{"id":"949e719b-4062-4d1e-98cc-0e26deca21db","arxiv_id":"2508.17726","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A single contrastive action encoder with diffusion-based motion augmentation performs few-shot human action anomaly detection on HumanAct12 with an average AUC of 86.9%, including three categories excluded from encoder training.","lead":"This paper trains a single contrastive encoder on normal human action examples and detects anomalies by comparing each new motion against a small support set of normal samples, including categories never seen during encoder training. It reports an average AUC of 86.9% on HumanAct12, but the evaluation is weakened by a generative model pretrained on the same dataset and by baselines that do not receive a support set.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unseen-category claim is undermined: HumanMAC, used to generate support samples for Phone/Boxing/Throw, was pretrained on the full HumanAct12 dataset, so those categories are not novel to the pipeline; a retraining test is required.","rationale":"The reader's weakest assumption correctly identifies the single most load-bearing gap. The paper's headline novelty is that a single encoder generalizes to action categories never seen anywhere in the pipeline. But Section 4.1 explicitly says HumanMAC is pretrained on HumanAct12, which includes the three 'unseen' categories, and Algorithm 2 uses HumanMAC to generate support samples for exactly those categories. Thus the evaluation does not establish novelty to the pipeline. My recommended verdict is CONDITIONAL rather than outright REJECT because the concern is empirically decidable: a retraining experiment with a leakage-free HumanMAC would either confirm or refute it. The architecture is clearly described, and the core components are logically consistent. The unmatched baseline protocol is a secondary concern: comparing against per-category models that are not few-shot makes the SOTA claim hard to interpret, but the more decisive issue is the leakage into the 'unseen' rows. If the retraining test shows no degradation, the paper's claims would be substantially rehabilitated; if it shows degradation, the central claim as stated should be rejected. I therefore set verdict_should_be to CONDITIONAL and agree with the reader on the weakest assumption.","tokens_in":13501,"tokens_out":8206,"duration_ms":90204,"concrete_test":"Retrain HumanMAC from scratch on HumanAct12 with all sequences from Phone, Boxing, and Throw removed, keeping the same 1000-epoch DDIM training configuration and the same 20 DCT bases. Then retrain the action encoder on the remaining nine categories and rerun the full few-shot evaluation of Table 2, with both training-time and inference-time augmentation using this leakage-free HumanMAC. Report mean AUC for the three held-out categories separately. If their average drops materially (e.g., by more than 2-3 AUC points) relative to Table 2, the reported unseen-category gains are inflated by pretraining leakage; if the drop is within seed noise, the current claim can stand. A useful secondary check is to run the same three categories with support-set augmentation disabled to isolate the effect of generated samples.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim of generalization to unseen categories rests on a training/pretraining separation that the paper does not actually enforce. Section 4.1 states that three categories (Phone, Boxing, and Throw) are 'excluded during training and used only for the testing phase,' but it also states that HumanMAC, the generative model, is 'pretrained on HumanAct12 for 1000 epochs.' HumanAct12 contains Phone, Boxing, and Throw. At inference, Algorithm 2 feeds every support sample through HumanMAC to produce augmented support samples (lines 2-7), so the few-shot normal reference set for each 'unseen' category is partly built from a model that has already learned the statistics of that exact category during pretraining. The reported unseen-category AUCs (Phone 0.921, Boxing 0.883, Throw 0.824) can therefore be explained by HumanMAC having memorized category-specific motion patterns rather than by the contrastive encoder's category-agnostic generalization. The ablation in Table 3 only reports mean AUC over all 12 categories; it does not decompose seen versus unseen categories, so it cannot rule out a leakage effect concentrated in the three rows where the novelty claim is made. Because contribution 3 and the 'unseen' rows of Table 2 are the paper's core evidence, this gap is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified, category-agnostic framework for few-shot human action anomaly detection (HAAD). The method uses a Res-GCN encoder trained with contrastive learning on multiple seen categories, and a pretrained diffusion-based generative model (HumanMAC) to augment both the training data and the support set at inference. Anomaly scores are computed by average Euclidean distance between the test embedding