{"id":"50f9908c-35d3-4305-848a-90066a9c224f","arxiv_id":"2411.11845","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A pipeline that reconstructs personalized MANO/NIMBLE hand meshes from arbitrary keypoints and derives a unified 25-joint set via a trained MLP.","lead":"UniHands turns hand keypoints from many different tracking systems into one standard, personalized 3D hand model. It combines two popular hand models and a small user study suggests the resulting joints look and move more naturally.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Precision claim is circular: Tab.1 fits MANO to MANO-derived keypoints, so sub-0.1mm errors show self-fit, not wild-keypoint generalization; the nine-person study does not provide mesh-level accuracy.","rationale":"I fully agree with the reader's diagnosis: the evaluation is self-referential because the test keypoints are MANO-derived, so the reported sub-0.1 mm means cannot support the paper's headline claim of reconstructing from wild-collected keypoints. The reader's weakest_assumption—that FreiHAND/InterHand keypoints are representative wild-collected data—is the same load-bearing concern. My stress test adds the observation that the negative mean errors in Tab. 1 are physically impossible if PJ/PV are Euclidean distances, further undermining the quantitative report. The user study, while a genuine pilot with nine participants, only measures subjective preference among joint sets, not reconstruction error on unseen keypoint sources. No code or baselines are provided, and the method's only objective numbers are from the circular setting. Therefore the correct verdict remains REJECT, and my read does not alter the reader's verdict. I would still credit the unified MANO-NIMBLE joint idea as a potentially useful contribution, but the paper's central accuracy claim is not currently supported by its evidence.","tokens_in":3665,"tokens_out":5350,"duration_ms":49960,"concrete_test":"Re-run the UniHands pipeline on raw keypoint tracks from non-MANO sources (e.g., the Ultraleap, HoloLens2, and MediaPipe sequences used in the Sec. 3.2 pilot), and compare the reconstructed meshes against ground-truth 3D hand scans of the same gestures captured with a high-resolution scanner. Report unsigned PJ/PV and 95th-percentile errors, and compare against a MANO-only baseline. If the median error is not below ~1 mm and better than the baseline, the 'precise reconstruction from wild-collected keypoints' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim (PJ/PV < 0.1 mm, Tab. 1) is computed by optimizing MANO pose/shape parameters against keypoints taken from FreiHAND and InterHand2.6M. Both datasets provide MANO annotations, so the 'wild-collected keypoints' are generated by the same parametric model family (MANO) that the optimizer is using. Recovering parameters from their own joint outputs makes near-zero mean error a fitting artifact rather than evidence of generalization to arbitrary keypoint sources. The paper itself qualifies Sec. 3.1 as 'when keypoints derived from wild-collected 3D hand models are aligned with ours,' which is exactly the circular condition. The large standard deviations (PJ-std 1.03--1.29 mm, PV-std 1.77--2.46 mm) and the negative mean errors for InterHand (PJ = -0.0047 mm, PV = -0.012 mm) are unexplained if errors are Euclidean distances, suggesting sign cancellation or a different metric. The nine-participant study (Sec. 3.2) is a real exploratory user preference test, but it rates rendered outputs subjectively and does not measure mesh-to-keypoint alignment error, so it does not rescue the quantitative accuracy claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniHands, a pipeline that reconstructs personalized parametric hand models from keypoints by coarse-to-fine optimization of MANO and NIMBLE parameters, and derives a unified joint set from the reconstructed mesh. The abstract claims sub-0.1mm alignment errors on FreiHAND and InterHand2.6M and reports a nine-participant user study that significantly prefers the unified joints for naturalism (p=0.016). The central quantitative claim is not supported by the presented evidence: the benchmark keypoints are MANO-derived, the reported errors include negative values, and no baseline comparisons are provided.","tokens_in":4019,"tokens_out":8750,"duration_ms":85355,"significance":"If validated, a method that maps heterogeneous 'wild-collected' keypoints to parametric MANO/NIMBLE meshes would be practically useful for hand tracking, animation, and XR applications. The proposed MANO-NIMBLE joint fusion via an MLP is an interesting design choice, and the user study is a genuine attempt to assess subjective quality. However, the paper's main