{"id":"31571db0-f459-4fb4-8f5c-866a489eb034","arxiv_id":"1908.04321","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Multi-timescale pose prediction, trained with supervision at intermediate layers, detects both short-term and long-term human anomalies and outperforms prior pose-based detectors on HR-ShanghaiTech and HR-Avenue.","lead":"This paper describes a surveillance-video model that predicts human body poses several steps into the future and the past at four different time horizons, then flags moments where the predictions fail as abnormal activity. It also introduces a large new corridor dataset and reports improved frame-level anomaly detection accuracy on two public benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline gain rests on a hand-picked, uniformly weighted timescale set; Table III shows the 25-step timescale hurts HR-Avenue, so the multi-timescale advantage may be configuration-specific rather than robust.","rationale":"The reader's conditional verdict and the identified weakest assumption are appropriate. My stress-test agrees that the most load-bearing point is Eq. (7)'s uniform average over the hand-picked set {3,5,13,25}, and Table III provides direct evidence that the marginal contribution of timescales is dataset-dependent. I do not see a more serious internal inconsistency that would invalidate the central idea: the architecture, the receptive-field mapping from layers to timescales, and the supervised training scheme are described plausibly, and the ablation table supports a real (if configuration-sensitive) multi-timescale effect. The absence of error bars and the post-hoc explanation for the Avenue decrease are serious but addressable weaknesses, consistent with a conditional acceptance. A concrete all-subset and weight-search experiment would settle whether the reported advantage is robust or an artifact of configuration choice. I therefore recommend leaving the reader's CONDITIONAL verdict unchanged.","tokens_in":8870,"tokens_out":6537,"duration_ms":62112,"concrete_test":"Re-run the Future+Past model on HR-ShanghaiTech and HR-Avenue for all non-empty subsets of the four supervised timescales, with at least five random seeds per configuration, and report mean±std Frame-AUC. Separately, learn per-timescale weights (e.g., logistic regression or a small validation set) on a held-out split of the training videos and apply the same weight estimator to both datasets. If the best subset or learned weights differ sharply across datasets, or if the uniform {3,5,13,25} configuration is not within one standard error of the best configuration, then the reported superiority is configuration-specific and the abstract's claim should be softened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that multi-timescale prediction is better than single-timescale prediction. Operationally, that claim is implemented by Eq. (7): the anomaly score averages L_j^2(t) uniformly over S={3,5,13,25}. This assumes every included timescale contributes equal discriminative information. Table III is the only quantitative evidence for the multi-timescale benefit, and it is not consistently supportive: on HR-Avenue, adding the 25-step timescale reduces the Future+Past AUC from 88.43 to 88.33, while on HR-ShanghaiTech it raises it from 75.65 to 77.04. The authors explain this as the dataset lacking long-term anomalies, but no per-anomaly or per-timescale analysis is provided to test that explanation. Moreover, the final configuration is selected after observing these test-set numbers; no validation-based model selection, non-uniform weighting, or error bars are reported. The claimed advantage over single-timescale baselines (72.05 to 77.04 on HR-ShanghaiTech; 84.99 to 88.33 on HR-Avenue) could therefore depend on the particular hand-picked set and the uniform average in Eq. (7), rather than on the multi-timescale principle itself. This is the load-bearing weakness: if the timescales or their weights were chosen differently, the headline result could shrink or reverse.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a multi-timescale framework for abnormal human activity detection in videos. The input is a human pose trajectory; two 1D-convolutional models predict future and past trajectories at four timescales (3, 5, 13, and 25 steps), with supervision at intermediate layers whose receptive fields match these horizons. At test time, prediction errors from all timescales and both directions are averaged (Eq. 7, with max over persons in Eq. 8) and thresholded to tag abnormal frames. The authors introduce a new Corridor dataset with 483,566 annotated frames and evaluate on HR-ShanghaiTech, HR-Avenue, Avenue, and ShanghaiTech. Table I reports improvements over prior methods on the HR variants, and Table III ablates the effect of adding timescales and the past-prediction branch.","tokens_in":9164,"tokens_out":6344,"duration_ms":60428,"significance":"If the results are reliable, the paper makes a useful contribution by showing that supervising intermediate convolutional layers at multiple temporal horizons yields