{"id":"225b5def-7281-4c0c-bd69-646a6a0bf58c","arxiv_id":"2512.00572","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Skeleton-based inputs beat raw photos for yoga pose classification, with VGG16 plus MediaPipe skeletons reaching 96.09% on the new 16-class Yoga-16 dataset.","lead":"This paper builds a new dataset of 16 yoga poses and tests three neural networks on three input styles: raw photos, MediaPipe skeleton drawings, and YOLOv8 skeleton drawings. It finds that feeding the network skeleton drawings instead of photos improves accuracy, with the best model reaching about 96% on the new dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 96.09% headline accuracy may be inflated by test-set hyperparameter selection and inconsistent training protocols; the skeleton-vs-raw comparison is not yet trustworthy.","rationale":"The reader correctly identified that the dataset is curated and the test set is a random split from the same pool, which limits external generalization claims. However, the more load-bearing concern for the paper's central claim is internal validity: hyperparameter selection appears to have used the test set, and the training protocols across models/inputs are inconsistent. If test-set tuning occurred, the 96.09% accuracy is not a fair estimate of generalization, and the superiority of skeleton inputs could be an artifact of tuning rather than a real property of the representation. This concern is concrete and checkable from the released code, and it directly undermines the main contribution. I agree with the reader that the verdict should be conditional rather than reject, because the concern is not yet confirmed and the paper provides some evidence (e.g., 5-fold CV, custom test set) that partially mitigates. The verdict remains CONDITIONAL, but with a sharper basis: the empirical headline awaits verification of the evaluation pipeline rather than merely external dataset diversity.","tokens_in":19504,"tokens_out":4406,"duration_ms":45869,"concrete_test":"Inspect the released GitHub script (mohiuddin2531/yoga-16) to determine which data split is used in the hyperparameter tuning loop corresponding to Table 14. If the test split is referenced in model selection, the 96.09% estimate is invalid. Then re-run the comparison using the held-out validation split (Table 5) for hyperparameter selection and matched training protocols (both from scratch or both fine-tuned) for VGG16 with direct images and MediaPipe skeletons; report test accuracy. If the gap between modalities drops to within a few percent, or the selected model no longer achieves 96.09% on the test set, the skeleton-advantage claim is unsupported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim—skeleton inputs beat raw images—rests on the numerical gap between VGG16+MediaPipe (96.09%) and VGG16+direct (86.33%). But Section 4.5 and Table 14 select hyperparameters by 'Overall Accuracy' without reporting which split is used. The best configuration in Table 14 (Experiment 1) yields exactly 96.09%, the same number as Table 12, suggesting the test split may have been used for model selection. No separate validation results are reported, and the later 5-fold CV (Table 15) is applied only after the fact to the already-chosen model. If the test set was used to pick hyperparameters, the headline number is an optimistic selection artifact rather than an unbiased estimate, and the claimed superiority of skeleton inputs may shrink or disappear under honest evaluation.\n\nAdditionally, the training protocol is internally inconsistent: Section 3.6 states models are trained from scratch, yet Section 3.6.1–3.6.3 and Tables 7–8 describe ImageNet-initialized fine-tuning with frozen backbone layers. If different input types or architectures used different initialization/fine-tuning schedules, the input-modality comparison is confounded with optimization choices. This is a more direct threat to the paper's central empirical claim than dataset curation alone, because it affects the internal validity of the very accuracy numbers being compared.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces a curated dataset, Yoga-16, containing 16 yoga pose classes with 1,280 images (80 per class), and reports a benchmark comparing VGG16, ResNet50, and Xception under three input modalities: direct RGB images, MediaPipe Pose skeleton images, and YOLOv8-Pose skeleton images. The paper's central empirical claim is that skeleton-based representations outperform raw images for yoga pose classification, with the best result being VGG16 on MediaPipe skeleton inputs at 96.09% accuracy. The authors also report 5-fold cross-validation (93.55±0.94%), a custom YouTube-derived test set (93.75% for the best model), Grad-CAM visualizations, and a qualitative comparison with prior work.","tokens_in":19764,"tokens_out":5709,"duration_ms":56736,"significance":"If the comparison is trustworthy, the paper provides a useful empirical benchmark and a reusable dataset for yoga pose