validation is circular relative to the MANO model family, the error metric appears miscalculated, and the absence of baselines leaves the central reconstruction claim unsupported. The current manuscript therefore does not establish its headline contribution.","major_comments":[{"comment":"The evaluation is circular. The keypoints in FreiHAND and InterHand2.6M are derived from or aligned with the MANO model family (FreiHAND provides MANO parameters; InterHand2.6M provides MANO annotations), and Eq. (1) optimizes MANO pose and shape to minimize the distance to those same keypoints. Near-zero mean errors therefore measure self-consistency, not generalization to wild-collected keypoints. The paper's own qualification in Sec. 3.1 ('when keypoints derived from wild-collected 3D hand models are aligned with ours') acknowledges this limitation. Please validate on keypoints from non-MANO sources (e.g., MediaPipe detections, marker-based captures, or other hand model families) and report errors against independent ground-truth meshes.","section":"Section 3.1, Table 1, Eq. (1)"},{"comment":"The reported mean errors for InterHand2.6M are negative (PJ = -0.0047 mm, PV = -0.012 mm). For Euclidean distances, mean errors are non-negative by definition. This indicates that the metric is either signed (e.g., per-coordinate signed differences averaged over axes) or miscalculated. The statement 'PJ/PV < 0.1mm' is therefore not supported without a precise and valid error definition. Please specify the exact metric and recompute all values.","section":"Table 1"},{"comment":"No baselines are reported. To support the claim of 'precise reconstruction,' the paper should compare against existing keypoint-to-mesh fitting methods, such as direct MANO fitting or recent learning-based hand mesh reconstruction approaches, on standard benchmarks with accepted error metrics (e.g., PA-MPJPE, MPVPE). Without such comparisons, the numbers in Table 1 provide no evidence that UniHands outperforms a straightforward optimization baseline.","section":"Section 3.1"},{"comment":"The nine-participant user study measures subjective ratings of rendered joint sets, not mesh-to-keypoint alignment or reconstruction accuracy. It therefore cannot substitute for a quantitative evaluation of the central precision claim. The reported significant result (p=0.016) applies only to the naturalism task, and the details of the statistical test, multiple-comparison handling, and participant blinding are not provided. This section should be framed as a pilot preference study, not as validation of reconstruction precision.","section":"Section 3.2"}],"minor_comments":[{"comment":"The regularization weights λ_reg and λ_smooth are never specified, and the optimization protocol (learning rate, number of iterations, initialization, early stopping) is not described. These details are necessary for reproducibility.","section":"Section 2.1, Eq. (3)"},{"comment":"The MLP training procedure is underspecified: the paper does not state what data are used to generate training pairs, how the NIMBLE joints are obtained for supervision, the loss function, or the MLP's prediction error. Please provide these details.","section":"Section 2.2"},{"comment":"Reference [4] is incorrect: the cited Zimmermann and Brox paper is 'Learning to Estimate 3D Hand Pose from Single RGB Images' (ICCV 2017), which is not the FreiHAND dataset. The FreiHAND dataset is introduced in Zimmermann et al., 'FreiHAND: A Dataset for Markerless Capture of Hand Pose and Shape from Single RGB Images' (ICCV 2019).","section":"References"},{"comment":"The abstract states that experiments 'demonstrate its ability to precisely reconstruct hand mesh vertices and keypoints,' but the body (Sec. 3.1) restricts this to data 'aligned with our optimization models.' The abstract should reflect this limitation to avoid overclaiming.","section":"Abstract"},{"comment":"The box plots and the star markers in Fig. 3 are difficult to interpret. Please specify which statistical test was used, whether a correction for multiple comparisons was applied, and what the stars denote (e.g., p<0.05? p<0.016?).","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central quantitative evaluation is circular and the metric appears mis-specified, so the headline reconstruction claim is currently unsupported. The user study is exploratory and does not compensate for the missing quantitative validation. I could see this becoming publishable if the authors redo the experimental evaluation with independent keypoint sources, correct the error metric, and add baselines; otherwise the paper should be rejected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing you should know: UniHands has a genuinely useful idea — a unified 25-joint set that fuses MANO and NIMBLE skeleton definitions, with an MLP that derives NIMBLE-like joints directly from the MANO mesh. That part is new relative to the cited literature and could be practically handy for XR and animation pipelines that need to combine keypoint data from different trackers. The MLP-based joint derivation is a clean, reproducible contribution, and the paper is careful to position it as a mesh-to-joint mapping rather than a full reconstruction method.