{"id":"ce942261-5223-42d0-93df-070444f872ee","arxiv_id":"1908.00646","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A skeleton-only action recognition method on PSD matrices with a quotient metric, Bezier curve fitting, and global alignment kernel achieves 97.99% on UTKinect, 96.16% on KTH, and 92.44% on UAV-Gesture.","lead":"This paper builds a skeleton-based action recognition pipeline where each video becomes a curve on the manifold of fixed-rank positive semidefinite matrices, using a different metric, curve smoothing, and a global alignment kernel. It reports competitive accuracies on three public datasets, slightly beating its immediate predecessor but trailing deep learning methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4) in §3.2 is false: the closed-form d=2 distance minimizes over rotations only, not O_d, so it overestimates distances when det(Z_j^T Z_i)<0, potentially corrupting the KTH/UAV-Gesture results.","rationale":"The reader's PSD concern is valid but conditional: the local kernel's PSD-ness is asserted, not proven. The distance-formula error is a demonstrable false theorem that can directly change the numeric results on two of the three datasets. The paper's central empirical claim ('competitive results') is at stake if the distances used to build the GAK kernel are wrong. Because the paper does not identify which of M1/M2 produced Tables 3 and 5, a condition is necessary: correct Eq. (4), clarify the metric used, and re-evaluate. This does not reject the paper outright since the method might still be competitive when using the correct metric, but it is not acceptable as-is.","tokens_in":14060,"tokens_out":24453,"duration_ms":219431,"concrete_test":"Run the distance counterexample: take Z_i=[[1,0],[0,1]], Z_j=[[0,1],[1,0]]. Compute Eq. (3) via SVD (should be 0) and Eq. (4) (should be 2). Then re-run the KTH and UAV-Gesture experiments using Eq. (3) instead of Eq. (4) for all distance computations and compare the resulting accuracies; if the 96.16% and 92.44% figures change, the reported results depend on the erroneous formula.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing weakness is the closed-form distance for d=2 in Section 3.2, Eq. (4). The paper identifies S_+(d,n) with R^{n*d}_*/O_d and defines d(G_i,G_j)=min_{Q∈O_d}||Z_j Q - Z_i||_F. The maximum of tr(Z_j^T Z_i Q^T) over Q∈O_2 is the nuclear norm of M=Z_j^T Z_i, i.e. sqrt(tr(M^T M)+2|det M|). Eq. (4) instead uses sqrt((a+d)^2+(c-b)^2), which equals the nuclear norm only when det(M)≥0. The appendix proof parameterizes Q as a rotation matrix only, omitting reflections, which are in O_2. Counterexample: Z_i=I_2, Z_j=[[0,1],[1,0]]; both give Gram matrix I (same point), so the true distance is 0, but Eq. (4) gives 2. KTH and UAV-Gesture are 2D (d=2) and the paper does not state whether the results in Tables 3 and 5 used M1 (Eq. 3) or M2 (Eq. 4). If M2 was used, the kernel and GAK alignment are computed with a corrupted distance, so the reported accuracies, including the UAV-Gesture improvement over P-CNN, may not hold under the correct metric.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a skeleton-based action recognition pipeline. Each frame's landmarks are mapped to a Gram matrix in the fixed-rank positive semidefinite cone S_+(d,n); a sequence becomes a trajectory on that manifold. The trajectory is denoised with composite Bézier curve fitting, sequences are aligned with a Global Alignment Kernel built from a Riemannian distance, and the resulting kernel matrix is fed to a standard SVM. Experiments on UTKinect-Action3D, KTH-Action, and UAV-Gesture report accuracies of 97.99%, 96.16%, and 92.44% respectively, which the authors describe as competitive with state-of-the-art methods and better than their direct predecessor [18] on UTKinect and better than the P-CNN baseline on UAV-Gesture.","tokens_in":14360,"tokens_out":7744,"duration_ms":73178,"significance":"The paper is a useful engineering contribution if the mathematical components hold: it provides a non-deep, interpretable alternative that achieves competitive accuracy from skeletons only, and the use of GAK rather than DTW is a sensible step toward a valid SVM kernel. However, the closed-form distance in Section 3.2 is incorrect for d=2, and the PSD property of the GAK on this manifold is asserted rather than proved; both issues bear directly on the reported numbers. The paper also includes detailed timing tables and comparison tables, which help reproducibility.","major_comments":[{"comment":"Theorem 3.2 is false as stated. The distance in Eq. (3) minimizes over the full orthogonal group O_2, which includes reflections, but the proof in the Appendix parameterizes Q only as a rotation matrix. For Z_i = I_2 and Z_j = [[0,1],[1,0]], both Gram matrices equal I_2, so the true quotient distance is 0, whereas Eq. (4) returns 2. The correct maximization of tr(Z_i^T Z_j Q^T) over O_2 is the nuclear norm of Z_i^T Z_j, which equals