered (Section 5). Because partial clouds are generated by dropping camera subsets before merging (Section 3.3.2), any region that the retained cameras cannot see is absent from both input and ground truth; if such regions are systematic (e.g., the arch under weight bearing, the heel at strike, or areas affected by specular reflections cropped from the transparent-plate bottom view), every completion model is trained and evaluated against a target that is itself incomplete. Visual inspection cannot rule out these systematic missing regions, and model rankings could change once the true complete surface is known. This is the load-bearing assumption: the benchmark's usefulness depends on the 5-view cloud being a reliable complete surface, and that is not quantitatively established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"FootGait3D is a dataset paper presenting a multi-view, high-resolution point cloud collection of the foot-ankle complex during the stance phase of gait. The data were acquired with a five-camera depth-sensing system (the previously reported PFA system): four horizontal cameras and one camera beneath a transparent walkway plate. The dataset contains 8,403 frames from 46 subjects; for each frame, a five-view merged point cloud serves as ground truth, and partial point clouds are generated from all combinations of four, three, and two horizontal cameras (with the bottom camera always omitted), yielding 100,836 point cloud frames in total. The authors benchmark eleven point cloud completion methods (PCN, PoinTr, AnchorFormer, SnowflakeNet, PointAttn, SVDFormer, PointSea, CSDN, EGIINet, XMFNet, MAENet) using Chamfer Distance and F1-score, and report that SnowflakeNet, PointAttn, SVDFormer, and PointSea perform best. The dataset is publicly available on HuggingFace.","tokens_in":14533,"tokens_out":16339,"duration_ms":175135,"significance":"The resource is potentially valuable: it is the first real-world dataset, to my knowledge, that provides paired partial and complete dense point clouds of the foot-ankle complex during dynamic gait, with graded occlusion levels. If the ground-truth completeness is validated, it would enable direct training and evaluation of completion models on realistic dynamic anatomical data, complementing static foot datasets (Foot3D, SynFoot) and synthetic completion benchmarks (PCN, ShapeNet). The public release and the systematic benchmarking against standard architectures are strengths, as is the use of a previously described capture system with known hardware characteristics. However, the paper's central assumption—that the five-view reconstruction is a complete and accurate surface—is not quantitatively established, and the evaluation protocol contains ambiguities that must be resolved before the benchmark numbers can be used.","major_comments":[{"comment":"The ground-truth completeness claim is not quantitatively demonstrated. Section 3.3.3 states that full-view clouds were visually inspected and that the system has 'surface reconstruction accuracy below 2 mm RMSE,' but no measurement protocol is given for this accuracy and it is not shown to apply to the dynamic foot captures in FootGait3D. Because partial clouds are generated by dropping camera subsets before merging (Section 3.3.2), any region invisible to the retained cameras and also absent from the five-view merge—for example, the plantar contact patch under weight bearing, the arch during mid-stance, or surfaces occluded by the contralateral swing foot before removal—is missing from both the input and the ground truth. The paper's own Section 5 acknowledges residual synchronization offsets that were manually filtered. If such systematic missing regions exist, the completion target is biased and model rankings could change. Please add quantitative completeness evaluation (e.g., comparison with independent high-resolution static scans, repeated-trial reconstruction consistency, or per-frame hole statistics) and specify how the 2 mm RMSE figure was obtained for the dynamic data.","section":"§3.3.3; §3.3.2; §5"},{"comment":"The F1-score threshold is inconsistently defined. Equation (2) and Section 4.1 specify τ = 0.001 (1e-3), while the captions of Tables 1 and 2 define F1 at '1‰ of the bounding-box diagonal.' These two definitions coincide only if the point clouds are normalized to a unit bounding box, but no such normalization is described. Without normalization, a threshold of 1 mm combined with CD-L1 values of 17–19 × 10^-3 (i.e., average distances in the 17–19 mm range if coordinates are in meters) is hard to reconcile with F1 values near 0.97. Please state explicitly whether and how the coordinates were normalized before metric computation, provide the exact threshold used, and recompute the affected numbers if needed.","section":"§4.1; Tables 1–2"},{"comment":"The evaluation protocol is under-specified in two respects that affect the benchmark conclusions. First, the text says all 16384-point predictions and ground-truth clouds were downsampled to 2048 points with FPS before metric computation, yet the tables also present parenthetical values for 16384 predicted points; the paper does not describe how the 16384-point ground truth is constructed from the original clouds (which contain roughly 78k–120k points) nor how the