{"id":"f70b3639-afe9-4717-8d69-36d1f92ba1f4","arxiv_id":"2501.01689","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A dual-image CNN trained on OpenPose keypoints is reported to estimate GDI and knee flexion angle from single RGB gait videos with lower MAE than two previous methods, though cadence accuracy is worse.","lead":"The paper proposes a compact dual-image CNN that estimates gait metrics such as GDI and knee flexion angle from single RGB videos using OpenPose keypoints. It reports better MAE than two prior methods for GDI and knee flexion, but worse cadence, with no error bars or code link.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed superiority over prior methods depends on Table 1 baselines being evaluated under the same patient-disjoint split and preprocessing; the paper never reports how the 1D-CNN and STT numbers were obtained, so the 10.6%/12.0% gains may be an artifact of split or pipeline mismatch.","rationale":"I agree with the reader's weakest_assumption: the comparability of baselines is the single most load-bearing condition. The absolute DPG numbers could be true; the dataset description and preprocessing are detailed enough that the DPG pipeline could be reproduced. What cannot be verified from the paper is that the comparison is apples-to-apples. The percentages in the abstract and Section 3.2 are defined relative to Table 1 baseline numbers, and the paper does not state whether those numbers were reproduced with the same subject-disjoint split and preprocessing. This is a correctness question about the central claim, not a stylistic issue. I also note two supporting weaknesses: the architecture is described as a 'Transformer' but Section 2.2 contains no self-attention or transformer block (only Conv + FC layers), which undermines the stated novelty; and the paper claims released code and trained models but provides no repository link, which hampers the proposed verification. Neither changes my verdict beyond the reader's CONDITIONAL, because the empirical result may still hold and the public dataset would allow a controlled rerun. Thus I keep the verdict UNCHANGED: the paper should not be accepted as-is, but the concern is testable and does not warrant outright rejection.","tokens_in":6380,"tokens_out":4942,"duration_ms":50238,"concrete_test":"Run the released implementations of 1D-CNN [11] and STT [17] on DPG's exact test split and preprocessing (same 232 test videos, same OpenPose keypoints, same two-image encoding—or their native inputs if the comparison is intended to be end-to-end) and recompute MAE. If the values differ from 6.3137/6.5469 (GDI) and 5.8220/5.9129 (knee flexion), the reported 10.6%/12.0% improvements are not established. A minimal first step is to obtain DPG's test-video identifier list and the baseline authors' test predictions to check whether the same subjects and clips are compared.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that DPG lowers MAE by 10.6% on GDI and 12.0% on knee flexion relative to state-of-the-art methods (Section 3.2). This claim is only meaningful if the 1D-CNN [11] and STT [17] entries in Table 1 were evaluated on the same test set, with the same patient-disjoint split, clip windowing (124 frames, 31-frame overlap), OpenPose keypoint representation, and preprocessing, and with the same model-selection rule. Section 3.1 reports the DPG pipeline and split counts (1,768/920 training, 212/106 validation, 232/112 test) but says nothing about how the baseline MAE values were produced—whether they were rerun under this protocol or copied from the original papers. The original Kidzinski et al. dataset contains 2,212 recordings of 1,138 patients; whether that split and the split used in [17] coincide with the one used here is not established. If the baseline numbers come from different splits or input representations, the percentage reductions are not estimates of the same quantity and the headline claim collapses. This is the load-bearing assumption because even fully reproducible DPG MAEs do not imply superiority without controlled baselines. A secondary issue is that Section 2.2 describes only convolutional and fully connected layers, with no self-attention or transformer block, despite the method being called a 'dual-input Transformer'; this weakens the stated novelty but does not by itself invalidate the empirical comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DPG, a dual-input convolutional network for estimating gait deviation index (GDI), knee flexion angle, and cadence from single-view RGB video. OpenPose keypoints from 124-frame clips are encoded