sqrt((a+d)^2+(c-b)^2) only when det(Z_j^T Z_i) >= 0; otherwise it is larger. This error directly affects the d=2 experiments on KTH and UAV-Gesture.","section":"§3.2, Eq. (4), and Appendix"},{"comment":"The paper asserts without proof that the resulting global alignment kernel is positive semi-definite on S_+(d,n). The cited PSD result in Cuturi et al. [8] applies to local kernels that are positive definite; no argument or reference is given that exp(-d^2/sigma^2) is positive definite for the quotient/Bures metric on fixed-rank PSD matrices. Since the SVM step relies on the kernel matrix being PSD, this is a load-bearing gap. Please provide a proof or a specific citation, or state explicitly that a non-PSD proximity matrix is used with a suitable variant such as ppfSVM.","section":"§5.1, Eqs. (6)-(8)"},{"comment":"The accuracy tables for KTH and UAV-Gesture do not state whether Eq. (3) (M1) or Eq. (4) (M2) was used. Because Eq. (4) is incorrect, the reported accuracies are ambiguous; if M2 was used, the numbers may change under the correct O_2 distance. Please specify the metric used for each reported result and rerun the experiments if M2 was used.","section":"§6, Tables 3 and 5"}],"minor_comments":[{"comment":"There is an indexing mismatch: the loops that fill k(i,j) range over i=0..tau1 and j=0..tau2, while the recurrence for M uses indices i=1..tau1+1 and j=1..tau2+1 without a consistent offset; please clarify the indexing convention.","section":"Algorithm 1"},{"comment":"In Eq. (10) the orthogonal group is written as O_p; it should be O_2, consistent with Section 3.2.","section":"Appendix, Eq. (10)"},{"comment":"The hyperparameters lambda and sigma are tuned per dataset but the selection criterion is not described; please report how these values were chosen (for example, via a validation split) to mitigate overfitting concerns.","section":"§6"},{"comment":"The sentence 'This strategy assures us that the kernel is positive semi-definite' overstates the conclusion; at most the construction inherits the property if the local kernel is positive definite, which is exactly the point that needs to be established.","section":"§5.1"}],"recommendation":"major_revision","confidential_remarks":"The main correctness concerns are the false d=2 distance theorem and the unproven PSD property of the GAK on this manifold. If the authors correct Eq. (4) and re-run the d=2 experiments, the empirical claims may still hold; I would not recommend rejection because the overall pipeline is reasonable. As written, however, the reported accuracies may be affected by an invalid distance, so the manuscript cannot be accepted in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: Eq. (4) is not the quotient distance defined in Eq. (3). The quotient is over all O_2, including reflections, but the appendix proof maximizes only over rotations. For any Z_i, Z_j with det(Z_j^T Z_i) < 0, the true distance is the nuclear norm, and Eq. (4) overestimates it. Example: Z_i = I_2 and Z_j = [[0,1],[1,0]] give the same Gram matrix, so the true distance is 0, but Eq. (4) gives 2. Since the paper does not state whether the KTH and UAV-Gesture results used M1 (Eq. 3) or M2 (Eq. 4), the empirical claims are not reproducible.\n\nThere is real substance here. The paper is a clean, honest incremental extension of Kacem et al. [18]: a different metric on the fixed-rank PSD manifold, composite cubic Bezier fitting, and GAK for alignment. The geometric framework is sound when the correct distance (Eq. 3) is used, and the experiments show competitive results on three benchmarks. The comparison to [18] is fair, and the writing is clear.\n\nThe soft spots are three. First and most serious: the theorem error above. If M2 was used in the experiments, the distances, the GAK kernel, and the reported accuracies are all corrupted. Second, the PSD property of the global alignment kernel is assumed via citation to Cuturi et al., but that theorem requires the local kernel to be PSD on the space; this is not proven for the Bures-type distance on the quotient manifold. Third, hyperparameters (lambda, sigma) are selected on the test set with no validation protocol or error bars, which inflates accuracy. No code is released, so reimplementation is the only check.