{"id":"b684b209-11ff-4630-8f14-32030119f496","arxiv_id":"1908.02422","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ASSG, a seed-growing framework with an erasing classifier, achieves state-of-the-art weakly-supervised temporal action localization, with 25.4% mAP at IoU 0.5 on THUMOS'14 and 32.3% on ActivityNet1.3.","lead":"This paper presents a method to find when actions happen in videos using only the video's action tags, not frame-level start and end times. It grows small trusted action regions outward and erases them to force the network to notice less obvious action parts, improving localization on two standard benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Erasing mechanism's causal role is untested; added classifier loss alone could explain the gains.","rationale":"The reader's weakest_assumption correctly identifies the lack of isolation of the erasing mechanism as the core unvalidated condition. My analysis agrees: the paper's central claim depends on erasing being the causal driver of improved localization, but the provided ablations (Table 2) only show that adding the classifier branch helps, not that erasing specifically is responsible. The improvement could equally arise from the extra classification loss regularizing the shared feature extractor, an alternative that is neither tested nor discussed. This is a load-bearing concern because it targets the paper's stated contribution, not a peripheral detail. Other issues, such as threshold tuning on the validation set and missing code, are real but secondary; they affect reproducibility rather than the validity of the causal mechanism. The verdict should remain CONDITIONAL because the empirical results are plausible and the method may work, but the central mechanism must be verified with a targeted ablation. No change to the reader's verdict is needed.","tokens_in":184,"tokens_out":4731,"duration_ms":152992,"concrete_test":"Run the following ablation on THUMOS'14: (A) SSG only, as reported; (B) ASSG but with the erasing threshold set to 0 (i.e., the classifier trains on the full shared feature map, no erasing); (C) ASSG as proposed. If variant (B) achieves an average mAP within 1 point of variant (C), then the improvement is not caused by erasing-based mining, and the adversarial-mechanism claim collapses. Additionally, for variants (A) and (C), compute the recall of ground-truth action segments that are not covered by the initial CAS seeds: if variant (C) does not recover substantially more of these initially missed segments than variant (A), the claimed discovery of trivial or missing actions is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that adversarial erasing of SSG-activated regions forces the classifier to mine previously missed action regions, which in turn drives seed growth and yields the large improvement (SSG 34.2% to ASSG 43.5% average mAP on THUMOS'14, Table 2). For this claim to hold, the erasing operation must be the causal driver of the improvement, not merely the addition of a second classification loss on shared features. The paper's ablations only compare SSG-only against the full ASSG; they never isolate the erasing step. A plausible alternative explanation is that the classifier branch, trained with a standard video-level cross-entropy loss, simply provides a stronger gradient signal that refines the shared feature extractor, improving the SSG's heatmaps regardless of whether any features are erased. The paper also provides no quantitative evidence that action instances absent from the initial seeds are actually recovered: the qualitative example in Figure 3 shows improved boundary filling and the emergence of one missing instance, but this is a single cherry-picked case. Without an ablation that trains the classifier on unerased features, or with random erasing, or that measures recall of initially missed instances, the 'adversarial mining' mechanism remains unsupported. The term 'adversarial' is also used loosely, as Algorithm 1 alternates two losses rather than optimizing a minimax objective, further weakening the causal story.