as two grayscale images: one stacks normalized landmark coordinate rows over time and the other plots hip, knee, and ankle landmarks from all frames. A three-layer CNN processes each image, the resulting features are concatenated and regressed through four fully connected layers. Training and evaluation use the Gillette Children's dataset with an 8:1:1 patient-disjoint split. Table 1 reports MAE values of 5.6450, 5.1203, and 0.1418 for GDI, knee flexion, and cadence, respectively, and the text claims 10.6% and 12.0% improvements over 1D-CNN and STT for the first two metrics while cadence is worse. The authors state that code and trained models are released, but no repository link is given.","tokens_in":6734,"tokens_out":4268,"duration_ms":42280,"significance":"If the reported comparison were controlled, the result would be a modest but practically useful demonstration that a simple image-encoding network can estimate clinical gait metrics from RGB video with few parameters. The main strengths are the use of a real clinical dataset with patient-disjoint splitting and a clearly described preprocessing pipeline. However, the paper's central novelty claim, a 'dual-input Transformer' with self-attention, is not reflected in the architecture, which is described as a CNN with no attention mechanism. The significance therefore hinges on whether the empirical superiority claim survives a controlled baseline comparison; in its current form the evidence is not sufficient.","major_comments":[{"comment":"The headline claim of 10.6% reduction on GDI and 12.0% on knee flexion is not supported unless the 1D-CNN and STT baseline MAEs were obtained under exactly the same protocol. The manuscript reports the DPG split (8:1:1, patient-disjoint; 1,768/920 training, 212/106 validation, 232/112 test) and preprocessing, but never states whether the baseline numbers in Table 1 were reproduced with that split, the same 124-frame windowing with 31-frame overlap, the same OpenPose keypoint representation, and the same model-selection rule, or whether they were copied from [11] and [17]. Because the original dataset splits and preprocessing in those papers may differ, the percentage improvements may be artifacts of evaluation mismatch. Please rerun baselines under the identical pipeline or state precisely the source of each baseline number and justify comparability, and report variability over repeated runs.","section":"Section 3.2 / Table 1"},{"comment":"The architecture described in Section 2.2 and Figure 2 consists of three convolutional layers followed by four fully connected layers. There is no self-attention, multi-head attention, positional encoding, or transformer block anywhere in the model description. The Abstract, Introduction, and title repeatedly call the model a 'dual-input Transformer', and the Introduction states that 'the self-attention mechanism enables the model to focus on relevant anatomical keypoints', which is contradicted by the presented method. Either the model must be changed to include an attention component, or the claims and name should be revised to describe a dual-input CNN; as written, the central modeling contribution is not accurately characterized.","section":"Section 2.2 / Figure 2"},{"comment":"The dual-input representation encodes the 124-frame sequence as two static images: one row-per-frame normalized keypoint image and one scatter of hip-knee-ankle coordinates over all frames. This choice discards the explicit temporal ordering in the second input and only implicitly preserves it in the first. No ablation compares this encoding to a sequence-based input or to using a single image, and no analysis shows that the two inputs contribute complementary information. Given that the claimed advantage over 1D-CNN is attributed to temporal modeling, the absence of such an ablation leaves the core design choice unvalidated.","section":"Section 3.1 / preprocessing"},{"comment":"All results are single-point MAE values with no standard deviation, number of repeated runs, or statistical significance testing. Differences between methods, for example 6.3137 versus 5.6450 for GDI and 5.9129 versus 5.1203 for knee flexion, may be meaningful but cannot be assessed without variability estimates or paired tests across patients. The paper also does not specify the units or clinical interpretation of the cadence error, which makes the reported failure on cadence difficult to evaluate.","section":"Section 3.2 / Table 1"}],"minor_comments":[{"comment":"There are typographical errors in the title page and author list, including 'Quantita tive' and 'Vuilerrme'; the