prediction errors that are more discriminative for anomaly detection than a single timescale. The Corridor dataset is a substantial new resource with multi-person anomalies and class labels. The ablations in Table III provide direct evidence for the multi-timescale principle, and the architecture and loss are described in sufficient detail to be reimplemented. The main weaknesses are the lack of statistical validation and the hand-picked, uniformly weighted timescale set, which make the quantitative claims less robust than the text suggests.","major_comments":[{"comment":"The central claim rests on the specific timescale set S = {3, 5, 13, 25} and the uniform averaging in Eq. (7). Table III shows that adding the 25-step timescale reduces HR-Avenue AUC from 88.43 to 88.33, which the authors attribute to the dataset lacking long-term anomalies, but no per-anomaly or per-timescale analysis is provided to test that explanation. Because the final configuration appears to be selected after observing the test-set results, the reported gains could be configuration-specific. Please provide validation-based selection of S, a sensitivity analysis over S and the weights in Eq. (7), or a per-anomaly breakdown demonstrating that different timescales contribute to different anomaly types.","section":"V-E (Table III) and Eq. (7)"},{"comment":"All reported Frame-AUC values are point estimates from a single run. The margins between configurations are small (e.g., 75.65 vs 77.04 on HR-ShanghaiTech; 88.43 vs 88.33 on HR-Avenue) and may be within run-to-run variance for a deep model trained with this protocol. Please report the mean and standard deviation over multiple training runs, or confidence intervals via bootstrap over the test set, and state whether the differences versus the baselines and ablations are statistically significant.","section":"V-D (Tables I and III)"}],"minor_comments":[{"comment":"The abstract claims the model 'outperforms existing methods' without qualification, but Table I shows lower Frame-AUC on the full Avenue dataset (82.85) than Liu et al. (84.90). The body qualifies the claim to HR-ShanghaiTech and HR-Avenue; the abstract should be corrected to avoid overclaim.","section":"Abstract and Section V-D"},{"comment":"The definition of L_2^j in Eq. (3) is inconsistent with Eqs. (1) and (7): as written it defines a per-time average error, yet Eq. (1) uses it as a sum over t and Eq. (7) uses L_2^j(t). Please clarify the notation, e.g., define L_2^j(t) explicitly and write the layer-level loss as a sum over t.","section":"III-C (Eq. 3)"},{"comment":"The sentence 'The loss after an epoch is equal to the loss incurred at the last sub-epoch' is ambiguous about whether gradients from earlier sub-epochs are retained. Please specify the optimizer update rule and how the multi-timescale loss in Eq. (4) is used during training.","section":"V-B"},{"comment":"The dataset description gives the total number of frames but not the distribution of abnormal frames across activity classes; a per-class frame count would help readers assess the difficulty and potential biases.","section":"IV (Corridor dataset)"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is sound and the dataset is a valuable contribution, but the quantitative evidence needs strengthening. The absence of error bars and the post-hoc selection of the timescale set are the main risks. If the authors can provide repeated-run statistics and a validation-based or sensitivity analysis for the timescale set, the paper would likely meet the bar for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something straightforward and useful: it takes Morais et al.'s pose-trajectory prediction approach and gets a single 1D convolutional stack to predict at several timescales by supervising intermediate layers whose receptive fields are 3, 5, 13, and 25 steps. That is a real architectural idea, and the ablations in Table III do show that adding timescales helps on HR-ShanghaiTech (72.05 to 77.04 AUC for Future+Past). The new Corridor dataset is the other solid contribution: 483k frames, larger than the usual single-camera benchmarks, and it includes group-level anomalies (protest, chasing, fighting) that other datasets lack. I would want the dataset released, but the scale and activity variety are valuable on their own.\n\nThe soft spots are about robustness, not about the core idea. The stress-test note is on target: Eq. (7) averages prediction errors uniformly over the fixed, hand-picked set {3,5,13,25}. Table III is the only evidence for the multi-timescale benefit, and it is not monotonic—on HR-Avenue, adding the 25-step timescale drops AUC from 88.43 to 88.33. The authors blame the absence of long-term anomalies in Avenue, but they provide no per-timescale or per-anomaly analysis to verify that explanation. They also appear to have selected the final configuration after looking at test-set numbers, with no validation split, no error bars, and no significance tests. All results are single runs. The abstract's claim that the model \"outperforms existing methods\" is too strong: on the full Avenue benchmark they are at 82.85, below Liu et al.'s 84.90. The weaknesses are addressable—error bars, a validation-based timescale choice or learned weights, a softened abstract, and released code and data—but as written the paper's central claim depends on choices that could reasonably shift the outcome.