{"id":"d95b4ee7-4ce5-4e54-8dd3-4a59bdd80a15","arxiv_id":"2505.21943","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A point-to-region loss that propagates pseudo-label confidence to background pixels makes semi-supervised point-based crowd counting work and beats prior methods on ShTech, UCF-QNRF, and JHU++.","lead":"This paper replaces point-to-point matching with point-to-region matching in semi-supervised crowd counting, letting background pixels around each predicted person share the predicted point's confidence. This stabilizes pseudo-label training and improves counting and localization accuracy on four crowd datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (23) marks every pixel farther than μ from any pseudo-point as reliable background; if teacher recall is imperfect, missed people are trained as background, and the paper never quantifies this failure mode.","rationale":"The reader's weakest assumption is also the one I find most load-bearing. The method's novelty is precisely the confidence-propagation scheme of Eq. (23), and the central SOTA claim rests on it. The paper argues from PSAM that background supervision is needed to suppress over-activation and then treats 'far from any pseudo-point' as background. But the PSAM evidence is about over-activation around existing foreground points, not about the correctness of labeling unannotated regions. The missing-person scenario is a direct failure mode: pseudo-label extraction at threshold 0.5 guarantees recall below 1, especially in dense crowds, and nothing in the loss distinguishes 'no person here' from 'teacher did not see this person.' The empirical tables are broad and the method is plausible, which is why I do not move the verdict to REJECT; however, the failure mode is unquantified and easy to test. The reader's conditional verdict is appropriate, so verdict_should_be is UNCHANGED. My proposed check isolates the assumption by using ground-truth knowledge to mask only pixels near missed pseudo-points; this is a single controlled intervention rather than another hyperparameter sweep. If the masked variant performs the same, the Eq. (23) assumption is not the reason P2R works, and the conditional can be lifted.","tokens_in":19788,"tokens_out":8963,"duration_ms":95361,"concrete_test":"On ShTech A with the 5% labeled split, after the 100-epoch warmup, compute teacher pseudo-point recall against ground-truth points on a held-out labeled subset. Then rerun P2R with one change: in Eq. (23), set Z[i,i]=0 for any pixel i whose nearest ground-truth point has no pseudo-point within μ, instead of marking it reliable background; keep every other term identical. Compare MAE of the standard and oracle-masked variants over three seeds. If the masked variant is not materially better, the far-background assumption is not load-bearing; if it is substantially better, Eq. (23) is actively training true positives as background and the central claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that P2R supplies a correct confidence scheme for semi-supervised point-based counting. The load-bearing step is Eq. (23), Z = diag[M_st ζ + (1_n - β)], where β[i]=1 only for pixels within radius μ of some extracted pseudo-point. Consequently, every pixel farther than μ from all pseudo-points receives confidence 1 with pseudo-label 0. This is valid only if the teacher's pseudo-labels have near-perfect recall outside those radii. In dense crowds, a person whose head score falls below the 0.5 extraction threshold and who is not within μ of a detected person is silently converted into a confident background example. Since the teacher is an EMA of the student, the error is potentially self-reinforcing: missed people are pushed toward background, lowering future teacher scores at those locations. The PSAM analysis documents over-activation near foreground points but does not measure teacher false negatives. The μ ablation in Fig. 6(c) varies the radius with real pseudo-labels; it does not isolate what happens when ground-truth points are missing from the pseudo-label set. If teacher recall is imperfect, part of the reported advantage (e.g., 5% ShTech A MAE 69.9 vs 83.7 for OT-M) may reflect suppression of difficult missed regions rather than correct learning. This is the least secure condition for the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a point-to-region (P2R) loss for semi-supervised point-based crowd counting. It first observes that the standard point-to-point (P2P) matching loss on unlabeled