{"id":"85f58698-1960-4c83-bbdc-a1233501d055","arxiv_id":"2412.18321","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A CNN-LSTM gesture recognizer with a decorative 3D skeleton visualization that reports unverifiable accuracy and speed numbers.","lead":"This paper describes a gesture recognition system built from standard CNN and LSTM components, with a 3D hand skeleton diagram presented as the main novelty. The claimed 95% accuracy and fast response times are presented without any reproducible evidence, and the cited dataset does not contain gesture images.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The experimental result rests on a misidentified dataset: GazeCapture contains eye-tracking data, not gesture images, so the 95% accuracy claim is unsupported.","rationale":"The reader's weakest_assumption identifies exactly the load-bearing flaw: the paper's experimental results are anchored to a dataset that does not contain gesture data. GazeCapture is a well-documented real dataset for gaze estimation, and its public release includes face/eye videos and gaze labels, not hand gestures. Since Section III-A explicitly claims the opposite and no other dataset is named, the training process, the reported accuracy, and the timing table are all ungrounded. This is not merely a question of external consensus or contested baselines; it is an internally testable factual mismatch between the described data source and the actual dataset. The paper also lacks a reproducible protocol, but the dataset error is sufficient by itself to reject the central claim. I therefore concur with the reader's REJECT verdict, and no change to that verdict is needed.","tokens_in":7206,"tokens_out":1416,"duration_ms":14724,"concrete_test":"Verify the contents of the GazeCapture dataset from its official release and documentation (Krafka et al., 2016): list all recorded modalities and label fields. If the dataset contains only eye images, face images, and gaze target coordinates — with no hand images or gesture category labels — then the Section III-A premise is false and the 95% accuracy result is unsupported. As a secondary check, attempt to locate the training/validation split and any gesture classifier labels in the dataset files; their absence would conclusively invalidate the reported experiment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — a CNN-LSTM gesture recognizer with 3D hand-skeleton modeling achieves 95% accuracy and 0.12–0.25 s response times — is supported only by the experiment described in Section III. That experiment's data source is stated in Section III-A to be the GazeCapture dataset, described as containing 'a large number of gesture images and corresponding user gaze data.' This description is factually incorrect: GazeCapture is a Stanford eye-tracking dataset collected for gaze estimation, containing images of faces/eyes and gaze target coordinates, not hand-gesture images or gesture labels. Because the training set, the multimodal gesture-plus-gaze fusion, and the 95%/90%/85% comparative accuracies all depend on this dataset, the premise failure removes the experimental grounding for the paper's headline claim. No alternative dataset, training protocol, code, or raw results are supplied, so the reported numbers cannot be independently checked or attributed to an actual experiment. The 3D skeleton model is described visually but is never shown as an input to or component of the CNN-LSTM pipeline, so even the proposed method's mechanism is not demonstrated. Thus the central claim is a claim-without-evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a computer-vision gesture recognition system based on a three-dimensional hand skeleton model, integrating a CNN-LSTM architecture and optional eye-tracking data for multimodal fusion. The abstract and Section III-B claim that the method achieves 95% recognition accuracy, outperforming CNN-based methods (90%) and manual-feature approaches (85%), with system response times between 0.12 and 0.25 seconds across VR, AR, smart home, gaming, and complex-background scenarios. The experimental section reports these results in Table 1, but provides no training protocol, sample sizes, error bars, hyperparameters, or code. The paper also describes a 3D hand-skeleton visualization and discusses future directions for HCI.","tokens_in":7405,"tokens_out":2022,"duration_ms":20055,"significance":"If the claimed results were reproducible, they would represent a modest incremental contribution to gesture recognition, combining well-known components (CNN, LSTM, dropout, data augmentation) with a 3D skeleton representation. However, the manuscript provides no empirical evidence that the proposed method works as described: the dataset on which all experiments rest is misidentified, the experimental sections contain no methodological detail, and the 3D skeleton model is never integrated into the recognition pipeline in a testable way. The paper ships no code, no dataset, no formal derivation, and no reproducible procedure, so its central quantitative