{"id":"2730b6c0-0561-42ec-a508-f122819dd192","arxiv_id":"1908.08216","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"3C-Net adds a center loss and a counting loss to weakly supervised action classification, improving temporal action localization state of the art on THUMOS14 and ActivityNet 1.2.","lead":"This paper introduces 3C-Net, a neural network that localizes actions in videos using only video-level labels and the number of action instances, without frame-level boundary labels. It combines three losses to make action features more separable and reports gains over prior weakly supervised methods on two video benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'predicted count' in Eq. 9 is a softmax-weighted mean of the T-CAM, not an instance count, because the temporal softmax in Eq. 4 sums to 1; the central counting-loss mechanism is therefore mathematically misdescribed.","rationale":"The reader's weakest_assumption exactly identifies the most load-bearing concern: Eq. 9 is mathematically a softmax-weighted mean, not a count, because the temporal attention sums to 1. This attacks the central explanatory mechanism of the paper while leaving the empirical SOTA claim potentially intact. The ablation gain from the counting loss could survive even if the loss is only an auxiliary regularizer, so the appropriate disposition remains conditional rather than rejection. The additional issue that count labels were derived from temporal segment annotations in Section 4.1 further weakens the weak-supervision narrative, but it is secondary to the misdescribed counting mechanism. The concrete test above would settle whether m_i(j) actually tracks instance counts and whether an unnormalized count formulation changes the result.","tokens_in":12701,"tokens_out":4968,"duration_ms":54744,"concrete_test":"Run the released 3C-Net code on the THUMOS14 validation videos and record m_i(j) from Eq. 9 together with ground-truth counts n_i(j). If m_i(j) approximates n_i(j) across videos with counts ranging from 1 to 15, the quantity might behave as a count; if m_i(j) remains O(1) and does not scale with the number of action segments, Eq. 9 is a softmax-weighted mean. As an analytic check, replace the temporal softmax in Eq. 4 with an unnormalized sigmoid activation so that the summed quantity can genuinely represent a count, then retrain and compare mAP at IoU=0.5 against CLS+CL; if there is no improvement, the claimed counting mechanism is not responsible for the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.4, Eq. 4 defines a^F_i(t,j) = exp(C^F_i(t,j)) / sum_l exp(C^F_i(l,j)), so sum_t a^F_i(t,j) = 1. Equation 9 then computes m_i(j) = sum_t a^F_i(t,j) C^F_i(t,j), which is a weighted average of the T-CAM over time, not a sum of per-segment densities. A weighted average cannot represent the number of action instances: a video with one action instance and a video with fifteen instances can yield the same m_i(j) if the T-CAM values are rescaled. Consequently, the counting loss in Eq. 10 drives a weighted mean toward the integer label, not the total activation mass toward the count. The paper's claim that the loss makes prediction scores sum to a positive value within action instances and zero otherwise, and that this delineates adjacent action instances by matching counts, is unsupported. The observed gain from the counting term in Table 4 (26.6 vs 24.6 mAP at IoU=0.5) may come from an implicit regularizer, but it is not evidence for count-based delineation. This is an internal inconsistency in the stated mechanism, not merely a disagreement with external consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 3C-Net for weakly-supervised temporal action localization using video-level class labels and action instance counts. The architecture is a two-stream I3D-based network with three losses: a classification loss on top-k pooled T-CAMs, a multi-label center loss with attention-based feature aggregation, and a counting loss. Experiments on THUMOS14 and ActivityNet 1.2 report state-of-the-art results, with ablations showing progressive improvements from each term. The main claimed novelty is the counting loss, which is intended to separate adjacent action instances by matching a predicted count to the ground-truth count.","tokens_in":12971,"tokens_out":4364,"duration_ms":40605,"significance":"If the counting mechanism were correctly formulated, the idea of using instance counts as weak supervision for temporal action localization would be a useful contribution, especially since count labels are cheap relative to frame-level annotations. The multi-label adaptation of the center loss is also a sensible extension. The paper provides source code, thorough ablations, and comparisons on two benchmarks. However, the central counting loss as defined is mathematically a weighted average rather than a count, so the claimed mechanistic explanation is not supported; the empirical gains may still hold but require a corrected formulation and a re-evaluation of the underlying claim.","major_comments":[{"comment":"Equation (9) defines m_i(j) = sum_t a^F_i(t,j) C^F_i(t,j). Since a^F_i(t,j) is a temporal softmax over t, sum_t a^F_i(t,j) = 1, so m_i(j) is a weighted average of the T-CAM values over time. A weighted average cannot represent an instance count: one can rescale all C^F_i(t,j) and obtain any desired