data has an identically zero background term (Eq. 9), which makes training ill-posed. It then introduces PSAM, a gradient-based visualization tool, to show that the unlabeled P2P loss causes over-activation around foreground points. Based on this observation, the paper replaces P2P with P2R: each pseudo-point supervises a local circular region, and pixels farther than a radius \\mu from all pseudo-points are treated as confident background. Experiments on ShTech A/B, UCF-QNRF, and JHU++ under 5%, 10%, and 40% labeled protocols, plus unsupervised domain adaptation, report state-of-the-art results and a large speed-up from removing the Hungarian algorithm.","tokens_in":20109,"tokens_out":6665,"duration_ms":69240,"significance":"If the results hold, the paper makes a useful contribution: it identifies a concrete failure mode of pseudo-labeling for point-based counters, provides a correct mathematical diagnosis of the missing background gradient, and offers a simple and faster alternative. The method is clearly specified, the code is released, and the experiments cover four datasets and UDA. The derivation in Supplement Sec. 11 is sound. The main weakness is that the confident-background term in Eq. (23) relies on an untested assumption about pseudo-label recall; the reported gains could be partially due to suppressing missed people as background. The absence of error bars or multiple-seed runs also tempers the state-of-the-art claims.","major_comments":[{"comment":"The term (1_n - \\beta) assigns confidence 1 and pseudo-label 0 to every pixel farther than \\mu from all extracted pseudo-points. This is valid only if the teacher's pseudo-labels extracted with the 0.5 threshold in Eq. (6) have near-perfect recall outside those radii. In sparse or semi-dense crowd regions, a person whose teacher score is below 0.5 and who is not within \\mu of a detected person is silently converted into a confident background example. Because the teacher is an EMA of the student, this error can be self-reinforcing across iterations. The paper does not quantify teacher recall or sensitivity to it; the \\mu ablation in Fig. 6(c) varies the radius with real pseudo-labels and cannot isolate the effect of missing pseudo-points. I request an explicit experiment in which known fractions of ground-truth points are removed before constructing pseudo-labels, with P2R MAE reported as a function of the removal rate, or an alternative formulation that gates the background confidence term on a teacher background-probability estimate.","section":"Sec. 5.2, Eq. (23)"},{"comment":"All reported numbers are single runs without error bars or multiple seeds, and the key hyperparameters \\tau, \\eta, \\mu, and \\alpha are selected on the same datasets and protocols used for the final comparisons. This creates a selection-bias risk for the state-of-the-art claims (e.g., 5% ShTech A MAE 69.9 vs. 83.7 for OT-M). I recommend reporting mean and standard deviation over at least three seeds and stating how hyperparameters are chosen, ideally on a held-out validation split rather than the test protocol.","section":"Sec. 6, Tables 1-2 and Fig. 6"}],"minor_comments":[{"comment":"In the 10% protocol, the MT [59] row lists MAE 319.3 and MSE 94.5 for ShTech A; this ordering is impossible for non-negative errors and appears to be a typo or table-formatting error. Please correct and re-check all numbers in that row.","section":"Table 1"},{"comment":"The sentence 'map to \\hat p_t by M_st' is hard to parse; please define the vector z = M_st \\zeta explicitly and make the dimensions of \\zeta and \\beta clear in Eq. (23) and Algorithm 2.","section":"Eq. (7) and Algorithm 2"},{"comment":"The algorithms do not specify behavior when the pseudo-label set is empty (m=0) or when a matched region contains no pixel with positive score; please add a guard or discuss how such cases are handled in practice.","section":"Algorithms 1-2"},{"comment":"The 68x speed-up of P2R over P2P for loss computation is reported without specifying hardware or whether the timing includes GPU transfer and Hungarian matching; please add these details.","section":"Sec. 6.1, efficiency comparison"},{"comment":"The inverse-sigmoid versus identity ablation is reported on ShTech A only; if space permits, adding a second dataset would make the choice of S(\\cdot) more convincing.