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Adversarial Seeded Sequence Growing (ASSG) for weakly-supervised temporal action localization. The method takes Class Activation Sequence (CAS) peaks as initial seeds, grows these seeds with a Seeded Sequence Growing (SSG) network, and adds a classifier branch that erases the high-confidence SSG-activated regions from the shared features and then predicts video-level classes via a self-adaptive pooling (SAP) aggregation. The two losses are alternated with the claim that this constitutes adversarial training. Experiments on THUMOS'14 and ActivityNet1.3 report large improvements over the CAS baseline (e.g., average mAP 0.1--0.5 from 24.4% to 34.2% with SSG and to 43.5% with ASSG in Table 2) and state-of-the-art or competitive results against other weakly supervised methods, including 25.4% mAP at IoU 0.5 on THUMOS'14 and 32.3% at IoU 0.5 on ActivityNet1.3.","tokens_in":12642,"tokens_out":5022,"duration_ms":57030,"significance":"If the causal mechanism claimed by the paper holds, ASSG would be a valuable advance: it combines seed growing with an erasing classifier to mine less discriminative action regions, and it reports consistent gains on two standard benchmarks. The paper has concrete strengths: ablations in Tables 1 and 2 show the contribution of the SAP pooling and of the overall two-module system, the evaluation is against external baselines, and the implementation details are sufficiently specific to be reproduced. However, the central claim that adversarial erasing is the driver of the improvement is not experimentally isolated, and the adversarial interpretation of the training procedure is not substantiated by analysis or ablation. These issues affect the validity of the paper's main explanatory narrative rather than merely its presentation.","major_comments":[{"comment":"The causal role of the erasing step is untested. The only architectural ablation compares SSG (CAS w/ SSG) with ASSG (CAS w/ SSG w/ classifier), so the 9.3% average-mAP gain (34.2% to 43.5% in Table 2) could come entirely from adding a second classification loss on shared features, independent of whether any features are erased. A convincing experiment must compare the full ASSG against a variant that trains the same classifier on unerased features, or against a random-erasing control, to show that erasing the SSG-activated regions is what drives the improvement.","section":"§3.3, Algorithm 1 lines 14–16, Table 2"},{"comment":"The term 'adversarial' is used more strongly than the algorithm supports. Algorithm 1 simply alternates optimization of the seeding loss L_seed and the classification loss L_class; there is no minimax objective, gradient reversal, or equilibrium analysis. The authors should either provide evidence that the erasing step creates a harder classification distribution and that the alternation yields the claimed mining behavior, or revise the claims to describe the procedure as alternating optimization with feature erasure rather than true adversarial training.","section":"§3.4, Algorithm 1, Section 1"},{"comment":"No quantitative evidence is provided that the classifier actually recovers trivial or previously missed action instances. The only support is a single qualitative example in Figure 3 showing one missing instance emerging. The authors should measure, for example, the recall of ground-truth instances that are missed by the initial CAS seeds or by the SSG output, and compare it before and after the erasing-classifier branch is added. Without such a measure, the paper's central 'mining' claim remains a plausible but unverified interpretation of the aggregate mAP gain.","section":"§3.3, §4.3.4, Figure 3"},{"comment":"The thresholds θ_d, θ_a, and λ are selected on the same benchmark used for the final state-of-the-art comparison, with no held-out validation split and no error bars or multiple-run statistics. This is a robustness concern for the reported margins, especially the 9.3% gain in Table 2. The authors should report standard deviations over at least a few runs, or select hyperparameters on a validation portion and report the corresponding test results, to rule out selection bias.","section":"§4.2, §4.3.1"}],"minor_comments":[{"comment":"The notation is inconsistent: G(H_{c,t}, S_c, θ_d) is written with a condition l ∈ N(S_c) but the heatmap index is t, and the neighbor variable l is not defined in terms of t. Please clarify the neighbor set and the indexing.","section":"Eq. (1)"},{"comment":"The definition of A_{c,t}(X) is ambiguous: the meaning of |f_{c,t}(X)| and the sum over i = 1 to |f_{c,t}(X)| of f^i_{c,t}(X) is unclear, and the relationship between f(·) and the earlier feature map F is not stated precisely.","section":"Eq. (3)"},{"comment":"There are several typos and formatting artifacts: 'earsing' for 'erasing', 'Evalution' for 'Evaluation', 'Hanmmer' for 'Hammer', 'erasion' for 'erasure', and a stray glyph after the NMS sentence in Section 3.5. These should be corrected.","section":"Throughout"},{"comment":"The description of 'a pair of opposite ReLU activations' used to separate foreground and background features is too brief; the authors should explain how the positive and negative activation branches are constructed and how they are used to form the foreground