{"id":"27dd00bf-4c24-4af5-b7b1-9fc386665834","arxiv_id":"2501.11124","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"NoCo corrects noisy pseudo-labels for weakly supervised temporal action localization, reporting state-of-the-art accuracy on THUMOS14 and ActivityNet v1.2.","lead":"The authors propose NoCo, a video system that first guesses where actions happen in videos using only category labels, then automatically corrects three common types of mistakes in those guesses, and trains a faster detector on the corrected guesses. On the THUMOS14 and ActivityNet v1.2 benchmarks, it reports better localization accuracy and higher inference speed than prior weakly supervised methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central noise-correction claim lacks direct evidence that teacher predictions are more accurate than the pseudo-labels they correct; if teacher and pseudo-labels share the same noise, the reported gains may not result from the proposed correction mechanism.","rationale":"The reader's weakest assumption identifies exactly the point where the central argument is least secure: the teacher-student loop is trusted to correct the same kind of noise it was trained on, without independent evidence of superior label quality. My analysis adds a concrete internal inconsistency: the paper explicitly promises pseudo-label quality metrics (mIoU) but never reports them, so the mechanism is asserted rather than demonstrated. The ablations are internally consistent and show that each module contributes, which is genuine supporting evidence, and the generalization experiments across UM and CO2-Net strengthen the claim that the framework transfers. However, those ablations only show end-task mAP, which could improve because the student model is stronger or because extra training iterations help, not necessarily because the labels become cleaner. A direct quality check on the pseudo-labels would settle this. The speed claim is less load-bearing because it is not compared against the previous SOTA detector, only against the WTAL baseline, but that does not invalidate the accuracy claim. Overall, the paper merits a conditional accept rather than outright rejection, and the reader's conditional verdict should stand pending the proposed quality measurement.","tokens_in":13635,"tokens_out":5720,"duration_ms":56312,"concrete_test":"On THUMOS14, compute the mIoU and short-instance recall of (a) initial ASM-Loc pseudo-labels, (b) CALA-augmented pseudo-labels, and (c) the final online pseudo-labels after AIC/MIC updates, each against ground-truth instances, and separately for the three noise types (boundary error, missing short actions, merged adjacent actions). If (c) is not measurably closer to ground truth than (a), the proposed noise-correction mechanism is not the source of the reported improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"NoCo's performance claim rests on the assumption that the teacher model, pretrained on CALA-enhanced pseudo-labels, provides correction signals that are more trustworthy than the noisy pseudo-labels themselves. This is load-bearing because both AIC and MIC consume teacher predictions to reshape or add instances to the online pseudo-label set. If the teacher has learned the same boundary errors, missing short actions, and merged instances present in its training labels, then AIC merely averages noisy boundaries and MIC has no new short instances to propose. The paper states in the Implementation section that it reports mIoU for pseudo-label quality, but no such numbers appear in the main text; the ablation studies measure final mAP only, which conflates label-quality improvement with other effects such as the stronger TriDet student architecture. Without evidence that the corrected pseudo-labels are closer to ground truth than the initial pseudo-labels, the attribution of the 6.3% THUMOS14 (0.3:0.7) and 4.2% ActivityNet gains to noise correction is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NoCo, a weakly supervised temporal action localization method built on pseudo-label noise correction. It first generates instance-level pseudo-labels with a WTAL base model and refines their boundaries with a Context-Aware Label Augmentation (CALA) module, which aggregates neighboring same-class proposals using IoU- and confidence-based weights. It then trains a fully supervised student detector under an online teacher-student framework, where the teacher is an EMA of historical students and provides correction signals through two modules: Ambiguous Instance Correction (AIC), which refines existing pseudo-labels by aggregating high-IoU teacher predictions, and Missing Instance Compensation (MIC), which adds low-IoU teacher proposals as new instances. A High-quality Pseudo-label Mining (HPM) loss reweights positive samples during training. The method is evaluated on THUMOS14 and ActivityNet v1.2, reporting state-of-the-art average mAP of 50.9% and 30.7% respectively, and is shown to generalize when plugged into ASM-Loc, UM, CO2-Net, and