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The core derivation is sound and the proposed method is well specified. The main risk is the unvalidated background-reliability assumption in Eq. (23) and the lack of statistical robustness in the experimental claims. These are addressable with additional experiments, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this paper has a real, transferable idea: in semi-supervised point-based counting, the standard pseudo-label loss has zero background term under point-to-point matching, so the unlabeled loss only reinforces foreground. The derivation of Eq. (9) is correct, and the proposed point-to-region loss is a sensible repair that propagates confidence to local neighborhoods and marks distant pixels as reliable background. Second, the paper is better than the current abstract suggests: the PSAM visualization is not window dressing; it gives a concrete diagnosis of over-activation near foreground points, and P2R fixes the observed failure.\n\nWhat is genuinely new: replacing one-to-one Hungarian matching with local-region assignment in both supervised and semi-supervised training, and using region-level confidence masks for pseudo-labels. It also removes the Hungarian bottleneck, which matters for dense images. The experimental sweep is broad: four benchmarks, three label protocols, plus UDA. The gains over OT-M are large and consistent (e.g., 5% ShanghaiTech A MAE 69.9 vs 83.7). The code is public, which helps.\n\nThe soft spots, in order of seriousness. The background term in Eq. (23) treats every pixel farther than mu from any extracted pseudo-point as confident background. If teacher recall is imperfect—and it always is in dense crowds—missed people outside the pseudo-point neighborhoods are trained as background. Because the teacher is an EMA of the student, this error can be self-reinforcing. The paper never quantifies teacher false negatives or tests sensitivity to missing pseudo-points. This is a real caveat, but I don't think it sinks the paper: the labeled loss and foreground positives still provide countervailing signal, and the ablation of mu (Fig. 6c) shows MAE is fairly flat over a wide radius range, which suggests moderate robustness. Still, an ablation that artificially drops points from pseudo-labels would substantially strengthen the claim. Second, there are no error bars or multiple-seed results; all hyperparameters are tuned on the same benchmark validation sets. That's common practice in this subfield, but the absence of variance reporting makes it hard to know how much of the 5-15 MAE gap is real. Third, the PSAM method is adapted from ODAM; the novelty is in the application, not the visualization machinery. That's fine, but the paper should say so more plainly.\n\nBottom line: this is a solid, practical contribution, not a breakthrough. The central failure analysis is correct, the fix is simple and well-motivated, and the experiments support the main claim conditional on the teacher-recall caveat. I'd send it to review, and I'd ask the authors to address the recall sensitivity directly.","headline":"P2R is a genuinely useful fix for a real failure in semi-supervised point-based crowd counting; the central diagnosis is correct and the paper deserves refereeing, with one unquantified vulnerability around teacher recall.","tokens_in":20599,"tokens_out":2491,"would_cite":true,"duration_ms":25760,"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":"Point-to-region matching makes semi-supervised point-based crowd counting work.","keywords":["semi-supervised crowd counting","point detection","point-to-region matching","pseudo-labeling","point-specific activation map","crowd localization","domain adaptation","point-based counting"],"falsifier":"A direct test: in a fixed semi-supervised run, randomly drop a known fraction (say 10-20%) of the teacher's pseudo-points before the P2R loss is computed. If P2R's final MAE rises by roughly the dropped fraction compared to the full-pseudo-label run, that confirms the reliable-background assumption is the load-bearing part and that imperfect teacher recall would hurt. If the MAE is unchanged, the background confidence term is not the decisive mechanism.","tokens_in":19594,"feed_emoji":"🧍","tokens_out":7044,"duration_ms":64219,"temperature":0.7,"pith_summary":"This paper tries to establish that point-based crowd counters can be trained with only a few annotated images plus large amounts of unlabeled data, provided pseudo-label confidence is spread over local regions instead of single points. The authors show that the usual point-to-point matching is ill-posed for semi-supervised learning: the background term of the