phrase 'convolution Transformer' should likely be 'convolutional Transformer'.","section":"Title page / Header"},{"comment":"The Abstract and contribution list mention step length as an estimated gait parameter, but Table 1 and Section 3.2 report results only for GDI, knee flexion, and cadence; the step length result is missing.","section":"Abstract / Section 3.2"},{"comment":"Figure 1 is reused from Kidzinski et al. [11] for illustration purposes, but the manuscript does not include an explicit copyright or permission statement for the reused figure.","section":"Figure 1 caption"},{"comment":"The phrase 'Through thought experiments and comparisons' is unclear; it should be replaced with a more precise description such as 'Through qualitative analysis and experimental comparisons'.","section":"Section 3.2"},{"comment":"Reference [17] is a prior paper by Le and Pham, which appears to be from the same group as the current manuscript; if so, this relationship should be disclosed when STT is used as a baseline.","section":"References / Section 3.2"},{"comment":"The statement that 'we released our codes and trained models' is not accompanied by a repository link or availability statement, so the claim cannot be verified by readers.","section":"Section 1 / Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper reads as an extended abstract, and the key risk is the uncontrolled baseline comparison and the mismatch between the 'Transformer' label and the actual CNN architecture. The comparison to [17], which appears to be prior work by the same group, should be checked for disclosure. If the authors can rerun the baselines under the identical split and preprocessing and either add attention or rename the method, the empirical claim could become meaningful, but in the current form the central claims are not sufficiently supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper in two sentences: it takes 124 frames of OpenPose keypoints, draws them as two grayscale images, and runs a small CNN to regress GDI, knee flexion, step length, and cadence. The reported gains on GDI and knee flexion are plausible, but I do not trust the baseline numbers as reported, so the central claim is not yet established.\n\nWhat is genuinely new: the dual-image encoding is a cheap way to turn a sequence into a static input, and the model is small enough to run on modest hardware. That is a real plus for the stated clinical scenario. The preprocessing is described at a level that allows re-implementation on the public Kidzinski dataset, and the authors say they released code. If the code is actually available, that is a concrete contribution.\n\nSoft spots, in order of severity. First, Table 1. The paper never says whether the 1D-CNN and STT numbers were produced under the same patient-disjoint split, clip windowing (124 frames, 31 overlap), and OpenPose keypoint representation, or copied from earlier papers. The original Kidzinski dataset has 2,212 recordings of 1,138 patients; this paper uses an 8:1:1 patient split, but the STT baseline is the authors' own ISBI 2024 paper without any disclosure that it is self-comparison. Without controlled baseline reproduction, the 10.6%/12.0% reductions cannot be read as superiority. This is the load-bearing gap, and the reader's stress-test note lands.\n\nSecond, the name. The paper calls the model a Transformer, but Section 2.2 describes only convolution, max-pooling, and fully connected layers. There is no self-attention. That is not just marketing; it obscures what the method actually is and inflates the novelty. The contribution is a dual-input CNN, and it should be described as such.\n\nThird, the evaluation is thin: no error bars, no repeated runs, no statistical tests. The differences might be noise. Fourth, there is no ablation to justify the dual input. Without testing the model on each image alone, we do not know whether the second pattern (hip-knee-ankle) adds anything. The cadence result is worse, and the explanation that the model cannot detect gait cycles is reasonable but should be part of the design, not an excuse.\n\nWho should read this: people working on low-cost markerless gait analysis will find the encoding idea useful, but they should not change their methods based on these numbers until the baseline issue is resolved. This deserves peer review in the sense that the question is legitimate and the paper is re-implementable; a good referee would require a clean comparison before publication.