{"id":"fd005d58-bfb8-4baf-ba7b-411d7fd98b14","arxiv_id":"2509.03170","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Count2Density maps per-image crowd counts into spatial density maps via an EMA historical bank and hypergeometric sampling, eliminating the need for point annotations.","lead":"Count2Density is a training pipeline that produces crowd density maps using only per-image total counts, not per-person point labels. It builds pseudo-density labels from past model predictions plus a saliency prior, then self-trains on them, reporting large gains over cross-domain methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablations use only MAE, so the central claim of recovering spatial density from counts is not established; spatial metrics are needed across configurations.","rationale":"The reader's weakest assumption focuses on the saliency prior misaligning and the self-training loop amplifying the mismatch. I agree that the initial spatial prior is a fragile point, but the paper's own Table 6 already shows that MAE is almost insensitive to the initialisation (149.2 vs 151.4 vs 152.5). This suggests the more precise issue is not simply that saliency can be wrong, but that the evaluation never measures whether the proposed components contribute to spatial accuracy at all. The central claim is explicitly about spatial information, yet every ablation table uses global counting errors that do not penalize spatial misplacement. A method that always outputs a centered blob of the right total mass would do well on MAE but would fail on subregion counting and localization. Because those spatial metrics are only reported for the full pipeline, the current paper cannot rule out the possibility that the spatial quality is inherited from the BASNet prior or even from incidental image statistics, rather than learned from count labels via the proposed mechanism. The cross-domain comparison is also not a clean test of spatial inference because Count2Density sees target-domain count labels while cross-domain baselines see none; the reported margin is therefore confounded by supervision level. For these reasons, the verdict should remain CONDITIONAL, but the condition should be tightened: the authors must provide spatial-metric ablations and a fair comparison against a count-only baseline. This is an honest non-finding in the sense that the method may well work as claimed, but the current evidence does not yet establish the central spatial-inference claim.","tokens_in":13292,"tokens_out":7438,"duration_ms":87615,"concrete_test":"Run the Table 5 ablations (full pipeline; -saliency init; -contrastive reg.; -historical bank; etc.) on UCF-QNRF and report the same spatial metrics as Table 3/4: per-tile subregion MAE and localization F1. Specifically, compare 'Unsup. bas-net', 'Centred blob', and 'None' initialisations. If F1 and subregion MAE do not degrade substantially (e.g., stay within 5%) when the saliency prior is removed or replaced by a centered blob, then the proposed bank/sampling loop itself is not the source of spatial accuracy; if they do degrade, then the spatial claim rests on the external BASNet prior rather than on count-level learning. Either outcome would clarify whether the central claim holds.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that Count2Density recovers quantitative spatial density information from count-level supervision. The evidence for spatial quality is Tables 3-4 (subregion counting and localization F1), but these are reported only for the full pipeline. The component ablations in Table 5 and initialization ablations in Table 6 are evaluated with MAE/MSE only, which are global count errors and are insensitive to where the predicted density is placed. This is particularly problematic because the only external spatial signal in the pipeline is the BASNet saliency initialisation of the historical map bank (Section 3.2). Table 6 shows that replacing BASNet with a centered blob or no initialisation changes MAE by <1-2% (151.4 and 152.5 vs 149.2). If the spatial prior were actually driving spatial accuracy, one would expect a much larger MAE change; conversely, if it is not, then the source of spatial signal in the self-training loop is unclear. Either way, the paper's ablations cannot discriminate: a configuration could achieve similar MAE while producing spatially wrong density maps. Thus the claim that the proposed EMA bank, hypergeometric sampling, and contrastive regulariser 'allow Count2Density to infer detailed spatial density information' is not supported by the reported experiments. In addition, the cross-domain comparison in Table 2 is confounded: Count2Density is trained with target-domain count labels, whereas cd-cc and bla use no target labels at all, so the 49.1 MAE reduction may reflect the extra supervision rather than spatial inference.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Count2Density, a crowd density estimation pipeline trained only with image-level count annotations. It generates pseudo-density maps by sampling a number of points equal to the count label from a Historical Map Bank that stores an EMA of past predictions, initialized with an unsupervised saliency estimator. A contrastive spatial regularizer is added to separate crowd and background features. The method is evaluated on UCF-QNRF, ShanghaiTech-A/B, JHU-Crowd++, and NWPU-Crowd using multiple backbones, and compared with cross-domain and semi-supervised baselines; ablations and qualitative pseudo-map evolution are also provided. The main claim is that count-level supervision, combined with the bank and regularizer, is sufficient to recover quantitative spatial density information, enabling subregion counting and