two point-supervised baselines. The paper also claims a roughly threefold inference-speed improvement over its baseline.","tokens_in":13945,"tokens_out":5952,"duration_ms":59066,"significance":"If the results hold, the paper addresses a real and under-studied problem in pseudo-label-based WTAL: the noise in automatically generated instance labels. The design is modular and decoupled from the WTAL base model, and the generalization experiments across four base models strengthen the practical relevance. The two-stage pipeline with CALA, AIC, MIC, and HPM is coherent, and the ablations are internally consistent with the claimed contributions. The paper also reports a useful practical benefit: inference uses only the lightweight student detector. The main weaknesses are evidentiary: the central claim that gains come from noise correction is not directly supported by pseudo-label quality measurements, the speed comparison is not controlled across architectures, and the method's many thresholds are not subjected to sensitivity analysis or variance reporting. These issues leave the quantitative claims plausible but not fully established.","major_comments":[{"comment":"The Implementation section states that the authors 'report mIoU for pseudo-label quality,' but no such numbers appear in the main text or tables. The paper's central claim is that CALA, AIC, and MIC correct the three identified types of pseudo-label noise. Without direct measurements showing that the corrected pseudo-labels are closer to ground truth (e.g., mIoU before CALA, after CALA, and after AIC/MIC, and separately for boundary error, missing instances, and many-to-one cases), the attribution of the reported mAP gains to noise correction is not established. This is especially important because the teacher model is pretrained on the same CALA-enhanced pseudo-labels that AIC and MIC are supposed to correct; if the teacher inherits the same error patterns, correction could mainly average or duplicate existing noise. Please provide pseudo-label quality metrics and, if possible, a comparison between teacher prediction quality and initial pseudo-label quality.","section":"Implementation; Online Noise Correction Framework"},{"comment":"The speed comparison is not apples-to-apples. Table 5 compares a 'Baseline' at 35.8 mAP, 1.64 video/s, and 124.79 GFLOPs with 'NoCo' at 42.1 mAP, 6.67 video/s, and 20.12 GFLOPs. The baseline is the ASM-Loc-based pseudo-label retraining pipeline, while NoCo uses the lightweight TriDet student at inference. The 'threefold improvement' therefore conflates the effect of the proposed modules with the choice of a different, more efficient detector architecture. To support the speed claim, report inference speed of the same student detector with and without NoCo's correction modules, and also report speed for prior state-of-the-art methods under the same hardware and feature-extraction setup.","section":"Table 5; Analysis of Inference Speed"},{"comment":"The method introduces a large number of free hyperparameters: confidence threshold psi, NMS threshold rho, IoU thresholds eta0, eta1, and eta2, blending weight alpha, exponent beta, and positive loss weight lambda. No sensitivity analysis is reported for any of these, and no standard deviations over multiple seeds are provided for the main results or the ablation rows. Given that the state-of-the-art margins are small (e.g., 1.2 mAP over PivoTAL on THUMOS14 and 1.1 mAP over Ju et al. on ActivityNet v1.2), it is important to show that the conclusions are robust to threshold choices and random initialization. Please report at least three-seed mean and standard deviation for the main tables and a sensitivity study for the most critical thresholds (eta1, eta2, alpha).","section":"Experiments; Ablation Studies"},{"comment":"The main text repeatedly defers algorithmic details to the supplementary material: 'We refer to Supplementary for the workflow of AIC' and 'We refer to Supplementary for the main workflow of the MIC.' In the provided manuscript, no supplementary file is present. Equations (4)-(7) give the aggregation and weighting formulas, but the full algorithms, the order in which AIC and MIC update the online pseudo-label set, and the handling of weights across iterations are not completely specified. This is a reproducibility issue for a central component of the method; the workflows should be included in the main text or the supplementary material should be provided.","section":"Methods; Online Noise Correction Framework"}],"minor_comments":[{"comment":"There is a duplicated word in the introduction: 'perceives the contextual information for each pseudo-label and and corrects them' should read 'and corrects them.'","section":"Methods; Framework Overview"},{"comment":"The section heading 'Conslusion' is a typo and should be 'Conclusion.'","section":"Conclusion"},{"comment":"The 'Baseline' row in Table 3 deserves a clearer label. The value 35.8 is lower than ASM-Loc's own reported 45.1 mAP in Table 1, which may confuse readers. Please