features F.","section":"§3.3"},{"comment":"The text states that STAR without frequency annotations reports 44.0% average mAP, but the STAR row in Table 3 implies an average of about 47.0% over IoU 0.1–0.5. Please clarify which configuration each reported STAR number corresponds to.","section":"Table 3 and §4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own prior STAR work: the CAS seeds are produced by a reimplementation of STAR, and the main state-of-the-art comparison is with STAR. This is not by itself a flaw, but it makes the comparison less independent than the other external baselines. The main technical gap is the missing ablation isolating the erasing mechanism, which is central to the paper's adversarial-mining narrative. I would be willing to reconsider after the authors provide that ablation and, ideally, a quantitative measure of missed-instance recall."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a legitimate and useful application of two known ideas—seed region growing from weakly supervised segmentation and adversarial erasing from object region mining—to temporal action localization with only video-level labels. Second, the headline empirical claim is real: on THUMOS'14 the full model reaches 25.4% mAP at IoU 0.5, above the STAR baseline, and ActivityNet1.3 shows consistent gains at higher IoUs. That is worth taking seriously.\n\nWhat the paper does well: the module ablations show a clear progression (CAS 24.4, CAS+SSG 34.2, full ASSG 43.5 average mAP), the threshold sensitivity analysis shows the method is not knife-edge, and the classifier branch adds no parameters. The qualitative figure shows boundary filling and one recovered instance. The writing is clear about where prior CAS-based methods fail.\n\nWhere the soft spots are, in rough order of importance. First, the causal story for the 'adversarial' classifier is untested. The ablation only compares SSG with and without the erasing classifier; it never trains the classifier on unerased features, or with random erasing. The stress-test worry is legitimate: the 9.3-point gain could come entirely from the extra classification loss sharpening the shared feature extractor, with the erasing operation doing little of the work. Second, the word 'adversarial' is doing more work than the mechanism supports—Algorithm 1 alternates two losses; there is no minimax objective. That is a framing issue, but it matters because it shapes the claimed contribution. Third, there are no error bars, thresholds are tuned on the same benchmark, and the STAR comparison is a bit tangled because STAR was also evaluated with and without frequency annotations. None of these are fatal to the empirical result, but they limit how precisely we can attribute the improvement.\n\nWho should read this: anyone working on weakly-supervised temporal action localization or on adapting image-domain weakly supervised methods to video. It deserves a serious peer review. My recommendation: send it out, but ask the authors to isolate the erasing mechanism with the obvious missing ablation and to tone down the 'adversarial' terminology. If that ablation comes back clean, this is a solid conditional accept.","headline":"A legitimate application of seed-growing and erasing to temporal action localization with real high-IoU gains, but the 'adversarial' erasing mechanism is never isolated, so the causal story is weaker than the ablation suggests.","tokens_in":13184,"tokens_out":2656,"would_cite":true,"duration_ms":28622,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adversarial seed growing recovers full action boundaries from sparse CAS peaks and mines missed action regions, giving state-of-the-art weakly-supervised localization on THUMOS'14 and ActivityNet1.3.","keywords":["temporal action localization","weak supervision","adversarial learning","seed growing","class activation sequence","erasing-based learning","action detection","video understanding"],"falsifier":"Train the full ASSG pipeline but replace the erased temporal segments with a randomly selected set of the same size as the SSG-activated regions; if the mAP gain over the SSG-only baseline is unchanged under this control, then the specific choice of which regions to erase is not what drives the classifier to recover missed actions, so the adversarial mechanism is not doing the claimed work.","tokens_in":12197,"feed_emoji":"🎬","tokens_out":11297,"duration_ms":101838,"temperature":0.7,"pith_summary":"The paper takes on a known weakness of weakly-supervised temporal action localization: a Class Activation Sequence (CAS) trained on video-level labels fires only on the most discriminative fragments of an action, so long actions are