cross-entropy loss vanishes for unlabeled data, so the model is trained with only positive examples and collapses. Their point-to-region (P2R) matching gives every pixel near a pseudo-point that point's confidence and explicitly marks distant pixels as reliable background, restoring the missing gradients. Under this scheme, semi-supervised counting reaches state-of-the-art mean absolute errors on ShTech A/B, UCF-QNRF, and JHU++, and the same loss transfers to unsupervised domain adaptation. A byproduct is that the Hungarian matching step is no longer needed, making loss computation roughly 68 times faster.","feed_headline":"Point-to-region pseudo-labels fix semi-supervised crowd counting","feed_subtitle":"Giving every pixel around a pseudo-point that point's confidence lets point-based counters learn from unlabeled crowds.","key_machinery":"The load-bearing mechanism is the point-to-region (P2R) matching matrix $M = M_f \\odot (\\beta 1_m^\\top)$, where $M_f$ is a many-to-one nearest-point assignment from pixels to pseudo-points and $\\beta$ is the indicator of pixels within radius $\\mu$ of any point. A per-column minimum of the cost matrix selects the representative foreground pixel in each region, and the confidence matrix $Z = \\mathrm{diag}(M_{st}\\zeta + (1-\\beta))$ propagates each pseudo-point's confidence to its whole region while marking far pixels as reliable background. The binary cross-entropy loss is then computed with $Z$ as weights, so the background term is no longer zero. PSAM, a gradient-based per-point activation map, is the diagnostic that shows the over-activation growth and motivates the region-based supervision.","core_discovery":"The paper's central claim is that the observed training collapse in semi-supervised point-based counting is caused by confidence never reaching background pixels. With point-to-point matching, the pseudo-label confidence vector is applied only to matched foreground pixels, so the second (background) term in the binary cross-entropy for unlabeled data is identically zero; the proof is given through equations (25)-(34), and the loss reduces to supervision with only positive samples. Using the point-specific activation map (PSAM), the authors visualize that the decoder's activation regions around each predicted person grow during this ill-posed training until neighboring pixels are misread as people. The P2R strategy replaces each one-pixel match with a local region: every pixel in the region inherits the pseudo-point's confidence, and pixels farther than a radius $\\mu$ from all pseudo-points are assigned confidence 1 with label 0. This re-introduces the background supervision that P2P drops, stabilizes training, and produces the reported gains in semi-supervised counting (5% ShTech A MAE 69.9 vs 83.7 for OT-M; 40% ShTech A MAE 55.6 vs 70.7) and in unsupervised domain adaptation.","pith_inferences":["Beyond the paper: the reliable-background rule means any person missed by the teacher (score below the 0.5 extraction threshold) is explicitly supervised as background, so P2R's performance should degrade as teacher recall drops; the paper does not measure this sensitivity.","Beyond the paper: because the ablation shows MAE is flat as the radius $\\mu$ grows from 32 to 256, P2R should transfer to other point-detection tasks such as cell or animal counting with little hyperparameter tuning.","Beyond the paper: PSAM is a general diagnostic for dense localization; it could be used in other self-training pipelines to detect when background supervision is being lost, not just in crowd counting."],"forward_implications":["With only 5% of labels, P2R reaches 69.9 MAE on ShTech A, beating the previous semi-supervised method OT-M's 83.7, and with 40% labels it reaches 55.6 vs 70.7.","Under the fully-supervised 100% protocol, P2R improves over P2P on all four benchmark datasets (e.g., ShTech A MAE 51.02 vs 52.74) while cutting loss computation time from 0.4307s to 0.0064s per image.","The same loss transfers to unsupervised domain adaptation, where P2R surpasses prior UDA methods on four cross-domain protocols (e.g., B->Q MAE 139.5 vs 194.9 for the best tabulated UDA baseline).","PSAM comparisons show the P2R-trained model keeps activations concentrated around true foreground pixels instead of spreading them onto neighbors, which is the mechanism behind the stable training."],"supporting_citations":[{"why":"Supplies the point-to-point matching baseline and the point-based counter architecture that P2R