parenthetical and non-parenthetical numbers are related. Second, the train/validation/test split is described only as 'based on the instance count' (5881/841/1681); it is not stated whether the split is subject-independent. If frames from the same subject appear in both the training and test sets, the reported performance overestimates generalization to new individuals. Please specify the split construction and the exact metric computation pipeline, including which resolution each reported number corresponds to.","section":"§4.3; Tables 1–2"},{"comment":"The paper does not document the format and organization of the released dataset. It only states that the dataset is available on HuggingFace and that sensor calibration matrices are provided. For a dataset contribution, the paper should describe the file formats for the point clouds and depth maps, coordinate units and axis conventions, directory and naming structure, the exact train/validation/test split used in the benchmark, and the license under which the data are released. Please add a data card or an appendix with this information.","section":"§3; Data Availability"}],"minor_comments":[{"comment":"The phrase 'All subjects all free from lower limb, foot, and ankle injuries' should be corrected to 'All subjects were free from lower limb, foot, and ankle injuries.'","section":"§3.2"},{"comment":"In the PointAttn row, the 2-view F1 value '0.997' is missing its closing parenthesis; the model name is written 'PointAttn' in the table but 'PointAttN' in the text.","section":"Table 1"},{"comment":"The claim that 'AnchorFormer's performance steadily degrades as the number of views decreases' is not supported by Table 1: the 4-view CD-L1 is 18.944 and the 3-view CD-L1 is 18.174, which is an improvement. Please reconcile the text with the table.","section":"§4.4"},{"comment":"The sentence claiming that 'PointAttn performs worst under the 4-view condition' contradicts Table 1, where PointAttn has the best (lowest) CD-L1 among single-modal methods at 4 views. This appears to be a typo and should be corrected.","section":"Supplementary B.1"},{"comment":"The caption contains grammatical errors: 'showed' should be 'show' and 'This reflect' should be 'This reflects.'","section":"Figure 6"},{"comment":"For the multi-modal methods other than SVDFormer, the paper does not specify how their architectures were adapted to the five-view depth maps and extrinsics. Please provide a brief description or refer to a code repository.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one-line take: FootGait3D is a genuinely new resource—the first dynamic foot-ankle point cloud dataset with paired partial and complete real captures and graded occlusion levels. If you do point cloud completion or biomechanical foot modeling, this is worth knowing about.\n\nWhat the paper does well: the capture and preprocessing pipeline is described in enough detail to reproduce, the partial views are generated sensibly (drop camera subsets before merging, always removing the bottom camera), and the benchmark covers a reasonable spread of single- and multi-modal baselines. The authors are candid about synchronization offsets and manual filtering, and the dataset is publicly released. That's real value.\n\nThe soft spots are real but patchable. The main one is the ground-truth completeness claim. The 5-view reconstruction is treated as the complete surface, but the only evidence is visual inspection in Sec. 3.3.3. The 'sub-2 mm RMSE' fidelity statement is inherited from the earlier PFA system paper, not measured on these dynamic captures. Because partial inputs are generated by dropping cameras, any region that none of the retained cameras can observe is absent from both the input and the target. If those regions are systematic—under the arch, at heel-strike—the benchmark trains and evaluates models against a target that is itself incomplete. The stress-test note pushes this hard, and the concern lands. The fix is straightforward: validate completeness against a high-fidelity static scan for a subset of subjects, or report per-frame coverage maps showing that missing regions are not concentrated in a few anatomical zones. I'd also like to see variance across train/test splits, since only a single split is reported. Minor issues: the RMSE measurement protocol is missing, Eq. (2) for F1 looks misprinted, and the baseline adaptation code isn't released. All addressable.\n\nThe central idea is sound and the dataset fills a real gap. The validation gaps do not sink it; they just need to be closed before the benchmark claims are bulletproof. This deserves serious peer review—conditional accept with a request for quantitative completeness validation. I'd bring it to a reading group.","headline":"First dynamic foot-ankle point cloud dataset with paired partial-complete views; a useful resource that needs quantitative ground-truth completeness validation before its benchmark claims are bulletproof.","tokens_in":15115,"tokens_out":3340,"would_cite":true,"duration_ms":37633,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FootGait3D pairs partial foot scans with full ground truth.","keywords":["FootGait3D","point cloud completion","gait analysis","foot-ankle complex","multi-view depth sensors","occlusion robustness","benchmark dataset","markerless