clarify that this row is the TriDet pseudo-label retraining baseline without any correction modules and explain why this pipeline underperforms the original WTAL model.","section":"Table 3; Ablation Studies"},{"comment":"The text says NoCo 'surpasses the latest methods CASE, AHIM, DDG-Net, and the current leading method PivoTAL,' but neither AHIM nor DDG-Net appears in Table 1 or in the reference list. Either add the corresponding rows and citations or remove the names.","section":"Results on THUMOS14"},{"comment":"The column header 'A VG' appears to be a formatting artifact for 'Avg.' The caption should spell out that this is the average mAP over the listed IoU thresholds, as the text in Section 5 does.","section":"Table 1; Table 2"},{"comment":"The abstract states that NoCo 'outperforms the previous state-of-the-art method in detection accuracy and inference speed greatly.' The accuracy margin over PivoTAL on THUMOS14 is 1.2 mAP, and the speed comparison is not controlled (see major comment on Table 5). Please temper the claim or provide additional supporting comparisons.","section":"Abstract; Results on THUMOS14"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the proposed framework is coherent, but the empirical support for the load-bearing 'noise correction' claim is incomplete. The absence of pseudo-label quality metrics, the uncontrolled speed comparison, and the lack of sensitivity/variance reporting are the main blockers. If the authors can supply the promised mIoU numbers and a controlled speed experiment, the paper would be much stronger. I would also encourage the authors to release code, since the method has many interacting thresholds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine, competent attempt to improve WTAL by treating pseudo-label noise as three specific failure modes, and the component-wise ablations hold together. It deserves refereeing, but the central attribution of the gains to 'noise correction' is under-supported because the paper never shows that the revised pseudo-labels are actually closer to ground truth.\n\nWhat's new: the paper names three concrete noise types—inaccurate boundaries, missing short actions, many-to-one merges—and builds a module for each: CALA for boundaries, MIC for missing instances, AIC for merges, plus an HPM loss to weight high-quality labels. The teacher-student online correction with EMA aggregation is a sensible way to get stable compensation signals. The strongest evidence is the generality: the framework improves not only ASM-Loc but also UM and CO2-Net on THUMOS14, and it transfers to point-supervised baselines like LACP and HR-Pro. Those numbers are consistent with the internal ablation, which adds each component and shows roughly monotonic gains.\n\nSoft spots: no error bars, no multiple seeds, and the AIC/MIC pseudocode and mIoU results are promised to a supplementary that isn't present. The speed claim only compares NoCo to the heavier ASM-Loc baseline, not to any other SOTA detector, so 'greatly faster' is an overstatement. The stress-test concern is fair: the teacher is trained on the same CALA-refined pseudo-labels, so its correction signals are not independent. That doesn't sink the method—AIC can still average out boundary jitter and MIC can add instances that the student's own predictions surface—but without mIoU numbers the reader can't tell if the labels are actually being corrected or if the gains come from the stronger TriDet head and extra training iterations. The ablation does control for the head, so that part is okay.\n\nWho it's for: anyone working on WTAL or pseudo-label-based detection. It's a buildable, practical contribution, not a paradigm shift. I'd send it to serious review conditional on the authors providing the supplementary, code or at least multiple-seed runs, and the promised mIoU comparison. If those come, this is a solid accept.","headline":"A solid, well-ablated WTAL method that reframes pseudo-label learning as noise correction, but the paper overclaims the mechanism without reporting pseudo-label quality metrics.","tokens_in":14402,"tokens_out":2670,"would_cite":false,"duration_ms":25976,"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":"NoCo corrects three structural types of pseudo-label noise—inaccurate boundaries, missing short clips, and merged adjacent instances—and sets new state-of-the-art results for weakly supervised temporal action localization, while running…","keywords":["weakly supervised temporal action localization","pseudo-label learning","noise correction","teacher-student framework","context-aware label augmentation","missing instance compensation","ambiguous instance correction","noisy label learning"],"falsifier":"On THUMOS14, where ground-truth annotations are available, compare per-instance IoU with ground truth for (a) the raw pseudo-labels, (b) the teacher's online predictions, and (c) the corrected labels after AIC and MIC: if the teacher's predictions are not systematically closer to ground truth than the pseudo-labels they correct—especially