and the augmented support-set embeddings. Experiments on HumanAct12 report an average AUC of 0.869, claimed to outperform prior one-model-per-category baselines, including on three action categories that are stated to be 'unseen' during training (Phone, Boxing, Throw). The paper also includes ablations of the generative augmentation and parameter sensitivity analyses.","tokens_in":13800,"tokens_out":6161,"duration_ms":65941,"significance":"If the central claims were established, the paper would represent a meaningful step toward scalable, few-shot HAAD: a single encoder that handles multiple categories and novel categories without retraining is a practically valuable capability, and the use of generative motion completion as a contrastive augmentation is a sensible idea. The paper is clearly written, reports mean and standard deviation over multiple seeds, and includes both quantitative and qualitative analyses. However, the key evidence for generalization to unseen categories is compromised by the fact that the generative model used for augmentation was pretrained on the full evaluation dataset, and there are ambiguities about hyperparameter selection. These issues directly affect the significance of the contributions.","major_comments":[{"comment":"The claim of generalization to unseen categories is circular. Section 4.1 states that Phone, Boxing, and Throw are 'excluded during training and used only for the testing phase' for the action encoder, but it also states that the motion generation module HumanMAC is 'pretrained on HumanAct12 for 1000 epochs.' Since HumanAct12 includes Phone, Boxing, and Throw, HumanMAC has seen these categories during its pretraining. At inference, Algorithm 2 applies HumanMAC to every support sample (lines 2–7), so the few-shot reference set for each 'unseen' category is augmented with plausible continuations that carry category-specific statistics learned from those exact categories. The reported unseen-category AUCs in Table 2 (Phone 0.921, Boxing 0.883, Throw 0.824) can therefore be explained by HumanMAC's prior knowledge rather than by the contrastive encoder's category-agnostic generalization. A valid evaluation would require HumanMAC to be pretrained without the held-out categories (e.g., on the 9 training categories only) or on a completely different dataset, and the paper should report seen-only and unseen-only results separately. This issue is load-bearing for Contribution 3 and for the abstract's claim of novelty in unseen-category settings.","section":"Section 4.4 and Table 2"},{"comment":"The support set size Ns and the number of generated samples Ng appear to be selected using the test set. Figure 5 shows how performance varies with Ns and Ng and marks the best-performing configuration with a star. The paper does not specify whether a validation set was used to choose these hyperparameters, nor does it state the exact Ns and Ng used for the main results in Table 2. If Ns (and possibly Ng) are chosen by maximizing AUC on the same test set used for final evaluation, the reported mean AUC of 0.869 is an optimistically biased estimate and the comparisons in Table 2 are not reliable. The authors need to clearly predefine the evaluation protocol, report the specific Ns and Ng used in Table 2, and justify that no test-set information is used for configuration selection.","section":"Section 4.4 and Table 2"},{"comment":"The comparison with baselines is not protocol-matched. STG-NF, MoCoDAD, and MultiLevel-NF are trained per-category on the full normal training data, while the proposed method is trained on a subset of categories and evaluated with a few-shot support set. Although one can argue that the proposed method faces a harder setting, this is not an apples-to-apples comparison, and calling the method 'state-of-the-art' based on a 2.6% improvement over MultiLevel-NF is not fully justified. At minimum, the paper should clarify that the baselines use full-shot data while the proposed method uses few-shot data, and ideally it should include few-shot baselines (e.g., k-NN with the same support set, or a backbone trained on seen categories only) to enable a controlled comparison. This affects the interpretation of the central quantitative claim.","section":"Tables 1–2"}],"minor_comments":[{"comment":"The notation in Eq. (5) is ambiguous: the expression appears to write \\(\\ell(i) = P_{j \\in P(i)} -\\log\\cdots\\), which should be a sum over the positive set \\(P(i)\\). Please write it explicitly as \\(\\sum_{j \\in P(i)} -\\log\\frac{...}{...