\n\nWhat is genuinely good: the intermediate-supervision mechanism is elegant, the loss design with node and layer terms is sensible, and the paper is clearly written. The citation pattern looks fine; the work builds directly on Morais et al. and says so. I would not desk-reject this. It deserves a serious referee and probably a conditional accept after experimental rigor is tightened. The multi-timescale principle is plausible, and the dataset alone justifies engaging with the paper.\n\nFor peer review: send it out, but the reviewers should push on the timescale selection, the uniform weighting, and the reproducibility of the numbers.","headline":"A clean multi-timescale extension of pose-based anomaly detection with a genuinely useful new dataset, but the headline gain rests on a hand-picked uniform timescale average and single-run numbers.","tokens_in":9660,"tokens_out":1013,"would_cite":false,"duration_ms":11790,"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 claims that detecting abnormal human activities requires predicting pose trajectories at several timescales at once, and that a multi-layered 1D convolutional network supervised at timescales 3, 5, 13, and 25 outperforms…","keywords":["abnormal activity detection","trajectory prediction","multi-timescale","pose trajectories","surveillance video","anomaly detection","convolutional neural network","Corridor dataset"],"falsifier":"A single experiment would settle it: train the same architecture with only the 25-step supervision and only the 3-step supervision, and compare both against the combined model on the Corridor dataset's loitering and sudden-running sequences. If the combined model does not beat the best single timescale on both anomaly types, the multi-timescale claim fails.","tokens_in":8662,"feed_emoji":"⏱","tokens_out":6942,"duration_ms":60114,"temperature":0.7,"pith_summary":"The paper argues that abnormal human activities have no single temporal scale: a jump is a short perturbation, while loitering becomes abnormal only after many frames. A detector trained at one fixed timescale will therefore miss anomalies whose duration it never saw. To fix this, the authors build a model that predicts past and future pose trajectories at four timescales simultaneously, with intermediate network layers supervised to predict 3, 5, 13, and 25 steps ahead. The prediction errors from all timescales are averaged into an anomaly score per frame. A sympathetic reader would take away that multi-timescale supervision is a simple, label-free way to make video anomaly detection cover both brief and prolonged human behaviors.","feed_headline":"Predicting poses at four timescales catches short and long anomalies","feed_subtitle":"Supervising a pose CNN at 3, 5, 13, and 25 steps lifts Frame-AUC to 77.04 on HR-ShanghaiTech.","key_machinery":"The load-bearing mechanism is a multi-layered 1D convolutional encoder-decoder in which intermediate layers are supervised, through node-level and layer-level losses, to predict trajectory offsets at chosen time horizons. Layer 1 has a receptive field of 3 frames and predicts 3 steps ahead; layers 2, 4, and 7 correspond to horizons 5, 13, and 25. A sliding window over the input produces several predictions per time instant, and these are averaged into layer losses, so training mimics the sliding-window test procedure. Past predictions come from running the same architecture on the reversed trajectory. The final anomaly score at time $t$ is the equal-weight average over timescales of the per-layer losses, taking the maximum over all people in the frame.","core_discovery":"The central claim is that prediction error signals abnormality only at the right timescale, so a model should predict at many timescales and combine the errors. The authors train a seven-layer 1D convolutional network on human pose trajectories, supervising layers 1, 2, 4, and 7 to predict 3, 5, 13, and 25 steps into the future, and a twin model that makes the same predictions into the past. At test time, per-timescale errors are averaged to score each frame, and frames above a threshold are tagged abnormal. On HR-ShanghaiTech the full model reaches 77.04 Frame-AUC and on HR-Avenue 88.33, exceeding the compared single-timescale baselines; the ablation shows AUC climbing from 72.05 with only 3-step prediction to 77.04 with all four timescales. The paper also introduces the Corridor dataset, 483,566 annotated frames of staged single- and group-level anomalies.","pith_inferences":["The equal-weight voting rule is the most brittle part of the design; a learned weighting of the four timescales would probably improve datasets with mixed anomaly durations, since Table III already shows the 25-step timescale hurting HR-Avenue.","The timescale set {3, 5, 13, 