localization.","tokens_in":13636,"tokens_out":6830,"duration_ms":69214,"significance":"If the central claim is substantiated, the contribution is significant: it would reduce annotation cost to count-level labels while retaining density-map outputs, a step beyond regression-based weakly supervised counting that does not support subregion counting. The paper's strengths are its extensive empirical scope (four datasets, multiple backbones), the clear integration of a historical bank and contrastive regularizer, the inclusion of subregion counting and localization evaluation (Tables 3-4), and honest ablation/initialization studies (Tables 5-6). However, the central claim about recovering spatial information is currently supported only by full-pipeline spatial metrics; the component ablations do not measure spatial accuracy, and several headline comparisons are not controlled for the supervision actually used. The method is novel and plausible, but the evidence as presented is not yet sufficient to establish the advertised superiority.","major_comments":[{"comment":"The central claim is that Count2Density 'infers detailed spatial density information from count-level supervision.' The ablations that attribute this to the Historical Map Bank, saliency initialization, and contrastive regularizer are evaluated only with MAE/MSE (Tables 5 and 6), which are global count errors and are insensitive to the spatial placement of density. A configuration could achieve the same MAE while producing a centered blob instead of the true crowd distribution. Spatial metrics (SSIM/PSNR, subregion counting, localization F1) are reported only for the full pipeline (Tables 3-4), not for ablated variants. This is especially consequential because Table 6 shows that replacing BASNet with a centered blob changes MAE by only ~1.5%, so the experiments cannot discriminate whether the initial spatial prior actually contributes to spatial accuracy. Please report subregion counting","section":"Section 4.2, Tables 5 and 6"},{"comment":"The comparison with cd-cc and bla is confounded by annotation budget: Count2Density is trained with target-domain count labels for every image, whereas cd-cc and bla use no target-domain labels at all. The advertised 49.1 MAE reduction on UCF-QNRF (198.3 vs 149.2) is therefore not an equivalent-setting comparison. The text's remark that cross-domain methods use source location labels does not remove this confound, because target count labels are still a form of target supervision. Moreover, on ShT-B the cross-domain baselines are better (cd-cc 11.4, bla 11.9 vs Count2Density 15.5), which is acknowledged in the text but not in the abstract. Please either add a target-unsupervised variant of Count2Density or recast the comparison explicitly as 'using count-level target annotations versus no target annotations.'","section":"Section 4.1, Table 2, 'Comparisons with Cross-Domain Approaches'"},{"comment":"The claim that Count2Density 'achieves better results than recent state-of-the-art approaches in semi-supervised settings' is not supported for the count-only configuration. In Table 2, Count2Density(man) has MAE 149.2 on UCF-QNRF, while irast achieves 135.6 and pal 128.1; on ShT-A several semi-supervised methods (l2r 90.3, irast 86.9, matt 80.1, pal 72.7, cu 70.7) beat Count2Density's 91.5; on ShT-B, cu (9.7) beats 15.5. The outperformance appears only for the Count2Density(Semi-sup.) variant that additionally receives 5-10% location-level labels. Please separate the claims: one for count-only training and one for the semi-supervised variant with a subset of location labels.","section":"Section 4.1, Table 2, 'Comparisons with Semi-Supervised Approaches'"},{"comment":"The pseudo-label generation is a self-training loop: the model's own predictions (averaged in the Historical Map Bank) generate the training targets, and the only external spatial grounding is the saliency initialization. Count labels constrain only the total mass, not the location of the density. The paper does not analyze the confirmation-bias risk beyond invoking EMA and the bank. Given Table 6's near-insensitivity of MAE to the initialization, it remains unclear what prevents the loop from converging to a count-accurate but spatially generic prediction. I suggest an additional experiment in which the bank is frozen at the BASNet initialization (no online updates) to isolate the contribution of the self-training dynamics, and reporting spatial metrics for that variant.","section":"Section 3.2, Eq. (2) and Figure 3"}],"minor_comments":[{"comment":"The summation bounds use W for the horizontal index and H for the vertical index; please make the notation consistent with the map dimensions (H x W) and clarify the coordinate convention.","section":"Eq. (1)"},{"comment":"The text says 'identifying local minima' when detecting people from density maps; this should presumably be local maxima, since people correspond to peaks in density.","section":"Table 4 description"},{"comment":"The sampling procedure is described as a 'hypergeometric distribution' and 'equivalent to performing yi Bernoulli samplings without replacement.' Weighted sampling without replacement is the more standard and precise description; the hypergeometric distribution typically refers to the number of successes in draws without replacement, not to the location-sampling operation itself.","section":"Section 3.2"},{"comment":"Minor grammatical issues: 'the such improvement is marginal' should be 'such improvement is marginal'; 'corrispond' should be 'correspond'; 'BASNET' should be 'BASNet' for consistency.","section":"Section 4.2"},{"comment":"Some rows have inconsistent symbol usage (e.g., the † marker is defined only in the caption for irast). Please ensure all markers are explained and that the 'Semi-supervised' vs 'Cross-Domain' categories are visually distinct for readers.