replaces.","marker":"[57]"},{"why":"The semi-supervised counting baseline (density-to-point pseudo-labels) whose performance P2R improves upon in the same experimental protocols.","marker":"[26]"},{"why":"Provides the generalized loss and coordinate-based localization perspective used to simplify the point counter and define matching costs.","marker":"[61]"},{"why":"Contributes the confidence-threshold pseudo-labeling rule that P2R adapts to spread confidence over regions.","marker":"[56]"},{"why":"Provides the mean-teacher EMA update that generates pseudo-labels in the semi-supervised framework.","marker":"[59]"},{"why":"Provides the pseudo-labeling and self-training idea underlying the confidence-based unlabeled loss.","marker":"[15]"},{"why":"The gradient-based visual explanation method that PSAM extends to point-level predictions.","marker":"[48]"},{"why":"Instance-specific gradient-based explanations for detection that PSAM adapts to point-based counting.","marker":"[74]"}],"fun_headline_variants":["Point-to-region loss stops false positives in crowd counting","Confidence spillover fixes semi-supervised crowd counting","New loss lets point counters learn from unlabeled crowds","Point-to-region supervision prevents pseudo-label collapse"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that every pixel farther than a fixed radius from every pseudo-point really is background; if the teacher misses a person (score below the 0.5 extraction threshold), that missed person is trained as background and the mistake can be reinforced.","fun_headline_variants_meta":{"raw":{"variants":["Point-to-region loss stops false positives in crowd counting","Confidence spillover fixes semi-supervised crowd counting","New loss lets point counters learn from unlabeled crowds","Point-to-region supervision prevents pseudo-label collapse"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000509,"raw_usage":{"total_tokens":2546,"prompt_tokens":1077,"completion_tokens":1469,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":693,"completion_tokens_details":{"reasoning_tokens":1406}},"tokens_in":693,"tokens_out":1469,"duration_ms":10895,"temperature":1.0,"reasoning_tokens":1406,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:19:11.986511+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test: in a fixed semi-supervised run, randomly drop a known fraction (say 10-20%) of the teacher's pseudo-points before the P2R loss is computed. If P2R's final MAE rises by roughly the dropped fraction compared to the full-pseudo-label run, that confirms the reliable-background assumption is the load-bearing part and that imperfect teacher recall would hurt. If the MAE is unchanged, the background confidence term is not the decisive mechanism.","supporting_citations":[{"cited_title":"Rethinking counting and localization in crowds: A purely point-based framework","cited_arxiv_id":null,"evidence_quote":"Supplies the point-to-point matching baseline and the point-based counter architecture that P2R replaces."},{"cited_title":"Optimal transport mini- mization: Crowd localization on density maps for semi- supervised counting","cited_arxiv_id":null,"evidence_quote":"The semi-supervised counting baseline (density-to-point pseudo-labels) whose performance P2R improves upon in the same experimental protocols."},{"cited_title":"A generalized loss function for crowd counting and localization","cited_arxiv_id":null,"evidence_quote":"Provides the generalized loss and coordinate-based localization perspective used to simplify the point counter and define matching costs."},{"cited_title":"Fixmatch: Simplifying semi-supervised learning with consistency and confidence","cited_arxiv_id":null,"evidence_quote":"Contributes the confidence-threshold pseudo-labeling rule that P2R adapts to spread confidence over regions."},{"cited_title":"Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results.Advances in neural information processing systems, 30, 2017","cited_arxiv_id":null,"evidence_quote":"Provides the mean-teacher EMA update that generates pseudo-labels in the semi-supervised framework."},{"cited_title":"Pseudo-label: The simple and effi- cient semi-supervised learning method for deep neural net- works","cited_arxiv_id":null,"evidence_quote":"Provides the pseudo-labeling and self-training idea underlying the confidence-based unlabeled loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Instance-specific gradient-based explanations for detection that PSAM adapts to point-based counting."}],"review_version":1}