classification. The public release of the dataset and code is a concrete strength that supports reproducibility. The contribution is primarily experimental and incremental rather than methodological: the architectures and skeleton extractors are established tools, and the main novelty lies in the systematic three-modality, three-architecture comparison. The dataset is small and deliberately curated, so the headline accuracy should be read as a proof-of-concept result on clean, single-subject, well-cropped images rather than as a claim about unconstrained real-world yoga videos.","major_comments":[{"comment":"The training protocol is stated inconsistently. Section 3.6 says the CNN models are trained from scratch because pretrained ImageNet weights are not well aligned with skeleton inputs. However, §3.6.1 describes fine-tuning VGG16 by unfreezing the last dense layers and partially the deeper convolutional blocks; Table 6 marks all VGG16 layers trainable, while Tables 7 and 8 mark the ResNet50 and Xception backbone stages/flow layers as not trainable. These are different training protocols, and if actually used, the accuracy differences in Table 12 confound input modality with initialization and layer-freezing choices. The authors should specify the exact protocol per experiment and, ideally, rerun the nine experiments with a single consistent training protocol before drawing the central conclusion that skeleton inputs are superior to raw images.","section":"§3.6, §3.6.1–3.6.3, Tables 6–8"},{"comment":"The hyperparameter tuning experiments are reported using 'Overall Accuracy' without stating which data split is used for model selection. The best configuration (Experiment 1) reports exactly 96.09%, the same number as the headline result in Table 12, which raises the concern that the test split was used to select hyperparameters. No validation-set accuracies are given for the eight tuning runs. Because Table 12 is the basis for the paper's central claim, this is load-bearing. The authors should clearly report the split used in Table 14 and provide the corresponding validation results; the final test accuracy should be computed only after the hyperparameters are fixed. The 5-fold cross-validation in Table 15 was applied after the model was already selected, so it does not remove this selection bias.","section":"§4.5, Table 14"},{"comment":"The external generalization evidence does not directly support the central modality comparison. The custom YouTube test set is evaluated only for VGG16 with MediaPipe skeleton inputs; no direct-image or YOLOv8-skeleton models are tested on it, so it cannot establish whether the claimed skeleton advantage persists under distribution shift. In addition, Yoga-16 curation explicitly excludes multi-subject, low-quality, non-human, and zoomed-out images (Section 3.2, Fig. 2), and the main test set is a random split from the same curated pool. The 96.09% vs. 86.33% gap therefore reflects a clean, curated setting. To support the broad conclusion that skeleton-based representations are more robust, the authors should compare raw-image and skeleton variants on the same external set, or explicitly scope the claim to the curated setting.","section":"§4.6, §3.2, Table 16"}],"minor_comments":[{"comment":"The first sentence after the VGG16 description says the confusion matrix, loss/accuracy curves, and ROC curve are for Xception with YOLOv8-Pose input, but the surrounding text is discussing VGG16. The figure references (S19–S21) appear to be misassigned.","section":"§4.3"},{"comment":"The section headings use the phrase 'Pretrained Deep Learning Models', which is inconsistent with Section 3.6's statement that models are trained from scratch. Please align the terminology.","section":"§4.2 and §4.3 headings"},{"comment":"The per-class 'Accuracy' column is unusual because, for a single-label test set with one class per image, per-class accuracy coincides with recall. Clarify the definition or rename it to avoid confusion.","section":"Table 13"},{"comment":"The comparison with prior work is qualitative ('R/NR'). State whether categories such as 'low resolution' and 'class imbalance' are measured objectively or are author judgments.","section":"Table 17"},{"comment":"The self-citations for the dataset, custom test set, and code are appropriate for data availability, but the custom test set's construction and selection criteria should be described in the main text rather than left entirely to the referenced repository.","section":"References [34], [36], [37]"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable empirical contribution, and the public release of data and code is commendable. The main concern is internal validity: the training-protocol inconsistency and the apparent use of the test split for hyperparameter selection undermine confidence in the headline accuracy comparison. These issues are fixable with additional reporting and possibly additional experiments, so I recommend major revision rather than rejection. The self-citations are legitimate for data availability and do not by themselves raise integrity concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper gives you a usable curated dataset and a systematic three-way comparison of raw images vs MediaPipe skeletons vs YOLOv8 skeletons across three CNNs, but the headline 96.09% accuracy is not something I would quote. The training-protocol description contradicts itself, and the hyperparameter tuning appears to have been done on the test set.