claims are unverifiable. The strengths are limited to a clearly stated architecture idea and a relevant application motivation, but these are not sufficient to support the reported accuracy and latency figures.","major_comments":[{"comment":"The dataset description is factually incorrect and load-bearing: the GazeCapture dataset is a well-known eye-tracking dataset collected at Stanford University, containing images of faces and gaze target coordinates; it does not contain gesture images or gesture labels. The paper states that GazeCapture 'contains a large number of gesture images and corresponding user gaze data' and that it 'consists of gesture and gaze data collected by more than 5,000 users.' Because the training set, the multimodal fusion, and the reported 95% accuracy all depend on this dataset, the premise failure removes the experimental grounding for the central claim. No alternative dataset is supplied, so this error alone invalidates the experiment as described.","section":"Section III-A"},{"comment":"The experimental results are asserted without any protocol: there is no description of the number of participants, number of gesture samples per class, train/test split, data preprocessing, hyperparameter settings (beyond an unspecified dropout probability p), training epochs, or hardware. Table 1 reports response speeds, fluency scores, and satisfaction ratings, but these are presented without error bars, confidence intervals, or any indication of how they were measured. The comparative accuracies (95% vs. 90% vs. 85%) are stated in prose with no citation to a specific experiment or baseline implementation. Consequently, the central claims of accuracy and real-time performance are unsupported and cannot be independently checked.","section":"Section III-B and Table 1"},{"comment":"The 3D hand skeleton model, which is the paper's main proposed innovation, is never shown to be a functional component of the recognition system. Section II describes a CNN-LSTM pipeline but does not specify how the 3D skeleton is fed into or fused with the network; the only mention of the skeleton is in the figure caption and the descriptive text in Section III-B, where it is characterized as a 'basic visualization framework.' The claimed improvement in accuracy and efficiency is therefore never tested, and the mechanism by which the 3D skeleton improves recognition remains unspecified.","section":"Section II and Figure 2"}],"minor_comments":[{"comment":"The abstract's phrase 'maintain high recognition accuracy and real-time response capabilities in different environments' is a general claim not tied to any specific quantitative result; consider stating the measured conditions or deferring such claims to the results section.","section":"Abstract and Section I"},{"comment":"Several references appear unrelated to gesture recognition or to the techniques they are cited for (e.g., [11]-[16], [20]-[26] on recommendation systems, NLP, and data mining), while the actual gesture-recognition literature is only sparsely cited; the authors should cite relevant prior work and remove padding references.","section":"References"},{"comment":"The Dropout equation is malformed in the rendering ('pBernoullihh ii' and other garbled characters); the formula should be typeset correctly using standard notation, e.g., h_i' = h_i * Bernoulli(p).","section":"Section II, Dropout formula"},{"comment":"Figure 1 is referenced as the overall framework, but the caption provides no description of the layers, dimensions, or the role of the two gray LSTM areas; a more detailed caption or an accompanying text description would improve clarity.","section":"Figure 1"}],"recommendation":"reject","confidential_remarks":"This manuscript appears to be a position/vision paper without actual experimental validation. The dataset error in Section III-A is not a minor slip; it invalidates the entire empirical section. The citation list includes many arXiv preprints that are topically unrelated to gesture recognition, which suggests the references may have been added to inflate the bibliography rather than to support the technical content. Given the lack of reproducible evidence and the misidentified dataset, rejection is appropriate. The authors could resubmit a substantially revised version with a correctly described dataset, a full experimental protocol, and actual results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: this paper is not publishable. The central experiment claims 95% gesture recognition accuracy on a dataset called GazeCapture, described as 'a large number of gesture images and corresponding user gaze data.' GazeCapture is a well-known eye-tracking dataset; it contains face images and gaze targets, not hand gestures. That single error collapses the experimental foundation. There is no protocol, no sample sizes, no hyperparameters, no code, no error bars. The comparative numbers (95% vs 90% vs 85%) appear in a paragraph with no provenance. The multimodal eye-tracking fusion is described in one sentence. The 3D hand skeleton model that appears in the title and abstract is never integrated into the CNN-LSTM pipeline; it is only a figure with a caption. The architecture itself—CNN for spatial features, LSTM for temporal modeling, dropout, data augmentation—is textbook material, covered by the paper's own references. There is no algorithmic novelty here.