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, this paper is worth a real look: it trains a crowd density estimator from image-level counts alone and still produces density maps that support subregion counting. That is genuinely new in this literature, where regression-based weak supervision predicts counts without spatial information. The pipeline is also a real combination, not a token novelty: an EMA historical map bank of past predictions, initialised with an unsupervised saliency estimator, used to sample pseudo-point locations from the normalised map, plus a contrastive regulariser over crowd/background features. I credit the authors for evaluating on four datasets, multiple backbones, and an honest modified count-only irast baseline, and for acknowledging that they lose to cross-domain methods on ShanghaiTech-B.\n\nThe soft spots, in proportion. The stress-test concern lands: the ablations in Tables 5 and 6 only report MAE/MSE, which are global count errors. A predicted density map can be spatially wrong and still integrate to the right count. Since the paper's central claim is that Count2Density recovers quantitative spatial information, the component ablations should report spatial metrics — subregion counting error, SSIM, or F1. Without that, Table 6 actually undercuts an easy story: replacing BASNet with a centred blob or no initialisation changes MAE by less than 2%. That suggests count accuracy is not driven by the spatial prior, and it leaves unclear where the spatial signal actually comes from. The spatial evidence in Tables 3 and 4 is only for the full pipeline, so it cannot attribute the behaviour to any component. This is a significant missing analysis, not a fatal flaw.\n\nThe cross-domain comparison is also partly confounded: Count2Density uses target-domain count labels, while cd-cc and bla use none. The headline 49.1 MAE reduction on UCF-QNRF likely reflects that extra supervision as much as the method. The abstract overclaims by saying it beats recent semi-supervised SOTA across datasets; on ShT-A, pal and cu do better, and on ShT-B cross-domain methods do better. The text is more honest than the abstract.\n\nMinor point: calling the sampling 'hypergeometric' is a bit off; it is sampling without replacement from a categorical distribution.\n\nWho this is for: researchers in crowd counting, label-efficient vision, and anyone needing density maps from cheap annotations. It deserves a serious referee. The core idea is plausible and novel, the experiments are broad, and the missing spatial ablations are fixable. I'd send it to review and ask for those ablations, plus code.","headline":"A genuinely new count-only density-estimation pipeline, but the spatial-quality claims rest on ablations that never measure spatial error — worth refereeing with that gap fixed.","tokens_in":14122,"tokens_out":3562,"would_cite":true,"duration_ms":40995,"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":"Crowd density maps from counts alone, no point labels","keywords":["crowd density estimation","count-level supervision","pseudo-density maps","historical map bank","contrastive learning","weakly supervised learning","subregion counting","saliency initialization"],"falsifier":"Take a crowd dataset where people are not visually salient (low contrast, heavy occlusion, or cluttered backgrounds), train Count2Density with only count labels, and compare its density maps against a count-regression baseline and against the same method with the bank seeded by ground-truth point locations; if MAE and subregion-counting error do not improve over regression, the central claim that spatial information can be recovered from counts fails in that regime.","tokens_in":13219,"feed_emoji":"👥","tokens_out":4022,"duration_ms":44129,"temperature":0.7,"pith_summary":"The paper claims that a training pipeline can produce spatially meaningful crowd density maps using only the total number of people per image as supervision. It does so by maintaining a Historical Map Bank of past density predictions, initialized with an unsupervised saliency map, and sampling as many pseudo-points as the count label from that bank via a hypergeometric distribution. A contrastive regularizer pushes the model to separate crowd from background features. If correct, this removes the main annotation bottleneck in density estimation and enables subregion counting without location-level labels. The paper reports large error reductions over cross-domain adaptation baselines, for example from 198.3 to 149.2 Mean Absolute Error on UCF-QNRF.","feed_headline":"Crowd density maps from counts alone, no point labels","feed_subtitle":"Count labels feed a memory bank that samples pseudo-points, yielding density maps that beat cross-domain baselines.","key_machinery":"Historical Map Bank: a per-image exponential moving average of the model's own past density predictions, initialized with the BASNet saliency estimator, which serves as a probability prior for sampling y_i locations with a hypergeometric distribution to generate the pseudo-density map used for supervision. The contrastive spatial regularizer selects crowded versus background patches from the predicted density map and optimizes a contrastive loss on feature representations, making the learned features spatially consistent.","core_discovery":"On its own terms, Count2Density establishes that count-level supervision is sufficient to learn density maps with quantitative spatial information. The key move is to convert each image's total count into a pseudo-density map: sample exactly y_i locations from a normalized historical average