on the low-IoU instances MIC adds and the overlapping instances AIC adjusts—then the correction signal is not actually more trustworthy, and the reported gains would not be attributable to noise correction.","tokens_in":13486,"feed_emoji":"🎬","tokens_out":10973,"duration_ms":83767,"temperature":0.7,"pith_summary":"Pseudo-labels generated by weakly supervised temporal action localization (WTAL) models are noisy in three systematic ways: boundaries are imprecise, short action clips are missed, and adjacent instances get merged into one. The paper argues that this noise—not the weakly supervised model itself—is the main bottleneck when such pseudo-labels are used to train a fully supervised detection head, and that each noise type can be corrected without any extra supervision. To that end it proposes NoCo, a two-stage pipeline: a context-aware label augmentation step that sharpens boundaries using neighboring proposals, followed by an online teacher–student framework whose teacher (an EMA of past students) compensates missing instances, resolves merged instances, and reweights labels by quality. On THUMOS14 and ActivityNet v1.2 the paper reports average mAP of 50.9% and 30.7%, respectively, surpassing prior WTAL state-of-the-art methods, with a roughly four-fold inference speedup because only the lightweight student runs at test time.","feed_headline":"Noise correction lifts WTAL to 50.9% average mAP","feed_subtitle":"By fixing boundaries, missing clips, and merged instances, NoCo beats prior state of the art and runs faster.","key_machinery":"The load-bearing mechanism is the online teacher–student correction loop, in which the teacher is an exponential moving average of historical student checkpoints, giving it stable confidence and IoU signals from which to correct the current pseudo-label set. Within that loop, two modules consume teacher predictions: the Ambiguous Instance Correction module (AIC) forms, for each existing pseudo-label, an associated set of teacher proposals sharing the same category with IoU above a threshold, then fuses them via weighted averaging of start and end timestamps; the Missing Instance Compensation module (MIC) adds teacher proposals whose maximum IoU against all existing pseudo-labels falls below a threshold, with an adaptive confidence-based weight. The High-quality Pseudo-label Mining loss (HPM) then scales the classification and regression loss of each positive proposal by a weight derived from its IoU and confidence, so low-quality labels contribute less. Before the loop, the Context-Aware Label Augmentation module (CALA) replaces plain NMS with a weighted aggregation of neighboring same-class proposals sharing IoU above a threshold, using the weight $w_r = e^{\\sqrt{\\mathrm{IoU}} \\cdot \\min(\\max(\\theta_r,0),1)}$ and separate fusion functions for start and end timestamps, which the paper claims recovers boundary context that NMS discards.","core_discovery":"The central discovery is that pseudo-label noise in WTAL is not random but consists of three identifiable structural errors—inaccurate instance boundaries, missing short action clips, and adjacent instances collapsed into one—and each can be corrected by a dedicated module that uses teacher predictions as a second opinion. The Ambiguous Instance Correction module mines teacher proposals with high IoU against an existing pseudo-label and re-aggregates their start and end times to tighten boundaries; the Missing Instance Compensation module adds teacher proposals whose IoU with all existing pseudo-labels is low, recovering missed clips; and the High-quality Pseudo-label Mining loss weights each positive proposal by a quality score so clean labels dominate optimization. The paper demonstrates this on two benchmarks: applying NoCo to the ASM-Loc baseline raises average mAP from 45.1% to 50.9% on THUMOS14 and from 26.5% to 30.7% on ActivityNet v1.2, surpassing the previous state-of-the-art on both benchmarks, and the same wrapper improves other WTAL bases (UM, CO2-Net) and point-supervised methods (LACP, HR-Pro, BackTAL). Because the noise-correction modules sit outside the WTAL base model, inference needs only the TriDet student, giving a 6.67 video/s speed versus 1.64 video/s for the baseline.","pith_inferences":["The three noise types are generic to any weakly supervised proposal-generation pipeline, so the AIC/MIC/HPM recipe could plausibly transfer to weakly supervised object detection in video or spatio-temporal action localization; the paper does not test this.","Because the teacher aggregates historical student checkpoints, correction quality likely depends on how quickly the student's errors decorrelate across iterations, so the EMA update rate is a critical hyperparameter that may need per-dataset tuning beyond the settings reported.","The HPM loss reweights only positive foreground proposals; a natural untested extension would apply the same quality signal to background proposals, which might reduce false positives from context-heavy scenes.","If the teacher's