\n\nWhat does the paper do well? Very little. It correctly states that gesture recognition matters for VR, AR, and smart home, and it lists common challenges like illumination changes and individual variation. That's background knowledge, not a contribution. The citation list is padded with unrelated work (recommendation systems, medical NER, stock price prediction), which suggests the references are there to meet a length requirement rather than to frame an argument.\n\nThe stress-test note is right on the money. I checked the manuscript text: Section III-A explicitly names GazeCapture and describes it as a gesture dataset. That is a factual error about a public dataset, and it vitiates the entire experimental section. The paper also fails to explain how the gesture images were labeled, how many samples were used, how the LSTM was trained, or how the reported response times were measured. The result is a claim-without-evidence.\n\nMy recommendation: desk reject. This does not deserve referee time because the load-bearing assumption is demonstrably false. If the authors had actually run this experiment, they would know they used a different dataset, which suggests the numbers are not from the described setup. The paper is not coherent enough for a serious review, and there is no reproducible artifact to check.","headline":"A gesture-recognition paper whose stated dataset does not contain gestures; the 95% accuracy claim is unsupported by any reproducible protocol.","tokens_in":7961,"tokens_out":2541,"would_cite":false,"duration_ms":22382,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims a CNN-LSTM model with a 3D hand-skeleton visualization recognizes gestures with 95% accuracy and 0.12–0.25 s response times, outperforming CNN-only and manual-feature methods.","keywords":["gesture recognition","computer vision","human-computer interaction","3D hand skeleton model","CNN-LSTM","eye tracking","multimodal fusion","real-time interaction"],"falsifier":"Open the GazeCapture dataset and count the images that carry gesture-category labels. If no gesture-labelled images exist, the training procedure described in Section III-A cannot be reproduced and the 95% accuracy claim has no empirical support from the stated dataset.","tokens_in":6984,"feed_emoji":"🖐️","tokens_out":5249,"duration_ms":43485,"temperature":0.7,"pith_summary":"The paper is trying to establish that natural hand gestures, captured by an ordinary camera and represented as a three-dimensional hand skeleton, can serve as a fast and accurate input channel for human-computer interaction. It proposes a CNN-LSTM pipeline: the CNN extracts spatial features from gesture images, the LSTM models the temporal sequence of a gesture, and the 3D skeleton visualization of the hand carries the structural information. The authors report 95% recognition accuracy, which they claim beats CNN-based methods at 90% and manual-feature approaches at 85%, with system response times of 0.12 to 0.25 seconds in VR, AR, smart home, and gaming scenarios. If these numbers hold, the work supports replacing controllers and menus with contactless gesture control in interactive systems.","feed_headline":"CNN-LSTM with 3D hand skeleton claims 95% gesture accuracy","feed_subtitle":"Virtual reality, AR, smart-home and game controls respond in 0.12-0.25 seconds with eye-tracking fusion.","key_machinery":"The central object is the 3D hand-skeleton model: a simplified graph of the hand in which the palm center connects to the base of each finger and the joints of each finger connect in sequence to the fingertip, with colored lines showing the skeleton. This model is the structural input that the CNN processes spatially and the LSTM processes temporally, so the skeleton carries the gesture-shape information that the networks classify. The other active components are cross-entropy loss for training, dropout for regularization, and an attention-weighted fusion layer that combines eye-tracking gaze data with gesture features.","core_discovery":"On its own terms, the paper's discovery is that a three-dimensional hand-skeleton graph, when combined with a CNN-LSTM network and an eye-tracking attention mechanism, yields a gesture recognition system that is both accurate and responsive enough for real-time interaction. The skeleton is built by connecting the palm center to each finger base and then linking finger joints sequentially, which the authors say captures finger bending and hand posture. The reported experiments give the system a 95% accuracy, besting a CNN-only system at 90% and a manual-feature system at 85%, with latency ranging from 0.12 seconds in a virtual-reality environment to 0.25 seconds in complex backgrounds. The paper further claims that gaze