25} is fixed by the network's receptive fields rather than by anomaly statistics; selecting horizons from the training data could reveal whether an intermediate duration such as 9 steps matters.","The same intermediate-supervision trick could be applied to other sequential abnormality signals, such as optical flow or object trajectories, not only human skeletons.","The paper reports only Frame-AUC, so per-anomaly recall by duration is untested; that measurement would directly validate or undermine the multi-timescale explanation."],"forward_implications":["A fixed-horizon predictor can be upgraded to a multi-timescale one by adding supervision at intermediate layers, without any new labels.","On HR-ShanghaiTech, Frame-AUC rises monotonically as timescales 3, 5, 13, and 25 are added, from 72.05 to 77.04, so long-horizon errors carry information that short-horizon errors miss.","Combining past and future predictions improves over either direction alone in most tested configurations.","Pose trajectories alone are sufficient to detect human anomalies; the max-over-people aggregation makes the method applicable to group scenes.","The chosen timescales must match the anomalies present: on HR-Avenue, adding the 25-step timescale slightly reduces AUC because that dataset lacks long-term anomalies."],"supporting_citations":[{"why":"Establishes the HR-ShanghaiTech and HR-Avenue pose-trajectory benchmarks and the reconstruction-prediction baseline the paper must beat.","marker":"[13]"},{"why":"Supplies the future-frame prediction baseline and the prediction-error paradigm for Frame-AUC comparison.","marker":"[9]"},{"why":"Provides the classical autoencoder temporal-regularity baseline compared in Table I.","marker":"[5]"},{"why":"Introduces the ShanghaiTech dataset and the sparse-coding/RNN baseline that defines the comparison setting.","marker":"[12]"},{"why":"Supplies the unmasking training-free baseline used as a comparison on Avenue.","marker":"[17]"},{"why":"Is the latent-space autoregression baseline compared on ShanghaiTech.","marker":"[1]"},{"why":"Supplies OpenPose, the pose detector that produces the 25-point trajectories and the confidence weights used in the weighted MSE loss.","marker":"[3]"},{"why":"Defines the Avenue dataset used for HR-Avenue evaluation.","marker":"[11]"}],"fun_headline_variants":["Multi-timescale pose prediction spots short and long anomalies","Four timescales beat one for abnormal activity detection","Pose CNN with four timescales lifts anomaly detection AUC","Predicting poses at multiple timescales improves anomaly detection","Multi-scale trajectory prediction catches anomalies of any duration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assumption that the four hand-picked timescales {3, 5, 13, 25} are the right ones and that averaging their prediction errors with equal weight is a good anomaly score.","fun_headline_variants_meta":{"raw":{"variants":["Multi-timescale pose prediction spots short and long anomalies","Four timescales beat one for abnormal activity detection","Pose CNN with four timescales lifts anomaly detection AUC","Predicting poses at multiple timescales improves anomaly detection","Multi-scale trajectory prediction catches anomalies of any duration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00043,"raw_usage":{"total_tokens":2230,"prompt_tokens":1011,"completion_tokens":1219,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":1141}},"tokens_in":627,"tokens_out":1219,"duration_ms":8043,"temperature":1.0,"reasoning_tokens":1141,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:45:38.968626+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A single experiment would settle it: train the same architecture with only the 25-step supervision and only the 3-step supervision, and compare both against the combined model on the Corridor dataset's loitering and sudden-running sequences. If the combined model does not beat the best single timescale on both anomaly types, the multi-timescale claim fails.","supporting_citations":[{"cited_title":"Morais, V","cited_arxiv_id":null,"evidence_quote":"Establishes the HR-ShanghaiTech and HR-Avenue pose-trajectory benchmarks and the reconstruction-prediction baseline the paper must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the future-frame prediction baseline and the prediction-error paradigm for Frame-AUC comparison."},{"cited_title":"Hasan, J","cited_arxiv_id":null,"evidence_quote":"Provides the classical autoencoder temporal-regularity baseline compared in Table I."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the ShanghaiTech dataset and the sparse-coding/RNN baseline that defines the comparison setting."},{"cited_title":"Tudor Ionescu, S","cited_arxiv_id":null,"evidence_quote":"Supplies the unmasking training-free baseline used as a comparison on Avenue."},{"cited_title":"Abati, A","cited_arxiv_id":null,"evidence_quote":"Is the latent-space autoregression baseline compared on ShanghaiTech."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Avenue dataset used for HR-Avenue evaluation."}],"review_version":1}