predictions remain more accurate than pseudo-labels on out-of-distribution videos, NoCo could serve as a label-correction wrapper for deploying WTAL to new domains with only video-level labels, but the reported experiments are limited to two standard benchmarks."],"forward_implications":["NoCo sets a new state of the art on THUMOS14 and ActivityNet v1.2 for weakly supervised temporal action localization, with average mAP of 50.9% and 30.7%.","Because the noise-correction wrapper is decoupled from the WTAL base model, it can be grafted onto other weakly supervised methods; the paper shows consistent gains for UM, CO2-Net, ASM-Loc, and the point-supervised methods LACP, HR-Pro, and BackTAL.","Inference requires only the lightweight TriDet student, so NoCo runs at 6.67 video/s versus 1.64 video/s for the baseline WTAL method, with FLOPS reduced from 124.79G to 20.12G.","Each noise type has a dedicated module, and the ablations attribute specific gains to each: CALA adds 1.3% average mAP, the teacher–student scheme 0.7%, AIC 2.2%, MIC 1.8%, and HPM 0.2% on THUMOS14.","On ActivityNet v1.2 NoCo also outperforms some fully supervised methods such as SSN, suggesting that correcting pseudo-label noise can narrow part of the gap between weakly and fully supervised temporal action localization."],"supporting_citations":[{"why":"The default weakly supervised base model (ASM-Loc) that generates the noisy pseudo-labels and serves as the primary baseline for the reported gains.","marker":"(He et al. 2022)"},{"why":"TriDet, the efficient fully supervised detector used as the student model and the only model run at inference.","marker":"(Shi et al. 2023)"},{"why":"PivoTAL, the previous state-of-the-art on THUMOS14 that NoCo reports surpassing.","marker":"(Rizve et al. 2023)"},{"why":"The prior method whose ActivityNet v1.2 average mAP NoCo reports surpassing.","marker":"(Ju et al. 2023)"},{"why":"DivideMix, the noisy-label learning approach that motivates the paper's noise-correction strategy.","marker":"(Li, Socher, and Hoi 2020)"},{"why":"CO2-Net, one of the WTAL baselines used to demonstrate the generalizability of the NoCo framework.","marker":"(Hong et al. 2021)"},{"why":"UM, another WTAL baseline used in the generalizability analysis.","marker":"(Lee et al. 2021)"},{"why":"NMS, the standard proposal suppression method that CALA replaces because it discards contextual information.","marker":"(Neubeck and Van Gool 2006)"}],"fun_headline_variants":["Fix noise, not just labels: WTAL mAP jumps to 50.9%","Pseudo-label noise decoded: 3 errors, one fix, 50.9% mAP","WTAL gets faster and better: noise correction hits 50.9% mAP","Teaching the teacher: WTAL noise correction beats all prior","Unmasking pseudo-label noise for 50.9% average mAP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The correction loop assumes that a teacher model trained on the very same noisy pseudo-labels it is asked to fix produces confidence and IoU signals that are more trustworthy than those labels; if the teacher's errors correlate with the pseudo-label noise, the corrections could amplify the original mistakes rather than remove them.","fun_headline_variants_meta":{"raw":{"variants":["Fix noise, not just labels: WTAL mAP jumps to 50.9%","Pseudo-label noise decoded: 3 errors, one fix, 50.9% mAP","WTAL gets faster and better: noise correction hits 50.9% mAP","Teaching the teacher: WTAL noise correction beats all prior","Unmasking pseudo-label noise for 50.9% average mAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000346,"raw_usage":{"total_tokens":1956,"prompt_tokens":1061,"completion_tokens":895,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":677,"completion_tokens_details":{"reasoning_tokens":788}},"tokens_in":677,"tokens_out":895,"duration_ms":7569,"temperature":1.0,"reasoning_tokens":788,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:37:17.730365+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On THUMOS14, where ground-truth annotations are available, compare per-instance IoU with ground truth for (a) the raw pseudo-labels, (b) the teacher's online predictions, and (c) the corrected labels after AIC and MIC: if the teacher's predictions are not systematically closer to ground truth than the pseudo-labels they correct—especially on the low-IoU instances MIC adds and the overlapping instances AIC adjusts—then the correction signal is not actually more trustworthy, and the reported gains would not be attributable to noise correction.","supporting_citations":[{"cited_title":"N.; Mittal, G.; Yu, Y.; Hall, M.; Sajeev, S.; Shah, M.; and Chen, M","cited_arxiv_id":null,"evidence_quote":"PivoTAL, the previous state-of-the-art on THUMOS14 that NoCo reports surpassing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CO2-Net, one of the WTAL baselines used to demonstrate the generalizability of the NoCo framework."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"NMS, the standard proposal suppression method that CALA replaces because it discards contextual information."}],"review_version":1}