\n\nThis paper is for researchers working on geometric skeleton-based action recognition. It deserves a serious referee, but not acceptance in its current form. The authors need to correct Eq. (4) or explicitly state that all reported results use Eq. (3), prove or empirically verify the PSD property of the GAK with this local kernel, and disclose their model selection procedure.","headline":"The closed-form distance in Eq. (4) is wrong as stated (it ignores reflections), and the paper never says whether the headline numbers used the correct M1 or the incorrect M2.","tokens_in":14931,"tokens_out":4375,"would_cite":false,"duration_ms":44129,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45","53B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a skeleton-only geometric pipeline—Gram-matrix trajectories on a fixed-rank PSD manifold, denoised by curve fitting and aligned with a Global Alignment Kernel—matches or beats deep learning baselines on three…","keywords":["action recognition","skeleton-based","Gram matrices","positive semidefinite manifold","Riemannian geometry","global alignment kernel","curve fitting","SVM"],"falsifier":"Take any dataset used in the paper, compute the full similarity matrix $K$ from the GAK scores, and test its smallest eigenvalue; if $K$ has a negative eigenvalue, the kernel is not PSD for this manifold distance and the standard SVM argument fails, even if classification accuracy stays high. A sharper test is to sample pairs of Gram matrices in $S_+(d,n)$, compute the local kernel matrix $k(i,j) = \\frac{1}{2}e^{-D(i,j)/\\sigma^2}/(1-\\frac{1}{2}e^{-D(i,j)/\\sigma^2})$, and check whether it is positive semidefinite for many random configurations.","tokens_in":13852,"feed_emoji":"🦴","tokens_out":7006,"duration_ms":58741,"temperature":0.7,"pith_summary":"This paper claims that human actions can be recognized from body skeletons alone by representing each video frame as a Gram matrix of joint coordinates and treating the whole action as a trajectory on the manifold of fixed-rank positive-semidefinite matrices. To compare and classify such trajectories, it replaces the Dynamic Time Warping used in prior work with a Global Alignment Kernel, built on a quotient-manifold distance that has a closed form when the skeleton is two-dimensional. The fitted, aligned trajectories are fed to a standard SVM through a kernel the paper asserts is positive semidefinite. If the claim holds, the approach provides a parameter-light, interpretable alternative to deep models: it reports 97.99% on UTKinect, 96.16% on KTH, and 92.44% on UAV-Gesture, the last improving on the provided deep baseline while using only skeletons.","feed_headline":"Skeleton-only action recognition hits 97.99% on UTKinect","feed_subtitle":"A geometric pipeline with a global alignment kernel beats the previous Gram-matrix method and a deep baseline on UAV-Gesture.","key_machinery":"The central object is the manifold $S_+(d,n)$ of $n\\times n$ positive-semidefinite matrices of fixed rank $d$, realized as the quotient $\\mathbb{R}^{n\\times d}_*/O_d$: each Gram matrix $G=ZZ^T$ corresponds to an equivalence class of landmark matrices under right multiplication by orthogonal matrices. The distance between two such classes is the minimal Frobenius distance $\\min_{Q\\in O_d}\\|Z_jQ-Z_i\\|_F$, which for $d=2$ admits the closed form involving the entries of $Z_j^T Z_i$. Around this geometry, two mechanisms do the work: curve fitting by composite cubic Bézier curves (with a parameter $\\lambda$ trading data fidelity against acceleration) denoises the trajectories, and the Global Alignment Kernel—a positive-semidefinite kernel that sums over all possible temporal alignments—is applied to the resulting distance matrix to produce the PSD kernel $k(i,j) = \\frac{1}{2}e^{-D(i,j)/\\sigma^2}/(1-\\frac{1}{2}e^{-D(i,j)/\\sigma^2})$ that is fed to an SVM.","core_discovery":"The central claim is that the combination of three ingredients—the quotient geometry of $S_+(d,n)$ identified with $\\mathbb{R}^{n\\times d}_*/O_d$, composite cubic Bézier curve fitting, and the Global Alignment Kernel—turns Gram-matrix trajectories into a competitive action classifier with a theoretically valid PSD kernel. The paper's key identity is the closed-form Riemannian distance for $d=2$: $d(G_i,G_j)^2 = \\operatorname{tr}(G_i) - 2\\sqrt{(a+d)^2+(c-b)^2} + \\operatorname{tr}(G_j)$, where $Z_j^T Z_i = \\begin{pmatrix} a & b \\\\ c & d \\end{pmatrix}$, which arises from minimizing the Frobenius distance over the orthogonal group. The paper claims this PSD kernel supports a standard SVM, rather than the proximity SVM required by DTW, and reports accuracy gains on all three datasets, with the largest margin on UAV-Gesture.","pith_inferences":["If the PSD claim transfers, this recipe of quotient metric plus global alignment should apply to any landmark-trajectory task, such as facial expression or sign-language recognition, without retraining a metric.","The paper's own results hint that curve fitting helps on noisy 2D data and can slightly hurt on clean 3D data; an adaptive choice of $\\lambda$ per dataset might push accuracy higher.","A natural testable extension is to swap the halved Gaussian local kernel for other positive-semidefinite local kernels and check whether the Global Alignment Kernel remains PSD on the quotient manifold, isolating how much of the gain comes from the kernel choice.","For three-dimensional skeletons, an analogous closed-form distance derived from the $3\\times 3$ matrix $Z_j^T Z_i$ would likely speed up 3D pipelines the same way the 2D formula does."],"forward_implications":["On UTKinect-Action3D, the best configuration reaches 