detected as disconnected peaks and inconspicuous action instances are missed entirely. The proposed ASSG framework treats those CAS peaks as seeds and grows them outward with a Seeded Sequence Growing (SSG) network, while a second classifier, trained on the shared features after the seeded regions are erased, is forced to discover the less obvious action regions. The two modules are trained adversarially, so the erasing drives the classifier to find new evidence and the classifier's difficulty in turn pushes the seeds to grow further. If this works as claimed, high-quality temporal boundaries become recoverable from video-level class labels alone, and the paper reports state-of-the-art detection on THUMOS'14 and ActivityNet1.3, with the largest gains at the stricter IoU thresholds.","feed_headline":"Seed-growing lifts weak-label detection to 25.4% mAP on THUMOS'14","feed_subtitle":"Two modules grow sparse activation peaks to boundaries and find missed actions, using video-level labels only.","key_machinery":"The load-bearing object is the adversarial pair of the Seeded Sequence Growing (SSG) network and the erasing video-level classifier sharing one feature map. SSG owns a per-class temporal heatmap and a growing criterion $G(H_{c,t}, S_c, \\theta_g)$ that labels a neighbor of an existing seed as class $c$ when the heatmap exceeds the growing threshold $\\theta_g$ and the class is the argmax at that segment; this extends seeds toward action boundaries under the seeding loss $L_{seed}$. The classifier erases all segments whose heatmap exceeds the erasing threshold $\\theta_a$, then aggregates the remaining features with Self-Adaptive Pooling (SAP)—a weighted sum whose attention weights come from the shared feature map itself—and is trained with cross-entropy $L_{class}$. The adversarial loop is the alternation: erasing forces the classifier to find class-relevant evidence in the leftover features, and the classifier's gradients, passed through the shared map, push the SSG to grow seeds into regions the classifier still needs, so the two losses alternately raise the difficulty for each other.","core_discovery":"The central claim is that the two failure modes of CAS-based weakly-supervised localization—peak fragmentation of long actions and omission of trivial instances—can be removed by an adversarial loop rather than by stronger supervision. The SSG network initializes each action category with the sparse high-confidence regions of a CAS heatmap and then progressively labels neighboring segments according to a growing criterion that requires the segment's heatmap value to exceed a threshold and its class to be the maximum over all classes. In parallel, a self-adaptive classifier erases those SSG-activated high-confidence regions from the shared feature map and classifies the video from what remains, using a Self-Adaptive Pooling (SAP) aggregation that re-weights segments. Because the classifier must make the correct video-level prediction without the most discriminative features, it is pushed to attend to less salient or previously missed action segments, and the shared parameters mean this pressure also refines the heatmaps. The combination raises the average mAP over IoU 0.1–0.5 on THUMOS'14 from 34.2% with SSG alone to 43.5%, and reaches 25.4% at IoU 0.5 and 32.3% at IoU 0.5 on ActivityNet1.3, both above prior weakly-supervised methods.","pith_inferences":["The same adversarial erasing trick could be carried back to 2D weakly-supervised object localization: take CAM peaks as seeds, grow them with a seeded region-growing network, and erase the grown regions so a classifier is forced to attend to the rest of the object; the 1D formulation is a direct special case.","The Self-Adaptive Pooling aggregation, which re-weights segments by their learned attention without extra parameters, may be useful for any weakly-supervised classification of long sequences (e.g., audio event detection) where background segments would otherwise dominate a global average pool.","A testable next step is to use the SSG's grown heatmaps as pseudo ground-truth segment labels for training a fully-supervised detector, which could close most of the remaining gap to strong supervision without any manual segment annotation.","Because the method inherits its seeds from a CAS model, its ceiling is tied to seed quality; running the adversarial loop on top of several diverse CAS initializations and fusing their heatmaps might improve robustness to seed failure."],"forward_implications":["Long and fragmented actions can be localized at high IoU without frame-level supervision, because boundary recovery is formulated as growing reliable seeds rather than as explicit boundary