\n\nWhat the paper does well: the coarse-to-fine optimization is standard MANO fitting, but the authors are transparent about the one-time mesh alignment between MANO and NIMBLE, and the fusion joint definition is concrete enough to reimplement. The pilot user study, while small, is a real attempt to test on data that does not match the optimization model, and the significant preference for the fused joints on naturalism (p=0.016) is a suggestive signal, not a fabricated one.\n\nWhere it falls down: the central quantitative claim — PJ/PV < 0.1mm on FreiHAND and InterHand2.6M — does not hold up. Those datasets provide MANO annotations, so the optimizer is fitting MANO to keypoints that were generated by the same model family. That makes the near-zero mean error a self-fit artifact, not evidence of generalization to wild-collected keypoints. The paper itself acknowledges this in Sec. 3.1 with the phrase \"when keypoints derived from wild-collected 3D hand models are aligned with ours,\" but the abstract overstates it. The numbers also look internally off: mean errors of -0.0047mm and -0.012mm for InterHand are impossible if PJ/PV are Euclidean distances, and the standard deviations (1.03–2.46mm) are an order of magnitude larger than the means, suggesting signed error or a different metric that is never defined. No baselines and no code are provided, so independent verification is not possible.\n\nThe user study is the only non-circular evidence, and it is a nine-person pilot with no multiple-comparison correction — acceptable as a pilot, but not enough to rescue the accuracy claim.\n\nBottom line: the fusion-joint idea deserves a serious referee, but not in its current form. If I were the editor, I would send it to peer review because the core idea is useful and the flaws are fixable — with an evaluation on truly held-out keypoint sources, per-sample error distributions, and a clearly defined metric. As written, the paper's main quantitative evidence is circular, and the current draft should not be accepted.","headline":"Useful joint-fusion idea, but the main accuracy claim is a self-fit artifact and the paper needs a reworked evaluation before it can be trusted.","tokens_in":4447,"tokens_out":2325,"would_cite":false,"duration_ms":23162,"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":"UniHands rebuilds full 3D hands from sparse wild keypoints","keywords":["3D hand reconstruction","hand keypoints","parametric hand model","MANO","NIMBLE","motion capture","joint fusion"],"falsifier":"Take keypoints produced by a detector trained on a different skeleton convention (e.g., MediaPipe or HoloLens-style joints) for hands with known ground-truth 3D scans, run UniHands, and measure vertex-to-scan distance. If the error is far above the reported sub-0.1mm level, the claimed wild-collection generalization fails; if it stays low, the claim survives.","tokens_in":3431,"feed_emoji":"🖐️","tokens_out":3280,"duration_ms":27918,"temperature":0.7,"pith_summary":"UniHands is a method for turning cheap, easy-to-collect 3D keypoint annotations—hand joint positions taken from mixed, 'wild' sources—into personalized, parameterized hand meshes. It does so by fitting the MANO and NIMBLE parametric hand models with a coarse-to-fine optimization, then deriving a unified joint set from the mesh so the result plugs into animation and XR pipelines. On the FreiHAND and InterHand2.6M benchmarks it reports sub-0.1mm mean joint/vertex alignment when the input keypoints match the models' joint definitions. In a nine-participant study, the unified MANO-NIMBLE joints were significantly preferred for naturalness of motion (p=0.016). The value is that data collection no longer needs expensive mesh capture; keypoints suffice.","feed_headline":"UniHands rebuilds full 3D hands from sparse wild keypoints","feed_subtitle":"Fitting MANO and NIMBLE to keypoints gives sub-0.1mm meshes and a unified joint set users prefer.","key_machinery":"The load-bearing mechanism is a two-stage optimization over MANO's pose (θ), shape (β), and global wrist rotation (Rw): a coarse stage starts from mean pose and shape to fit wrist rotation, and a fine stage alternately refines pose/shape and wrist rotation