}\\).","section":"Section 3.3"},{"comment":"The ablation study reports only the mean AUC over all 12 categories. Since the unseen-category claim is central, the table should also show average AUC split into seen and unseen categories to demonstrate that the benefit of generative augmentation is not driven solely by leakage.","section":"Table 3"},{"comment":"The qualitative examples in Fig. 3 would be more informative if the number of support samples and generated samples used in the visualization were stated, as the score distributions depend on the support set composition.","section":"Section 4.2"},{"comment":"Minor language issues: 'The number of generations Ng do not impose noticeable influence' should be 'does not impose'; the sentence would benefit from revision for clarity.","section":"Section 4.4"}],"recommendation":"reject","confidential_remarks":"The circularity issue raised in the review is likely decisive. The paper's main empirical contribution—generalization to unseen action categories—is not supported because the generative model used for augmentation was pretrained on the full HumanAct12 dataset, including the 'unseen' categories. This is not a minor fix; it requires re-running the experiments with a generative model that has no exposure to the held-out categories. Additionally, the test-set selection of support-set hyperparameters, if confirmed, would further invalidate the reported performance. The paper has a plausible idea and a reasonable framework, but the current evidence does not substantiate the claims, and the necessary re-evaluation is beyond the scope of a minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the paper has a real idea — a single contrastive encoder for few-shot skeleton-based HAAD with diffusion-based motion augmentation — but the main evidence for the 'unseen category' claim is undermined by the generative model having been pretrained on the full dataset, including those supposedly unseen categories, so the claim as stated does not hold.\n\nWhat's actually new: the combination of support-set contrastive AD and HumanMAC-style motion augmentation is new for skeleton-based HAAD. The method is described clearly and the components are sensible. The ablation shows training-time augmentation gives a substantial boost (0.760 to 0.866), so the core mechanism seems to be doing something real. The parameter sensitivity analysis is more thorough than many papers in this area.\n\nWhere the soft spots are: the load-bearing flaw is the unseen-category evaluation. HumanMAC is pretrained on HumanAct12 for 1000 epochs (Section 4.1) and then used at inference to generate support samples for Phone, Boxing, and Throw—the three categories the paper treats as novel. So those categories are not novel to the full pipeline. The encoder never sees them, but the reference distribution is built by a model that has seen them. The reported 0.921/0.883/0.824 AUCs for those categories could reflect HumanMAC's category statistics rather than encoder generalization. The fix is straightforward: retrain HumanMAC with those categories held out, or use a separate motion dataset, and report seen/unseen AUC separately in the ablation. Not doing this is a genuine gap, not a nitpick.\n\nSecond, the comparison in Table 2 is not matched. The baselines are one-model-per-category, trained on full normal data; the proposed method uses a few-shot support set. The proposed setting may be harder, but that also means the 2.6% gain over MultiLevel-NF is not a fair state-of-the-art claim under the same protocol. A support-set version of a baseline, or at least a clear statement that the comparison is across settings, is needed.\n\nMinor: only one small dataset (HumanAct12, 1191 clips), no cross-dataset test, and no code released. The t-SNE and qualitative examples are illustrative rather than evidence. The related work and citations look fine.\n\nRecommendation: I would send this to a serious referee. The flaw is real but fixable, and the idea is worth discussing. The authors need to address the leakage issue and the unmatched comparison before the paper's claims can stand. It's a borderline reject for the current version, but not a non-starter.","headline":"A real but untested idea: the few-shot HAAD contrastive framework is sensible and well-described, but the unseen-category evidence leaks through HumanMAC's pretraining on the same dataset, so the headline claim needs rework.","tokens_in":14324,"tokens_out":3420,"would_cite":false,"duration_ms":34333,"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 single contrastively trained encoder detects anomalies in both seen and unseen human actions from a few normal examples, reaching 86.9% average AUC on HumanAct12.","keywords":["few-shot anomaly detection","human action anomaly detection","contrastive learning","generative motion augmentation","diffusion model","skeleton-based action recognition","HumanAct12","category-agnostic representation"],"falsifier":"Retrain the generative augmentation model on only the nine training categories, then rerun the few-shot evaluation on Phone, Boxing, and Throw; if their AUC drops materially from the reported 0.921, 0.883, and 0.824, the pretrained generator had leaked category-specific statistics into the supposedly unseen evaluation.","tokens_in":13292,"feed_emoji":"🕵️","tokens_out":7744,"duration_ms":74680,"temperature":0.7,"pith_summary":"This paper tries to establish that human action