average, so a single-instance video and a fifteen-instance video can have identical m_i(j). Consequently, the counting loss in Eq. (10) does not force the activation mass to match the count; it forces a normalized average toward the integer label. This invalidates the claim in Section 1 that 'the prediction scores sum up to a positive value within action instances and zero otherwise' and the explanation that the loss delineates adjacent instances. The observed improvement in Table 4 (26.6 vs. 24.6 mAP) may be a regularization effect, but it is not evidence for count-based delineation.","section":"Section 3.4, Eq. (9)"},{"comment":"The paper states that the attention-weighted T-CAM is 'equivalent to a density map' and that its summation yields the predicted count. This analogy with object counting [6] is unsupported: in density-based counting, the density map is a spatial map whose integral equals the count, whereas here C^F is a class score produced by a classifier and the softmax weights are normalized to sum to one. Therefore the quantity in Eq. (9) is not an integral of a density. Furthermore, the claim that 'temporal attention weighting ignores the background video segments' is not implemented, because the softmax attention assigns nonzero (albeit small) weights to all segments and no threshold is applied in the counting module. The authors should either redefine m_i as a true sum over T-CAM values (with hard attention masking) or remove the density-map/count interpretation.","section":"Section 3.4, density-map claim"},{"comment":"The counting loss is computed as a per-video, per-class scalar after summing over time, so it cannot by itself separate adjacent instances; it provides only a global constraint on the T-CAM. The paper does not report any direct evidence that the counting loss improves instance-level segmentation, such as the number of detected instances versus the ground-truth count, or that it is responsible for the qualitative delineation shown in Fig. 1. The ablation in Table 4 isolates the loss, but it does not test the proposed mechanism. Without such evidence, the claim that the counting loss 'delineates adjacent action instances' is not established.","section":"Section 4.3 and Table 4"}],"minor_comments":[{"comment":"There is a typo: 'disrcriminative' should be 'discriminative'.","section":"Section 3.4"},{"comment":"The caption reads 'PoleVaultaction'; it should be 'PoleVault action'.","section":"Figure 1 caption"},{"comment":"The denominator '1 + sum_i y_i(j)' is unusual; standard center updates divide by the number of samples in the class. Please clarify whether this is intentional to avoid division by zero, and explain the impact on the center update.","section":"Equation (7)"},{"comment":"The notation n_i(j) is introduced as the ground-truth count, but earlier y_i is a binary multi-hot label. Please define n_i explicitly and state its relationship to y_i.","section":"Equation (10)"},{"comment":"The description of the detection threshold η for ActivityNet 1.2 is terse; please state explicitly that η is set to 0 for that dataset and how the detection algorithm then generates segments from the T-CAM.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The counting-loss flaw is the crux of the paper: Eq. (9) does not implement a count, so the central claim is not supported as written. The empirical results are plausible and the code is provided, but the authors need to correct the formulation and re-run experiments. This is a substantive but potentially fixable issue, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere’s my read of 3C-Net. The empirical recipe is promising: they combine classification, center, and counting losses, and get real gains on THUMOS14 and ActivityNet with only video-level labels plus counts. The center-loss adaptation to multi-label videos—attention-weighted feature aggregation per class—is genuine and well motivated. If that were the whole paper, I’d be fairly confident.\n\nThe problem is the counting loss, which is the main novelty. Eq. 9 defines the predicted count as the temporal softmax–weighted sum of T-CAM activations. Since the softmax sums to 1 over time, that is a weighted average, not a count of instances. A video with one action instance and a video with fifteen can give the same value if the activations are rescaled. The text calls it a density map and says summation yields the count, but the summation is over a normalized distribution. That is not what the equation does. The loss is then driving this average toward an integer ground-truth count, which can still act as a strange regularizer, but the paper’s explanation for why it separates adjacent instances is unsupported.\n\nThe other soft spot is supervision. The count labels are generated from the temporal segment annotations that the method is supposedly avoiding. That makes the “weakly-supervised” claim less clean. It may still be cheaper than full boundary annotation, but the paper should say so. Missing error bars and the fact that the code isn’t verified are minor; the experiments are reproducible in principle.\n\nOverall, the center-loss contribution is solid and the empirical trend is probably real, but the central counting mechanism needs to be re-described or redefined. If the authors can fix that, this is a useful paper for the weakly-supervised localization community. As is, it deserves a serious referee, but with major revision, and I’d want the counting claim clarified before accepting.