regression.","Any CAS-based weakly-supervised detector can be augmented with the SSG and erasing classifier to improve its high-IoU performance, since the added module needs no new labels and the erasing classifier adds no parameters.","The growing criterion's two thresholds ($\\theta_g$ for expansion, $\\theta_a$ for erasing) give a simple, dataset-robust control over the aggressiveness of growth and the amount of feature hiding, as the ablation shows performance changes little across threshold values.","The adversarial training's main contribution is instance completeness and boundary refinement rather than coarse recall, since the gains concentrate at IoU 0.3–0.5 on THUMOS'14."],"supporting_citations":[{"why":"Supplies the dynamic-supervision growing strategy that ASSG re-implements in one dimension for temporal sequences.","marker":"[8]"},{"why":"Introduces step-by-step erasion and one-by-one collection of action regions, the direct inspiration for the erasing classifier.","marker":"[31]"},{"why":"Provides the adversarial-erasing approach for mining complementary object regions that the classifier branch adapts to video features.","marker":"[25]"},{"why":"The STPN network serves both as a source of initial CAS seeds and as a baseline; STPN-CAS w/ ASSG shows the upgrade is seed-agnostic.","marker":"[15]"},{"why":"The STAR network is the main seed source and the strongest weakly-supervised baseline compared on both benchmarks.","marker":"[28]"},{"why":"Supplies the pre-trained I3D two-stream features on which the whole pipeline operates.","marker":"[3]"},{"why":"Provides the ActivityNet v1.3 benchmark and its evaluation protocol used for the second set of experiments.","marker":"[7]"},{"why":"Provides the THUMOS'14 dataset used for all ablations and the primary comparison.","marker":"[10]"}],"fun_headline_variants":["Adversarial seed-growing boosts weak-label action localization","Erase and grow: adversarial trick finds missed actions in videos","Seed-growing adversarial nets hit 25.4% mAP on THUMOS'14","Adversarial loop grows seeds to find all actions in weak-label video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that carries the whole method is that after the SSG's high-confidence regions are erased, the remaining feature sequence still contains enough class-specific evidence about the non-salient or previously missed action instances for the classifier to recover them; if the erased features are information-free for those instances, the adversarial mining collapses to the seed-growing component alone.","fun_headline_variants_meta":{"raw":{"variants":["Adversarial seed-growing boosts weak-label action localization","Erase and grow: adversarial trick finds missed actions in videos","Seed-growing adversarial nets hit 25.4% mAP on THUMOS'14","Adversarial loop grows seeds to find all actions in weak-label video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000676,"raw_usage":{"total_tokens":3159,"prompt_tokens":1113,"completion_tokens":2046,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":729,"completion_tokens_details":{"reasoning_tokens":1969}},"tokens_in":729,"tokens_out":2046,"duration_ms":14212,"temperature":1.0,"reasoning_tokens":1969,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:44:47.813308+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the full ASSG pipeline but replace the erased temporal segments with a randomly selected set of the same size as the SSG-activated regions; if the mAP gain over the SSG-only baseline is unchanged under this control, then the specific choice of which regions to erase is not what drives the classifier to recover missed actions, so the adversarial mechanism is not doing the claimed work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces step-by-step erasion and one-by-one collection of action regions, the direct inspiration for the erasing classifier."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the adversarial-erasing approach for mining complementary object regions that the classifier branch adapts to video features."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The STPN network serves both as a source of initial CAS seeds and as a baseline; STPN-CAS w/ ASSG shows the upgrade is seed-agnostic."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The STAR network is the main seed source and the strongest weakly-supervised baseline compared on both benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained I3D two-stream features on which the whole pipeline operates."},{"cited_title":"Jiang, J","cited_arxiv_id":null,"evidence_quote":"Provides the THUMOS'14 dataset used for all ablations and the primary comparison."}],"review_version":1}