anomaly detection can be done by a single, category-agnostic encoder trained once on several normal action categories, and then applied to any category — seen or unseen — using only a small support set of normal examples. Its aim is to replace the one-model-per-category paradigm, where each normal action needs its own model and large amounts of data, with a few-shot solution that scales to new or rare actions. The paper argues that a contrastive learning objective creates a shared representation space in which normal and anomalous actions are separable by a simple distance to the support set, and that augmenting training and support samples with a diffusion-based motion generator materially improves that separability. If the claim holds, anomaly detection for human actions becomes practical in data-scarce settings where novel categories appear frequently, because a new category requires only a handful of normal examples rather than a retrained model. The reported evidence is an average AUC of 86.9% on HumanAct12, 2.6 points above the prior per-category state of the art.","feed_headline":"One model flags unseen action anomalies from a few examples","feed_subtitle":"A shared contrastive space plus diffusion-generated motion samples lifts average AUC to 86.9% on HumanAct12.","key_machinery":"The load-bearing mechanism is a contrastively trained action encoder paired with a frozen generative motion augmentor. Input motion sequences are transformed with the discrete cosine transform to keep low-frequency components, then passed through a residual graph convolutional network (Res-GCN) that outputs a compact vector per sequence. Training uses a SimCLR-style contrastive loss extended so that each anchor has multiple positive counterparts: the real samples of its category plus the generated variants produced by the generative model. The generative model — a diffusion model that frames motion prediction as masked completion in the frequency domain — observes the first 30 frames of a 60-frame sequence and generates the remaining 30 frames, producing semantically consistent but diverse samples that serve as extra positives in training and as expansions of the few-shot support set at inference. Anomaly scoring is a parameter-free step: the mean Euclidean distance from the test embedding to all support embeddings, with larger distances indicating anomalies.","core_discovery":"The central claim is that few-shot human action anomaly detection, including for action categories never seen during training, can be solved by one unified contrastive model rather than by per-category anomaly detectors. The model encodes motions into a shared space in which same-category actions cluster, and then scores a test sample by its average Euclidean distance to a small support set of normal samples from the target category at inference. The paper further claims that using a pretrained diffusion-based motion-completion model to generate additional plausible continuations of each sequence — both during contrastive training and when expanding the support set — increases inter-category discrimination and intra-category robustness. On the HumanAct12 dataset this yields a mean AUC of 0.869 across twelve categories, including three categories (Phone, Boxing, Throw) excluded from encoder training, which is 2.6 percentage points higher than the previous state-of-the-art per-category method.","pith_inferences":["If the category-agnostic space transfers as claimed, a single frozen encoder and generator could be evaluated on completely different skeleton datasets or on held-out action families (for example, excluding all part-body actions), which would test the boundary of 'unseen' generalization more sharply than leaving out three individual categories.","The ablation shows that training-time augmentation matters most, which raises the possibility that a simpler hand-designed skeleton augmentation — temporal cropping, joint masking, or speed perturbation — might capture a large share of the benefit at lower compute; the paper does not test this.","A testable deployment corollary is that the same pretrained encoder could be shipped as a common backend, with each new site or market adding only a support set of locally normal actions, turning few-shot human action anomaly detection into a configuration step rather than a training step."],"forward_implications":["A deployment can cover a new action category by collecting only a few normal clips and running them through the frozen encoder and generator; no retraining or per-category model is needed.","The same encoder can detect anomalies in categories it never saw during training as long as the support set defines the normal class, as shown for Phone, Boxing, and Throw.","Generative augmentation at training time is the main source of the accuracy gain, while augmenting the support set at inference mainly stabilizes results across random support-set choices.","Increasing the number of generated samples per real sample improves