\n\nBring it to reading group maybe—it’s a good example of a loss that “works” for reasons other than what the paper says.","headline":"A solid empirical paper whose headline counting mechanism is mathematically misdescribed: Eq. 9 computes a softmax-weighted average, not an action count.","tokens_in":13524,"tokens_out":3374,"would_cite":false,"duration_ms":31731,"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":"Video-level labels plus action counts, trained with classification, center, and counting losses, sharpen temporal action localization to beat prior weak supervision by 4.6% mAP on THUMOS14 and 3.7% on ActivityNet 1.2.","keywords":["weakly supervised temporal action localization","action count","center loss","temporal class activation maps","video-level supervision","multi-label classification","THUMOS14","ActivityNet"],"falsifier":"Shuffle the ground-truth count labels across THUMOS14 training videos while keeping the same multiset of counts. If mAP at IoU=0.5 remains near 26.6, the specific count values are not carrying the signal. A complementary check is to correlate predicted counts $m_i(j)$ with true counts on held-out videos: near-zero correlation with unchanged localization would show the term acts as a generic regularizer rather than an actual count estimator.","tokens_in":12483,"feed_emoji":"🎬","tokens_out":8691,"duration_ms":80158,"temperature":0.7,"pith_summary":"Temporal action localization—deciding when an action happens in an untrimmed video—usually needs frame-level boundary annotations, which are expensive and subjective. This paper claims that video-level category labels plus a simple count of how many times each action appears are enough to train a strong localizer, and that adding a counting loss specifically helps separate adjacent action instances that otherwise merge into one detection. The proposed 3C-Net combines classification, multi-label center, and counting losses on a two-stream I3D feature pipeline. On THUMOS14 it raises mAP at IoU=0.5 from 19.1 (classification only) to 26.6, an absolute gain of 4.6% over the prior weakly supervised state of the art, and on ActivityNet 1.2 it gains 3.7 mean mAP. If the claim holds, annotating videos by category and count is a much cheaper route to usable localization than drawing temporal boxes.","feed_headline":"Counting actions adds 4.6% mAP to weakly supervised localization","feed_subtitle":"No frame-level labels needed: category labels plus action counts sharpen temporal boundaries and set a new state of the art.","key_machinery":"The load-bearing object is the temporal class activation map, a per-segment vector of class scores learned from video-level labels. Three mechanisms shape it: a top-k classification loss for video-level class separability; a multi-label center loss with class-specific centers $c_j$ and attention-weighted feature aggregation $f^a_i(j)=\\sum_t a^a_i(t,j)\\,x^a_i(t)/\\sum_t a^a_i(t,j)$ pulling same-class features together; and a counting loss, $L_{\\text{count}}$, that compares attention-weighted T-CAM sums to ground-truth counts, using relative error for present classes and absolute error for absent classes. The counting loss's inverse-count weighting makes low-count videos—exactly the cases where adjacent instances are easiest to merge—dominate the gradient.","core_discovery":"The central discovery is that jointly optimizing classification, center, and counting losses yields a temporal class activation map (T-CAM) whose high-scoring segments align with individual action instances rather than with the whole action-containing region. The classification term provides video-level category separability; the adapted multi-label center loss clusters features per class using attention-weighted aggregation over high-attention segments; the counting term compares the attention-weighted T-CAM sum $m_i(j)=\\sum_t a^F_i(t,j)\\,C^F_i(t,j)$ with the ground-truth count $n_i(j)$ and penalizes their relative difference. Together these losses push the T-CAM to be near zero in background and between-instance gaps. The paper reports 26.6 mAP at IoU=0.5 on THUMOS14 (up from 22.0 for [16]) and 21.7 mean mAP on ActivityNet 1.2 (up from 18.0 for [16]), with the center loss alone already beating prior weakly supervised methods.","pith_inferences":["If the count term is mostly a regularizer, a cheap way to test this is to replace true counts with a learned count-prediction head trained on a small labeled subset; most of the gain persisting would mean per-video counts do not need to be manually annotated.","The attention-weighted center loss is not tied to action localization; it could be applied to other multi-label temporal tasks such as sound event detection or gesture spotting, where category presence is known but instance boundaries are costly.","On ActivityNet 1.2 the count term adds much less than on THUMOS14, consistent with the paper's observation of fewer multi-instance videos; this predicts that datasets with many repeated actions will show the largest benefit from count supervision.","Because the counting module itself has no learnable parameters, a natural ablation is to replace it with a penalty on the temporal derivative of the T-CAM; if the two behave similarly, the