97.99% accuracy, above the 96.48% of the direct Gram-matrix predecessor [18].","On KTH-Action, using GAK instead of DTW improves accuracy from 94.49% to 96.16% in the same pipeline, and curve fitting adds about one point on this noisy 2D data.","On UAV-Gesture, the approach reaches 92.44%, outperforming the deep-learning P-CNN baseline of 91.9% while using only skeletons.","The closed-form distance for two-dimensional skeletons halves the alignment computation time relative to the general quotient distance, making the geometric pipeline practical on modest hardware.","Because the kernel is claimed positive semidefinite, the same pipeline can use a standard SVM, avoiding the pairwise-proximity SVM that DTW-based methods require."],"supporting_citations":[{"why":"Supplies the Gram-matrix trajectory representation and the DTW-based baseline whose accuracy this paper improves.","marker":"[18]"},{"why":"Defines the Global Alignment Kernel and the PSD property that the SVM classification step depends on.","marker":"[8]"},{"why":"Provides the composite cubic Bézier curve fitting algorithm used to denoise trajectories.","marker":"[13]"},{"why":"Gives the quotient geometry of fixed-rank PSD matrices, including the distance, exponential, and logarithm used here.","marker":"[28]"},{"why":"Describes the Bures-Wasserstein metric on the same manifold, the geometry the authors adopt for the quotient.","marker":"[30]"},{"why":"Supplies the numerical implementations of the Riemannian exponential and logarithm maps used in curve fitting.","marker":"[5]"},{"why":"Extracts the 2D body skeletons from KTH and UAV-Gesture videos that serve as input.","marker":"[6]"},{"why":"Provides the UAV-Gesture dataset and the deep P-CNN baseline that the proposed method outperforms.","marker":"[35]"}],"fun_headline_variants":["Skeleton action recognition via PSD manifold geometry hits 97.99% on UTKinect","Closed-form distance on PSD matrices enables SVM-based skeleton recognition","Bézier curve fitting on PSD manifold improves action recognition accuracy","Gram-matrix trajectories with quotient geometry outperform deep baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline's theoretical justification rests on the assumption that the Global Alignment Kernel built from the quotient-manifold distance is positive semidefinite; the paper cites this property from the Euclidean time-series setting [8] without proving it for this manifold distance.","fun_headline_variants_meta":{"raw":{"variants":["Skeleton action recognition via PSD manifold geometry hits 97.99% on UTKinect","Closed-form distance on PSD matrices enables SVM-based skeleton recognition","Bézier curve fitting on PSD manifold improves action recognition accuracy","Gram-matrix trajectories with quotient geometry outperform deep baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1397,"prompt_tokens":879,"completion_tokens":518,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":440}},"tokens_in":495,"tokens_out":518,"duration_ms":5642,"temperature":1.0,"reasoning_tokens":440,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:40:47.699115+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any dataset used in the paper, compute the full similarity matrix $K$ from the GAK scores, and test its smallest eigenvalue; if $K$ has a negative eigenvalue, the kernel is not PSD for this manifold distance and the standard SVM argument fails, even if classification accuracy stays high. A sharper test is to sample pairs of Gram matrices in $S_+(d,n)$, compute the local kernel matrix $k(i,j) = \\frac{1}{2}e^{-D(i,j)/\\sigma^2}/(1-\\frac{1}{2}e^{-D(i,j)/\\sigma^2})$, and check whether it is positive semidefinite for many random configurations.","supporting_citations":[{"cited_title":"Kacem, M","cited_arxiv_id":null,"evidence_quote":"Supplies the Gram-matrix trajectory representation and the DTW-based baseline whose accuracy this paper improves."},{"cited_title":"Cuturi, J","cited_arxiv_id":null,"evidence_quote":"Defines the Global Alignment Kernel and the PSD property that the SVM classification step depends on."},{"cited_title":"Gousenbourger, E","cited_arxiv_id":null,"evidence_quote":"Provides the composite cubic Bézier curve fitting algorithm used to denoise trajectories."},{"cited_title":"Massart and P.-A","cited_arxiv_id":null,"evidence_quote":"Gives the quotient geometry of fixed-rank PSD matrices, including the distance, exponential, and logarithm used here."},{"cited_title":"Massart, J","cited_arxiv_id":null,"evidence_quote":"Describes the Bures-Wasserstein metric on the same manifold, the geometry the authors adopt for the quotient."},{"cited_title":"Boumal, B","cited_arxiv_id":null,"evidence_quote":"Supplies the numerical implementations of the Riemannian exponential and logarithm maps used in curve fitting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the UAV-Gesture dataset and the deep P-CNN baseline that the proposed method outperforms."}],"review_version":1}