\n\nWhat is genuinely useful: Yoga-16 is small (1,280 images, 80 per class) but balanced and cleaned of multi-subject and low-quality frames, and the authors release data and code. Running 9 architecture-by-input combinations, adding Grad-CAM, and doing 5-fold CV plus a YouTube-based test set is more than most yoga-pose papers do. The qualitative direction — MediaPipe skeletons beat raw RGB — matches Garg et al. and is plausible.\n\nThe soft spots are real. Section 3.6 says models were trained from scratch because ImageNet weights are not aligned with skeletons, yet Sections 3.6.1–3.6.3 and Tables 7–8 describe fine-tuning with frozen backbones. That is not a minor wording issue: if different input types used different initialization schedules, the skeleton-vs-raw comparison is confounded. Second, Section 4.5's hyperparameter table gives no validation accuracy; the best row reproduces 96.09% exactly, the same number as the main evaluation, which strongly suggests the test split was used for model selection. The 5-fold CV mean of 93.55±0.94 is the more trustworthy figure. Third, the dataset is small and curated to remove hard cases, so the main benchmark is an upper-bound scenario; the custom YouTube set is a step toward realism but is only 320 images and is also self-authored. The generalization evidence is therefore not independent.\n\nSome smaller issues: Section 4.3's text appears to copy wrong model names (it discusses Xception where VGG16 results are being reported), Table 17's comparison is impressionistic, and the claimed novelty is modest — prior work already concluded skeletonization helps.\n\nWho this is for: people building practical yoga pose classifiers or looking for a compact benchmark dataset. It is not a methodological advance, and the empirical comparison needs re-doing with a separate validation split and a consistent training protocol. That said, the dataset and code availability make it worth a serious referee rather than a desk reject. I'd recommend major revision: fix the protocol, add real validation-based selection, report the honest distribution of results, and ideally get an external test set.","headline":"A useful curated dataset and a plausible skeleton-beats-RGB direction, but the 96.09% headline is inflated by test-set hyperparameter selection and a self-contradictory training protocol.","tokens_in":20325,"tokens_out":3124,"would_cite":false,"duration_ms":30760,"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":"The paper's central claim is that skeletal renderings of the body carry more pose-discriminating signal than full RGB images, and on the new Yoga-16 benchmark VGG16 with MediaPipe skeleton input reaches 96.09% accuracy.","keywords":["Human Pose Estimation","Yoga Pose Classification","Deep Learning","MediaPipe Pose","YOLOv8 Pose","Skeleton-Based Representations","Yoga-16 dataset","Transfer Learning"],"falsifier":"Run the identical VGG16 pipeline on uncurated or multi-person images where keypoint extraction is imperfect: if accuracy on skeleton inputs drops below raw-image accuracy, the claim that skeletons generally outperform photos is falsified. A second test: include a pose class that is distinguishable mainly by clothing or props (e.g., a pose holding a block); skeleton inputs would likely fail where raw images succeed.","tokens_in":19371,"feed_emoji":"🧘","tokens_out":5290,"duration_ms":50560,"temperature":0.7,"pith_summary":"This paper sets out to show that representing a yoga pose as a skeleton—body joints and the lines between them, stripped of background and clothing—makes pose classification more accurate than feeding a CNN the original photograph. To test this, the authors built Yoga-16, a curated and balanced dataset of 16 poses with overlapping visual features, and ran three architectures (VGG16, ResNet50, Xception) on three input types: raw images, MediaPipe Pose skeletons, and YOLOv8-Pose skeletons. Skeleton inputs won across the board; the best model, VGG16 on MediaPipe skeletons, reached 96.09% accuracy, versus 86.33% for the same network on raw images. If this holds, automated yoga coaching could rely on lightweight, real-time skeleton extraction rather than heavy image analysis, and the benchmark gives the field a common test bed for comparing approaches.","feed_headline":"Yoga pose AI hits 96% when fed skeletons, not photos","feed_subtitle":"On a balanced 16-pose dataset, skeleton input beats raw photos across three CNN models, pointing to reliable real-time yoga