motion capture"],"falsifier":"Take a subject's foot, obtain a high-resolution static 3D scan in a pose matching a FootGait3D frame, register the two surfaces, and compute per-region missing area; if the 5-view reconstruction shows consistent gaps under the arch or at heel-strike beyond the reported 2 mm RMSE, the ground truth is incomplete and completion rankings on FootGait3D would need re-evaluation.","tokens_in":14156,"feed_emoji":"🦶","tokens_out":5004,"duration_ms":49687,"temperature":0.7,"pith_summary":"This paper introduces FootGait3D, a real-world dataset of 8,403 point cloud frames of the foot and ankle captured from 46 subjects walking barefoot through a five-camera depth-sensing system. Each frame provides a full five-view reconstruction used as ground truth, plus partial point clouds produced by subsets of four, three, or two cameras, with the bottom plantar camera always removed to mimic ordinary floors. The central claim is that this graded occlusion structure makes FootGait3D the first benchmark where point cloud completion methods can be trained and evaluated on paired partial-and-complete real dynamic anatomy rather than on synthetic CAD models. The paper further reports baseline results showing that leading completion networks, both point-only and multi-modal, lose accuracy on this data relative to synthetic benchmarks, which is the expected cost of realistic noise, non-uniform density, and dynamic deformation. If the dataset is sound, it gives biomechanics and computer vision a shared testbed for occlusion-robust foot shape completion, with consequences for clinical gait analysis, prosthetic design, and robotics.","feed_headline":"FootGait3D pairs partial foot scans with full ground truth","feed_subtitle":"Graded 4/3/2-view inputs against true geometry let completion models face real dynamic feet.","key_machinery":"The organizing mechanism is the view-subset synthesis rule: every partial cloud is a genuine sensor-level subset of the same synchronized capture, so partial and complete clouds share the same global coordinate frame and time instance. The bottom camera (DS4) is always removed in every partial condition, guaranteeing that the plantar surface is missing just as it is on a non-transparent floor; then one, two, or all three of the remaining side views are dropped to create graded occlusion levels. This design converts the dataset into a controlled severity ladder for occlusion: 4-view, four 3-view combinations, and six 2-view combinations, all paired with the same full 5-view ground truth. The data acquisition itself is carried by the Point-cloud Foot Analysis (PFA) system, a five-camera rig whose sensor poses are fixed by coarse calibration with a physical object followed by point-to-plane ICP registration.","core_discovery":"The paper's central claim is that FootGait3D supplies what prior real-world point cloud datasets lack: paired partial and complete surface geometry of a dynamic biological structure, captured during actual gait. The ground truth is the merged five-view reconstruction from sensors arranged around the capture zone plus one beneath a transparent plate; the partial inputs are created by dropping the bottom camera and then subsets of the side cameras, producing 4-view, 3-view, and 2-view conditions that simulate plantar occlusion and swing-foot occlusion. Benchmark experiments with eleven completion methods show that multi-modal methods that fuse real depth maps (notably SVDFormer and PointSea) keep stable performance across all view counts, while single-modal point-only methods vary unpredictably, with some (PointAttn) doing better on sparser inputs. The paper argues this pattern reflects real-world complexity and makes FootGait3D a more discriminating benchmark than synthetic ShapeNet-derived sets.","pith_inferences":["Beyond the paper: the always-missing plantar surface makes FootGait3D a natural probe for whether completion models actually learn anatomical priors for the arch and heel, since those regions can never be observed in the input and must be hallucinated from side geometry.","Beyond the paper: the same capture setup could be repurposed as a self-supervised pretraining source for foot shape completion, using the 5-view reconstructions as pseudo-labels and the view subsets as augmentation, before fine-tuning on sparser clinical capture rigs.","Beyond the paper: one could test whether the reported F1 threshold (1e-3 of the bounding-box diagonal) rewards models that fill in the missing plantar region at all, or merely those that match the visible dorsal surface; a region-masked metric would settle this.","Beyond the paper: a static high-resolution scan of each subject's foot in a matching pose would provide an external check on the ground truth; if systematic missing regions appear under the arch, model rankings on completion quality could change."],"forward_implications":["Completion models trained on synthetic CAD data can be retrained or fine-tuned on real dynamic foot geometry and evaluated against true paired ground truth rather than proxy metrics.","The graded 4/3/2-view conditions make it possible to measure how each method's accuracy degrades with occlusion severity, separating models robust to missing plantar data from those that only need denser input.","Multi-modal