\n\nMy recommendation: send it to review, but flag the baseline comparison as the make-or-break issue. If the authors can reproduce the baselines under the same protocol and add an ablation, the paper would be a respectable incremental contribution. As it stands, the headline claim is unverified.","headline":"A small dual-image CNN that may improve GDI/knee-flexion MAE on the Kidzinski dataset, but the baseline comparison is uncontrolled enough that the headline gains aren't established.","tokens_in":7242,"tokens_out":3738,"would_cite":false,"duration_ms":34604,"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":"A dual-image Transformer can estimate gait parameters from ordinary video, reducing GDI error by 10.6%.","keywords":["gait analysis","Gait Deviation Index (GDI)","knee flexion","cadence","single RGB video","pose estimation","dual-input Transformer","cerebral palsy"],"falsifier":"A reader could re-train the 1D-CNN and STT baselines on the same 8:1:1 patient-exclusive split and dual-image preprocessing, and also shuffle frame order within the 124-frame clips before building the input images. If the reproduced baseline MAEs do not match the values of 6.5469 and 6.3137 (GDI) or 5.9129 and 5.8220 (knee flexion), the reported reductions do not hold; if shuffled inputs change predictions only slightly, the model is not actually using temporal gait dynamics.","tokens_in":6210,"feed_emoji":"🚶","tokens_out":6577,"duration_ms":56592,"temperature":0.7,"pith_summary":"Quantitative gait analysis normally requires a motion-capture laboratory, which is expensive and hard to staff. This paper claims that a compact dual-input convolutional Transformer network can estimate clinically useful gait parameters from an ordinary single-camera RGB video, by turning a 124-frame skeleton sequence into two grayscale images. On a public dataset of 2,212 videos from 1,138 children with cerebral palsy, the model reports mean absolute errors of 5.6450 for the Gait Deviation Index, 5.1203 degrees for knee flexion, and 0.1418 for cadence. Against the two published comparison models, this is a 10.6% lower GDI error and a 12.0% lower knee-flexion error, while cadence remains worse. If these numbers hold under identical evaluation conditions, the method offers a smaller, cheaper route to markerless gait assessment.","feed_headline":"Dual-image Transformer cuts gait-measurement error by 12%","feed_subtitle":"It reads a 124-frame skeleton clip as two pictures and beats published GDI and knee-flexion errors.","key_machinery":"The load-bearing object is the dual-pattern input: two grayscale images constructed from the same 124-frame skeleton sequence. The first image is a 128x64 matrix whose rows are the normalized (x,y) coordinates of 25 body landmarks, padded and duplicated to 128x128; the second is a 128x128 plot of the hip, knee, and ankle trajectories across all frames. Each image is processed by three convolutional blocks with max pooling, the resulting 32,768-dimensional feature maps are concatenated to a 65,536-unit vector, and four fully connected layers regress to a single scalar. In the paper's framing, the self-attention mechanism lets the model focus on relevant anatomical keypoints and their relationships over time, replacing the hand-crafted features and separate pose-estimation stages used by earlier pipelines.","core_discovery":"The paper's central finding is that encoding the temporal sequence of 2D body keypoints as two static 128x128 grayscale images is enough for a single CNN-based Transformer model to regress gait parameters from single-view RGB video. The first image arranges all 25 body landmarks across the 124 frames as rows, and the second plots only hip, knee, and ankle coordinates over the whole clip; the same network consumes both, concatenates their features, and outputs one gait metric. Measured against ground-truth laboratory values, the DPG model achieves mean absolute errors of 5.6450 (GDI), 5.1203 (knee flexion), and 0.1418 (cadence) on the test set. The authors interpret the lower GDI and knee-flexion errors as evidence that a single dual-pattern model can surpass the 1D-CNN and STT approaches for those two metrics while using fewer resources, and they attribute the cadence shortfall to the absence of explicit gait-cycle and stride detection.","pith_inferences":["The paper never checks whether discarding frame order matters: an ablation that shuffles the 124 frames before building the two images would show how much of the accuracy comes from temporal gait dynamics rather than from a static posture summary.","If the baseline MAE values were copied from the original publications rather than recomputed on the same patient-exclusive split and preprocessing, the 10.6% and 