coaching.","key_machinery":"The skeleton-image representation is the load-bearing mechanism: a pose-estimation model (MediaPipe Pose or YOLOv8-Pose) detects body keypoints and renders them as a clean skeletal overlay, which is then fed to a CNN trained from scratch. This transformation strips away background, clothing, and lighting so the classifier is forced to learn from joint geometry alone. VGG16's sequential stack of small 3×3 filters with 2×2 max pooling turned out to be the best exploiter of these skeletal images, and hyperparameter tuning (filter size, pooling, batch size, optimizer, dense width) confirmed those settings.","core_discovery":"The paper's central discovery is that input representation, not architecture, is the dominant factor in yoga pose classification: converting each image into a skeleton visualization before classification consistently improves accuracy across all three CNN backbones. The best configuration, VGG16 with MediaPipe Pose skeleton images, achieves 96.09% accuracy, 96.27% precision, 96.09% recall, and 96.10% F1-score on Yoga-16—about ten points above the same VGG16 on raw photographs. The authors report that MediaPipe's 33-keypoint skeletons, rendered without background, let the network attend to limb alignments and joint angles, and that this advantage persists under 5-fold cross-validation (93.55","pith_inferences":["A likely boundary condition: the skeleton advantage depends on the curation choices in Yoga-16 (single subject, cropped, high-quality, no clutter). On messier multi-person or heavily occluded images, keypoint extraction errors could shrink or reverse the gap, so real systems may need a person-detection front end before skeletonization.","The consistent gap between MediaPipe and YOLOv8-Pose skeletons (e.g., VGG16: 96.09% vs 91.41%) suggests keypoint stability matters more than keypoint count; re-running the same protocol with other keypoint estimators would test whether the ranking is estimator-agnostic.","Misclassifications between poses with overlapping skeletal features (Dolphin Plank vs. Fish, Dolphin Plank vs. Wide Angle Seated Forward Bend) point to a concrete next experiment: adding pose-specific data augmentation or an auxiliary loss on joint angles could directly target the failure modes the authors identified.","Because training was from scratch on only 896 training images, the accuracy ceiling may be much higher with more data; a scaled-up version of Yoga-16 could reveal whether skeleton inputs retain their edge when appearance cues become available in larger variety."],"forward_implications":["On Yoga-16, switching VGG16 from raw photos to MediaPipe skeletons raises accuracy from 86.33% to 96.09%, showing representation choice can matter more than model choice.","The skeleton advantage holds across all three architectures: every network scores higher on either skeleton input than on direct images, so the result is not specific to VGG16.","The 93.75% accuracy on a YouTube-derived test set suggests the pipeline transfers to less controlled imagery, supporting real-world deployment.","Because MediaPipe skeleton extraction is lightweight and real-time, the result implies on-device yoga coaching is feasible without sacrificing accuracy.","The 5-fold cross-validation mean of 93.55 ± 0.94% indicates the reported accuracy is stable across data splits, not a lucky split artifact."],"fun_headline_variants":["Skeleton input boosts yoga AI accuracy to 96%","Yoga pose AI: skeletons beat photos by 10 points","96% yoga pose accuracy with skeleton images","Skeleton-based yoga classification beats raw photos"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central claim rests on assuming that a clean skeleton image carries enough information to distinguish all 16 poses, an assumption the dataset curation was built to satisfy by keeping only single-subject, well-cropped, high-quality images.","fun_headline_variants_meta":{"raw":{"variants":["Skeleton input boosts yoga AI accuracy to 96%","Yoga pose AI: skeletons beat photos by 10 points","96% yoga pose accuracy with skeleton images","Skeleton-based yoga classification beats raw photos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000766,"raw_usage":{"total_tokens":3224,"prompt_tokens":724,"completion_tokens":2500,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":2438}},"tokens_in":468,"tokens_out":2500,"duration_ms":17114,"temperature":1.0,"reasoning_tokens":2438,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:22:13.471923+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical VGG16 pipeline on uncurated or multi-person images where keypoint extraction is imperfect: if accuracy on skeleton inputs drops below raw-image accuracy, the claim that skeletons generally outperform photos is falsified. A second test: include a pose class that is distinguishable mainly by clothing or props (e.g., a pose holding a block); skeleton inputs would likely fail where raw images succeed.","supporting_citations":[],"review_version":1}