methods that consume the real five-view depth maps and camera extrinsics can be assessed for cross-modal fusion on authentic data, not rendered pseudo-views.","Biomechanics studies gain a markerless source of dense surface geometry of the foot-ankle complex across stance phase, supporting multi-segment foot modeling and subject-specific simulation.","The defined train/validation/test split (5881/841/1681 instances) gives a common protocol so future completion results on FootGait3D are directly comparable."],"supporting_citations":[{"why":"Supplies the Point-cloud Foot Analysis (PFA) system, the five-camera hardware and registration foundation on which the dataset is built.","marker":"[Jiang et al., 2023]"},{"why":"Provides the point-to-plane ICP algorithm used for fine sensor registration, which is required for multi-view merging into the ground truth.","marker":"[Chen and Medioni, 1992]"},{"why":"Defines the PCN dataset and completion network, serving as both the synthetic-benchmark reference and one of the benchmarked baselines.","marker":"[Yuan et al., 2018]"},{"why":"Introduces PoinTr and the ShapeNet-derived benchmarks that FootGait3D is contrasted against for real-world generalization.","marker":"[Yu et al., 2021]"},{"why":"Provides Foot3D, the static foot scan dataset whose static nature highlights the dynamic gap FootGait3D fills.","marker":"[Boyne et al., 2022]"},{"why":"KITTI is cited as a real-world dataset that lacks paired completion ground truth, motivating the design of FootGait3D.","marker":"[Geiger et al., 2012]"},{"why":"RealPC is cited for the performance gap of completion methods on real-world industrial data, supporting the need for a real-world benchmark.","marker":"[Pathak et al., 2024]"}],"fun_headline_variants":["FootGait3D pairs partial foot scans with full ground truth for 3D completion","Real gait foot-ankle scans: partial to full, for occlusion-robust completion","8,403 frames, 46 subjects: FootGait3D benchmark for foot 3D completion","Occlusion-robust foot 3D completion: real gait data with partial scans","Benchmarking foot-ankle 3D completion with graded real-gait data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The five-view merged reconstruction, checked only by visual inspection, is treated as complete ground-truth surface geometry for the foot and ankle; if systematic missing regions (for instance under the arch or at heel-strike) persist despite the bottom camera, the completion target itself is biased.","fun_headline_variants_meta":{"raw":{"variants":["FootGait3D pairs partial foot scans with full ground truth for 3D completion","Real gait foot-ankle scans: partial to full, for occlusion-robust completion","8,403 frames, 46 subjects: FootGait3D benchmark for foot 3D completion","Occlusion-robust foot 3D completion: real gait data with partial scans","Benchmarking foot-ankle 3D completion with graded real-gait data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001611,"raw_usage":{"total_tokens":6475,"prompt_tokens":1068,"completion_tokens":5407,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":684,"completion_tokens_details":{"reasoning_tokens":5289}},"tokens_in":684,"tokens_out":5407,"duration_ms":46072,"temperature":1.0,"reasoning_tokens":5289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:18:15.611071+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a subject's foot, obtain a high-resolution static 3D scan in a pose matching a FootGait3D frame, register the two surfaces, and compute per-region missing area; if the 5-view reconstruction shows consistent gaps under the arch or at heel-strike beyond the reported 2 mm RMSE, the ground truth is incomplete and completion rankings on FootGait3D would need re-evaluation.","supporting_citations":[{"cited_title":"Fast tool to evaluate 3d movements of the foot-ankle complex using multi-view depth sensors","cited_arxiv_id":null,"evidence_quote":"Supplies the Point-cloud Foot Analysis (PFA) system, the five-camera hardware and registration foundation on which the dataset is built."},{"cited_title":"Object modelling by registration of multiple range images","cited_arxiv_id":null,"evidence_quote":"Provides the point-to-plane ICP algorithm used for fine sensor registration, which is required for multi-view merging into the ground truth."},{"cited_title":"Pcn: Point completion network","cited_arxiv_id":null,"evidence_quote":"Defines the PCN dataset and completion network, serving as both the synthetic-benchmark reference and one of the benchmarked baselines."},{"cited_title":"FIND: An Unsupervised Implicit 3D Model of Articulated Human Feet","cited_arxiv_id":"2210.12241","evidence_quote":"Provides Foot3D, the static foot scan dataset whose static nature highlights the dynamic gap FootGait3D fills."},{"cited_title":"Are we ready for autonomous driving? the kitti vision benchmark suite","cited_arxiv_id":null,"evidence_quote":"KITTI is cited as a real-world dataset that lacks paired completion ground truth, motivating the design of FootGait3D."},{"cited_title":"Revisiting Point Cloud Completion: Are We Ready For The Real-World?","cited_arxiv_id":"2411.17580","evidence_quote":"RealPC is cited for the performance gap of completion methods on real-world industrial data, supporting the need for a real-world benchmark."}],"review_version":1}