gain is smoothness pressure rather than information about instance number."],"forward_implications":["Video-level category labels plus count labels can replace frame-level boundary annotations for training localizers, lowering annotation cost and reducing boundary subjectivity.","The multi-label center loss alone—without count supervision—already surpasses prior weakly supervised localization, so the technique transfers to settings where counts are unavailable.","The counting loss with relative-error weighting concentrates learning on low-count videos, the regime where instance merging most hurts average precision.","Penalizing absent categories toward zero predicted count suppresses background and false detections; the paper's ablation attributes 1.5 mAP to this term.","Joint training also improves action classification, reaching 86.9 mAP on THUMOS14, so the same model serves both recognition and localization."],"supporting_citations":[{"why":"Provides the weakly supervised state-of-the-art baseline that 3C-Net must beat on both THUMOS14 and ActivityNet 1.2.","marker":"[16]"},{"why":"Introduces T-CAMs and temporal attention for weakly supervised localization; 3C-Net builds directly on this representation.","marker":"[14]"},{"why":"Supplies the original center loss that the paper adapts to multi-label video-level supervision.","marker":"[30]"},{"why":"Basis for treating attention-weighted activation sums as predicted counts, with summation yielding a density-map count.","marker":"[6]"},{"why":"Provides the I3D RGB and flow features used as fixed backbones for both streams.","marker":"[4]"},{"why":"Defines the classification-loss baseline for weakly supervised action detection that the paper's formulation extends.","marker":"[28]"},{"why":"Demonstrates count-guided weakly supervised localization in the image domain, motivating the use of count labels here.","marker":"[8]"}],"fun_headline_variants":["Action counts yield 4.6% mAP jump for weakly supervised localization","Category labels plus action counts sharpen temporal localization","3C-Net lifts weakly supervised action localization via count and center loss","Weakly supervised? Count actions to beat SOTA without frame labels","Counting across time: new SOTA for weakly supervised action spotting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the predicted quantity from the score curve really is the number of action instances, not just a weighted average of activation strengths; if that identification fails, the extra loss term may still regularize training but it is not performing the counting the paper describes.","fun_headline_variants_meta":{"raw":{"variants":["Action counts yield 4.6% mAP jump for weakly supervised localization","Category labels plus action counts sharpen temporal localization","3C-Net lifts weakly supervised action localization via count and center loss","Weakly supervised? Count actions to beat SOTA without frame labels","Counting across time: new SOTA for weakly supervised action spotting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000844,"raw_usage":{"total_tokens":3686,"prompt_tokens":967,"completion_tokens":2719,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":2631}},"tokens_in":583,"tokens_out":2719,"duration_ms":20021,"temperature":1.0,"reasoning_tokens":2631,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:45:40.675787+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Shuffle the ground-truth count labels across THUMOS14 training videos while keeping the same multiset of counts. If mAP at IoU=0.5 remains near 26.6, the specific count values are not carrying the signal. A complementary check is to correlate predicted counts $m_i(j)$ with true counts on held-out videos: near-zero correlation with unchanged localization would show the term acts as a generic regularizer rather than an actual count estimator.","supporting_citations":[{"cited_title":"W- talc: Weakly-supervised temporal activity localization and classification","cited_arxiv_id":null,"evidence_quote":"Provides the weakly supervised state-of-the-art baseline that 3C-Net must beat on both THUMOS14 and ActivityNet 1.2."},{"cited_title":"Weakly supervised action localization by sparse temporal pooling network","cited_arxiv_id":null,"evidence_quote":"Introduces T-CAMs and temporal attention for weakly supervised localization; 3C-Net builds directly on this representation."},{"cited_title":"A discriminative feature learning approach for deep face recog- nition","cited_arxiv_id":null,"evidence_quote":"Supplies the original center loss that the paper adapts to multi-label video-level supervision."},{"cited_title":"Object counting and instance segmentation with image-level supervision","cited_arxiv_id":null,"evidence_quote":"Basis for treating attention-weighted activation sums as predicted counts, with summation yielding a density-map count."},{"cited_title":"Untrimmednets for weakly supervised action recognition and detection","cited_arxiv_id":null,"evidence_quote":"Defines the classification-loss baseline for weakly supervised action detection that the paper's formulation extends."},{"cited_title":"C-wsl: Count-guided weakly supervised lo- calization","cited_arxiv_id":null,"evidence_quote":"Demonstrates count-guided weakly supervised localization in the image domain, motivating the use of count labels here."}],"review_version":1}