data, fused through an attention-weighted layer, helps the system infer user intent in cluttered scenes.","pith_inferences":["Editorial inference: the same CNN-LSTM skeleton architecture could be benchmarked on a labelled hand-gesture dataset to isolate how much of the 95% comes from the skeleton representation versus the dataset.","Editorial inference: if gaze fusion adds real signal, the attention-weighted fusion layer could be extended to other cues such as head pose or object context without changing the backbone.","Editorial inference: the reported 0.12–0.25 second response range spans VR to low-light clutter, which suggests latency will be environment-dependent; a deployment study would need to measure per-frame inference time on target hardware."],"forward_implications":["In virtual reality, the reported 0.12-second response means gesture commands can feel immediate rather than delayed.","A 95% accuracy versus 90% for CNN-only and 85% for manual features implies the skeleton representation adds measurable discriminative power.","Response times under 0.25 seconds in complex backgrounds imply the system can operate outside controlled lighting or clean backdrops.","Fusing gaze through an attention layer implies the same model can weigh visual regions the user is looking at, which may reduce ambiguity between similar gestures.","The five tested scenarios, from VR to smart-home control, indicate one pipeline can cover several interaction form factors."],"supporting_citations":[{"why":"Establishes the real-time skeleton-based gesture recognition baseline the paper aims to extend.","marker":"[2]"},{"why":"Provides the deep-learning gesture recognition context that motivates replacing manual features.","marker":"[6]"},{"why":"Supplies the multi-modal deep-learning framework whose principles the method extends.","marker":"[19]"},{"why":"Justifies the CNN-LSTM temporal modeling choice for gesture transitions.","marker":"[21]"},{"why":"Motivates the dropout regularization used in training.","marker":"[24]"},{"why":"Underpins the 3D skeleton-based hand model used for visualization and recognition.","marker":"[27]"}],"fun_headline_variants":["3D hand skeleton + eye tracking hits 95% gesture accuracy","Hand skeleton graph + gaze fusion drives real-time gesture control","Eye-tracking attention boosts gesture recognition to 95%","3D skeleton + CNN-LSTM: 95% accuracy, real-time hand gestures","Gesture AI: 3D hand model and gaze yield 95% real-time accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire experiment rests on the claim that the GazeCapture dataset contains a large number of gesture images paired with user gaze data; if that dataset has no gesture images, the reported training setup, multimodal fusion, and 95% accuracy are unsupported.","fun_headline_variants_meta":{"raw":{"variants":["3D hand skeleton + eye tracking hits 95% gesture accuracy","Hand skeleton graph + gaze fusion drives real-time gesture control","Eye-tracking attention boosts gesture recognition to 95%","3D skeleton + CNN-LSTM: 95% accuracy, real-time hand gestures","Gesture AI: 3D hand model and gaze yield 95% real-time accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000529,"raw_usage":{"total_tokens":2548,"prompt_tokens":942,"completion_tokens":1606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":1511}},"tokens_in":558,"tokens_out":1606,"duration_ms":11210,"temperature":1.0,"reasoning_tokens":1511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:48:28.980176+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Open the GazeCapture dataset and count the images that carry gesture-category labels. If no gesture-labelled images exist, the training procedure described in Section III-A cannot be reproduced and the 95% accuracy claim has no empirical support from the stated dataset.","supporting_citations":[{"cited_title":"Adaptive Cache Management for Complex Storage Systems Using CNN-LSTM-Based Spatiotemporal Prediction","cited_arxiv_id":"2411.12161","evidence_quote":"Justifies the CNN-LSTM temporal modeling choice for gesture transitions."},{"cited_title":"Development of skeleton-based gait models for human movement recognition based on neural networks,","cited_arxiv_id":null,"evidence_quote":"Underpins the 3D skeleton-based hand model used for visualization and recognition."},{"cited_title":"Real-Time Hand Gesture Recognition: Integrating Skeleton-Based Data Fusion and Multi-Stream CNN","cited_arxiv_id":"2406.15003","evidence_quote":"Establishes the real-time skeleton-based gesture recognition baseline the paper aims to extend."},{"cited_title":"3D Convolutional Network based micro-gesture recognition,","cited_arxiv_id":null,"evidence_quote":"Provides the deep-learning gesture recognition context that motivates replacing manual features."},{"cited_title":"Emotion-Aware Interaction Design in Intelligent User Interface Using Multi-Modal Deep Learning","cited_arxiv_id":"2411.06326","evidence_quote":"Supplies the multi-modal deep-learning framework whose principles the method extends."}],"review_version":1}