of past predictions, itself initialized by unsupervised saliency, place point masses at those locations, and train the density estimator against this pseudo-map. The bank's exponential moving average acts as an ensemble that dampens confirmation bias. A self-supervised contrastive term encourages the feature extractor to treat crowded areas and background as distinct, so spatial structure survives in the","pith_inferences":["Editorial inference: the sampling-from-bank recipe is generic; any task where total mass is known but locations are not, such as counting cells, animals, or vehicles from aggregate counts, could use the same EMA bank plus hypergeometric sampling.","Editorial inference: if saliency is a poor prior for the target domain, the self-training loop could amplify the mismatch; the paper's own ablations suggest the bank still helps with trivial initialization, so the loop itself does substantial work.","Editorial inference: storing one map per training image limits scalability, so a compressed or sampled bank, for example storing prototypes or features rather than full maps, is a natural extension.","Editorial inference: the stronger performance on dense scenes suggests the hypergeometric sampling prior is better matched to crowded imagery; sparse scenes may need a different sampling distribution."],"forward_implications":["Density estimation models can be trained on datasets that only have total counts, removing the point-annotation bottleneck.","Predicted density maps from count-only training support subregion counting by integrating over image tiles.","The pseudo-label bank approach works across multiple density-estimation backbones, including NCC, Bayesian loss, generalized loss, and MAN.","In a semi-supervised variant with 5–10% location labels, the method outperforms several recent semi-supervised approaches on benchmark datasets.","Cross-domain adaptation baselines that use location labels in a source domain are outperformed on dense-scene benchmarks such as UCF-QNRF."],"supporting_citations":[{"why":"Supplies the integration formula that turns a predicted density map into a count, the basis for both global and subregion counting.","marker":"[1]"},{"why":"BASNet is the unsupervised saliency estimator used to initialize the Historical Map Bank with an initial spatial prior.","marker":"[31]"},{"why":"Noisy-label learning motivates the exponential moving average bank as a way to reduce confirmation bias in self-training.","marker":"[28]"},{"why":"Contrastive predictive coding provides the contrastive learning formulation used by the spatial regularizer.","marker":"[32]"},{"why":"Supplies the temperature value used in the contrastive loss.","marker":"[35]"},{"why":"Cross-domain baseline (BLA) that Count2Density outperforms on ShanghaiTech-A and UCF-QNRF, used as a main comparison.","marker":"[13]"},{"why":"Multifaceted Attention Network backbone that yields the best Count2Density results and demonstrates backbone-agnostic integration.","marker":"[10]"}],"fun_headline_variants":["Counting people without dots: density maps from totals only","No point labels needed: crowd density from headcounts","Turn headcounts into density maps, no dot annotations","Crowd density from count labels, not point locations","Count-only training yields spatial density maps for crowds"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The whole training loop leans on the assumption that the saliency-seeded Historical Map Bank points at pixels where people actually are; if that spatial prior is wrong in a new domain, the self-training loop could reinforce the error instead of correcting it.","fun_headline_variants_meta":{"raw":{"variants":["Counting people without dots: density maps from totals only","No point labels needed: crowd density from headcounts","Turn headcounts into density maps, no dot annotations","Crowd density from count labels, not point locations","Count-only training yields spatial density maps for crowds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000551,"raw_usage":{"total_tokens":2494,"prompt_tokens":803,"completion_tokens":1691,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":1627}},"tokens_in":547,"tokens_out":1691,"duration_ms":14311,"temperature":1.0,"reasoning_tokens":1627,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:05:31.394940+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a crowd dataset where people are not visually salient (low contrast, heavy occlusion, or cluttered backgrounds), train Count2Density with only count labels, and compare its density maps against a count-regression baseline and against the same method with the bank seeded by ground-truth point locations; if MAE and subregion-counting error do not improve over regression, the central claim that spatial information can be recovered from counts fails in that regime.","supporting_citations":[{"cited_title":"Lempitsky, A","cited_arxiv_id":null,"evidence_quote":"Supplies the integration formula that turns a predicted density map into a count, the basis for both global and subregion counting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BASNet is the unsupervised saliency estimator used to initialize the Historical Map Bank with an initial spatial prior."},{"cited_title":"Litrico, A","cited_arxiv_id":null,"evidence_quote":"Supplies the temperature value used in the contrastive loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cross-domain baseline (BLA) that Count2Density outperforms on ShanghaiTech-A and UCF-QNRF, used as a main comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Multifaceted Attention Network backbone that yields the best Count2Density results and demonstrates backbone-agnostic integration."}],"review_version":1}