12.0% reductions could be an artifact of different evaluation settings; reproducing the baselines is the decisive test.","The same two-image encoding might generalize to other clinical video tasks, but robustness to occlusions, camera angles, and atypical gait patterns remains untested, so real-clinic performance could differ from lab-test error."],"forward_implications":["Clinics with only a standard camera could obtain GDI and knee-flexion estimates within roughly 5-6 units or degrees of laboratory values, bypassing marker-based motion capture.","A single network replaces the multi-network pipeline of the earlier approach, reducing training and inference cost, which matters for telehealth and low-resource settings.","Cadence prediction at 0.1418 MAE is weaker than both baselines, so the method as presented is not yet a complete gait-assessment tool; adding stride and gait-cycle detection is the natural next step.","The dual-image encoding could be adapted to other clinical parameters derived from body landmarks, such as step length or trunk kinematics, with the same small input representation."],"supporting_citations":[{"why":"Extracts the 2D body keypoints from each RGB frame that form the model's input representation.","marker":"[9]"},{"why":"Supplies the Transformer and self-attention concept the network is named after and that the authors say captures spatial-temporal dependencies.","marker":"[10]"},{"why":"Provides the public clinical gait dataset, the 124-frame clip preprocessing, and the 1D-CNN baseline whose MAE is compared against DPG.","marker":"[11]"},{"why":"Gives the STT transformer baseline compared against, the other method whose GDI and knee-flexion errors DPG claims to reduce.","marker":"[17]"}],"fun_headline_variants":["Dual-image Transformer matches lab-grade gait metrics","Skeleton clip as two images: better GDI, knee flexion","Single video gait analysis: dual-input Transformer wins","One RGB video, two images, accurate gait parameters","Dual-pattern Transformer slashes gait error for GDI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed accuracy gains assume the two comparison models were evaluated on exactly the same train/test split and preprocessing; if their published numbers came from different settings, the reported improvements are not meaningful.","fun_headline_variants_meta":{"raw":{"variants":["Dual-image Transformer matches lab-grade gait metrics","Skeleton clip as two images: better GDI, knee flexion","Single video gait analysis: dual-input Transformer wins","One RGB video, two images, accurate gait parameters","Dual-pattern Transformer slashes gait error for GDI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000258,"raw_usage":{"total_tokens":1581,"prompt_tokens":941,"completion_tokens":640,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":561}},"tokens_in":557,"tokens_out":640,"duration_ms":6425,"temperature":1.0,"reasoning_tokens":561,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:22:12.125344+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could re-train the 1D-CNN and STT baselines on the same 8:1:1 patient-exclusive split and dual-image preprocessing, and also shuffle frame order within the 124-frame clips before building the input images. If the reproduced baseline MAEs do not match the values of 6.5469 and 6.3137 (GDI) or 5.9129 and 5.8220 (knee flexion), the reported reductions do not hold; if shuffled inputs change predictions only slightly, the model is not actually using temporal gait dynamics.","supporting_citations":[{"cited_title":"Tools for observational gait analysis in pa- tients with stroke: A systematic review,","cited_arxiv_id":null,"evidence_quote":"Extracts the 2D body keypoints from each RGB frame that form the model's input representation."},{"cited_title":"Acceleration gait measures as proxies for motor skill of walking: A narrative review,","cited_arxiv_id":null,"evidence_quote":"Supplies the Transformer and self-attention concept the network is named after and that the authors say captures spatial-temporal dependencies."},{"cited_title":"Biometric recognition by gait: A survey of modalities and features,","cited_arxiv_id":null,"evidence_quote":"Provides the public clinical gait dataset, the 124-frame clip preprocessing, and the 1D-CNN baseline whose MAE is compared against DPG."},{"cited_title":"Deep neural networks enable quantitative movement analysis using single- camera videos,","cited_arxiv_id":null,"evidence_quote":"Gives the STT transformer baseline compared against, the other method whose GDI and knee-flexion errors DPG claims to reduce."}],"review_version":1}