with Adam, using a keypoint-to-vertex distance term plus regularization. A second component is an MLP that maps the 778-vertex MANO mesh to NIMBLE-style joints, letting the pipeline fuse MANO and NIMBLE skeletons into a single 25-joint set. This joint-from-mesh derivation is what gives the output a standard format usable across tasks.","core_discovery":"The central claim is that a parametric-model fitting pipeline can convert wild-collected keypoints from diverse sources into precise, personalized 3D hand meshes, and that the same mesh can serve as a source of standardized hand joints. The authors show that minimizing keypoint-to-mesh-joint distance in a coarse-to-fine scheme, with MANO and NIMBLE aligned to each other, yields PJ/PV errors below 0.1mm on FreiHAND and InterHand2.6M. They also introduce a fused joint set—10 MANO joints, 10 NIMBLE joints, and 5 fingertip joints (25 total)—derived from the mesh by a trained MLP, which users rated as more natural than either model's native joints. The paper's own caveat is that the near-zero error occurs when the keypoint format is consistent with the optimization models.","pith_inferences":["The near-zero benchmark errors likely reflect self-consistency with the parametric model rather than true wild generalization; a stronger test would use keypoint detectors trained on different skeletons and compare against independent 3D scans.","The MLP that predicts NIMBLE joints from the MANO mesh could be trained entirely from synthetic mesh pairs, making the fused joint set extensible to other hand models.","If adopted, the fused joint set could serve as a normalization target for multi-source hand datasets, enabling cross-dataset training.","The small user study (nine participants) hints at a perceptual effect; a larger study with animation playback would be needed to confirm the naturalism advantage."],"forward_implications":["Hand tracking systems that output only keypoints can be upgraded to full personalized meshes without new hardware.","The unified 25-joint set gives animation and AR/VR pipelines a single skeleton format, removing the need to match MANO versus NIMBLE conventions.","Synthetic data generation and hand-pose training can use keypoint-only wild annotations as supervision for mesh-level outputs.","The method suggests that costly mesh annotations may be replaceable by keypoint annotations for many hand tasks."],"supporting_citations":[{"why":"Introduces the NIMBLE non-rigid hand model with bones and muscles, used as the second parametric model and source of anatomical joints.","marker":"[1]"},{"why":"Provides the InterHand2.6M dataset and baseline for 3D interacting hand pose estimation, used as an evaluation benchmark.","marker":"[2]"},{"why":"Supplies the MANO parametric hand model, the primary optimization target and mesh source in the reconstruction pipeline.","marker":"[3]"},{"why":"Provides the FreiHAND dataset and evaluation protocol used to measure mesh reconstruction accuracy.","marker":"[4]"}],"fun_headline_variants":["UniHands converts wild keypoints into precise 3D hand meshes","UniHands standardizes wild keypoints into personalized hand models","UniHands: from sparse keypoints to unified hand joints and meshes","UniHands fits MANO and NIMBLE to create precise 3D hands"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that the keypoint sets in FreiHAND and InterHand2.6M represent 'wild-collected' keypoints and match the MANO/NIMBLE joint definitions, so that sub-0.1mm errors demonstrate generalization rather than fitting the same parametric model that generated the data.","fun_headline_variants_meta":{"raw":{"variants":["UniHands converts wild keypoints into precise 3D hand meshes","UniHands standardizes wild keypoints into personalized hand models","UniHands: from sparse keypoints to unified hand joints and meshes","UniHands fits MANO and NIMBLE to create precise 3D hands"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000833,"raw_usage":{"total_tokens":3615,"prompt_tokens":903,"completion_tokens":2712,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":2626}},"tokens_in":519,"tokens_out":2712,"duration_ms":19039,"temperature":1.0,"reasoning_tokens":2626,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:04:10.105255+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take keypoints produced by a detector trained on a different skeleton convention (e.g., MediaPipe or HoloLens-style joints) for hands with known ground-truth 3D scans, run UniHands, and measure vertex-to-scan distance. If the error is far above the reported sub-0.1mm level, the claimed wild-collection generalization fails; if it stays low, the claim survives.","supporting_citations":[],"review_version":1}