accuracy up to about 3–5 generations and then degrades it, so the benefit is a balance between diversity and redundancy rather than simply more data.","Distance-based scoring in the learned space removes the need for a per-category density model or reconstruction network, simplifying the inference pipeline."],"supporting_citations":[{"why":"defines the frequency-guided multi-level HAAD baseline that the proposed method compares against and reports a 2.6-point improvement over.","marker":"[13]"},{"why":"supplies the per-category reconstruction/diffusion baseline that motivates the unified few-shot design.","marker":"[14]"},{"why":"supplies the per-category normalizing-flow baseline used to show the limitations of one-model-per-category methods.","marker":"[15]"},{"why":"introduces the few-shot unified anomaly detection idea of matching test samples to a support set, which the paper adapts to skeleton actions.","marker":"[17]"},{"why":"provides the SimCLR contrastive learning objective that the paper extends to multiple positive counterparts per anchor.","marker":"[25]"},{"why":"is the pretrained diffusion-based motion-completion model used for training-time and inference-time generative augmentation.","marker":"[28]"},{"why":"is the HumanAct12 dataset on which the encoder is trained and the few-shot evaluation is run.","marker":"[29]"},{"why":"supplies the DCT-based trajectory encoding and residual graph convolutional network architecture used for the action encoder.","marker":"[45]"}],"fun_headline_variants":["One model detects anomalies in unseen actions with few examples","Unified contrastive learning enables few-shot action anomaly detection","Diffusion-based augmentation boosts few-shot anomaly detection","Category-agnostic model spots anomalies from tiny support sets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire claim depends on the held-out action categories being unseen by every component of the pipeline; in particular, the pretrained motion generator used to create support and training samples was itself trained on the same dataset that contains those categories, so if it has memorized them, the reported 'unseen' results are not a test of genuine novelty.","fun_headline_variants_meta":{"raw":{"variants":["One model detects anomalies in unseen actions with few examples","Unified contrastive learning enables few-shot action anomaly detection","Diffusion-based augmentation boosts few-shot anomaly detection","Category-agnostic model spots anomalies from tiny support sets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000301,"raw_usage":{"total_tokens":1729,"prompt_tokens":931,"completion_tokens":798,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":734}},"tokens_in":547,"tokens_out":798,"duration_ms":7970,"temperature":1.0,"reasoning_tokens":734,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:00:49.238705+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the generative augmentation model on only the nine training categories, then rerun the few-shot evaluation on Phone, Boxing, and Throw; if their AUC drops materially from the reported 0.921, 0.883, and 0.824, the pretrained generator had leaked category-specific statistics into the supposedly unseen evaluation.","supporting_citations":[{"cited_title":"Frequency-Guided Multi-Level Human Action Anomaly Detection with Normalizing Flows","cited_arxiv_id":"2404.17381","evidence_quote":"defines the frequency-guided multi-level HAAD baseline that the proposed method compares against and reports a 2.6-point improvement over."},{"cited_title":"Multimodal motion con- ditioned diffusion model for skeleton-based video anomaly detection","cited_arxiv_id":null,"evidence_quote":"supplies the per-category reconstruction/diffusion baseline that motivates the unified few-shot design."},{"cited_title":"Normalizing flows for human pose anomaly detection","cited_arxiv_id":null,"evidence_quote":"supplies the per-category normalizing-flow baseline used to show the limitations of one-model-per-category methods."},{"cited_title":"Registration based few-shot anomaly detection","cited_arxiv_id":null,"evidence_quote":"introduces the few-shot unified anomaly detection idea of matching test samples to a support set, which the paper adapts to skeleton actions."},{"cited_title":"Humanmac: Masked motion completion for human motion prediction","cited_arxiv_id":null,"evidence_quote":"is the pretrained diffusion-based motion-completion model used for training-time and inference-time generative augmentation."},{"cited_title":"Ac- tion2motion: Conditioned generation of 3d human motions","cited_arxiv_id":null,"evidence_quote":"is the HumanAct12 dataset on which the encoder is trained and the few-shot evaluation is run."},{"cited_title":"Learning trajectory dependencies for human motion pre- diction","cited_arxiv_id":null,"evidence_quote":"supplies the DCT-based trajectory encoding and residual graph convolutional network architecture used for the action encoder."}],"review_version":2}