{"id":"3cf94161-e461-405c-a9e3-c81314312d35","arxiv_id":"1908.03684","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"A soft-assignment Bayesian loss that supervises expected counts at each annotated point improves crowd counting accuracy over pixel-wise density-map regression on four benchmarks.","lead":"This paper proposes a new training loss, called Bayesian loss, for counting people in crowded images from point annotations. Instead of training a network to reproduce a Gaussian-smoothed density map, the loss supervises the expected count assigned to each annotated head.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: fixed-Gaussian posterior is a modeling choice, and L_Bayes+ directly bounds count error; only single-run metrics remain unquantified.","rationale":"I read the paper's central claim as an empirical one: replacing pixel-wise density regression with the expectation-based loss in Eqs. (5)-(12) improves count accuracy on four benchmarks. The formal structure of the loss supports this: the posterior normalization in Eq. (8) makes the loss an L1 upper bound on count error, which I verified by summing the per-label expectations. This is independent support that the reader's weakest assumption about Gaussian misspecification, while valid for per-head localization, does not transfer to the global count claim. The experiments include a controlled baseline, ablations on sigma and d, cross-dataset transfer, and a different backbone; all point in the same direction. What is missing is any repeated-run variance estimate; without it, the magnitudes of the 8-17% improvements are not statistically quantified. This is a common reporting limitation and does not change the accept-level verdict, but it is the only place where a skeptical reader could reasonably push.","tokens_in":12764,"tokens_out":17950,"duration_ms":219018,"concrete_test":"Re-run the VGG-19 BASELINE and BAYESIAN+ training on UCF-QNRF with at least three random seeds and report mean plus/minus standard deviation for MAE and MSE; if the 17-point MAE gap (106.8 vs 88.7) is not larger than the combined seed variance, the 'substantial improvement' claim would need to be softened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that L_Bayes+ improves count accuracy over the baseline. I find no load-bearing flaw in this argument. For any image, because the posteriors in Eq. (8) sum to one over labels n=0..N, the estimated count C = sum_m D_est(x_m) satisfies C - N = sum_n (E[c_n]-1) + E[c0]. Hence |C-N| <= sum_n |1-E[c_n]| + |E[c0]| = L_Bayes+. The loss is therefore an upper bound on the per-image absolute count error, so even if the fixed Gaussian likelihood in Eq. (3) is misspecified for some head sizes or densities, minimizing the loss is still a sound surrogate for the quantity actually evaluated. The reader's concern that a correct density map could yield E[c_n]<1 is real for individual heads in dense clusters, but the network can redistribute density among pixels; the sum identity protects the global count. The only unquantified risk is statistical: Table 1 reports single runs with no variance, so the size of the claimed improvements over BASELINE is not established to be beyond run-to-run noise. This is a reporting gap, not a correctness defect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Bayesian loss for crowd counting with point supervision. Instead of regressing a fixed Gaussian-smoothed density map pixelwise, the loss computes, for each annotated head, an expected count by summing the predicted density weighted by the posterior probability that each pixel belongs to that head (Eqs. 5 and 10). A background label is introduced via a dummy background point (Eq. 13) to suppress density away from heads, yielding the enhanced loss L_Bayes+ (Eq. 12). Experiments on UCF-QNRF, ShanghaiTech A/B, and UCF CC50 show consistent MAE/MSE improvements over a VGG-19 baseline and favorable comparison with prior state-of-the-art methods, with the largest gain on UCF-QNRF. Ablations analyze the effect of σ, d, annotation noise, cross-dataset transfer, resolution limits, and backbone choice.","tokens_in":13056,"tokens_out":14987,"duration_ms":144417,"significance":"The contribution is conceptually clean and practically useful. The key theoretical observation, though not stated explicitly, is that because the posteriors sum to one, the per-image absolute count error is upper-bounded by L_Bayes+ (the sum of |1-E[c_n]| and |E[c0]|), so minimizing the loss is a sound surrogate for the evaluation metric. The paper also provides code, uses a standard backbone without extra detectors or multi-scale modules, and includes extensive ablations including robustness to annotation noise and cross-dataset generalization. If the reported gains hold under repeated runs, the method is a meaningful improvement over pixel-wise density-map regression, especially on the highly challenging UCF-QNRF dataset.","major_comments":[{"comment":"All reported numbers are from a single training run per configuration, with no standard deviation, confidence interval, or significance test. Because the paper's central claim is that the Bayesian loss yields 'substantial improvements' over BASELINE, and some improvements are modest (e.g., ShanghaiTechB MAE 8.5 to 7.7, UCF CC50 251.6 to 229.3), the reader cannot assess whether these differences exceed run-to-run variation. Please report means and standard deviations over at least 3-5 independent runs for the main tables (Tables 1-4) and for the ablation curves (Figs. 4-6), or provide a paired statistical test (e.g., bootstrap over test images) for the headline comparisons. This is a reporting gap rather than a flaw in the derivation, but it is load-bearing for the empirical claims.","section":"Sec. 4.4, Table 1"}],"minor_comments":[{"comment":"The authors should explicitly state the inequality |C - N| ≤ sum_n |1 - E[c_n]| + |E[c0]|, which follows from Eqs. (7), (10)-(12). This upper-bound property is a strong theoretical justification for the loss and deserves to be highlighted rather than left implicit.","section":"Sec. 3.2, after Eq. (7)"},{"comment":"The dummy background point definition involves division by ||x_m - z_n^m||, which is undefined when a pixel coincides with the nearest head point. Please specify how this case is handled in the implementation (e.g., by adding a small epsilon).","section":"Sec. 3.3, Eq. (13)"},{"comment":"The paper does not report the number of training epochs, batch size, or learning-rate schedule. Including these details would improve reproducibility, especially since the baseline is already stronger than several previously published methods.","section":"Sec. 4.3, Implementation Details"},{"comment":"There is a typo in 'additational parameter' and a repetition in 'wide range of of values of d'; these should be corrected.","section":"Sec. 4.5"},{"comment":"The x-axis label 'Deviation' should be clarified as 'annotation deviation (percentage of image height)' to match the description in the text.","section":"Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically sound and the proposed loss is well motivated, with a nice theoretical property that is not fully exploited in the presentation. The main barrier to acceptance is the lack of multi-run variance/statistical support for the central empirical claim. If the authors can supply error bars or a significance analysis without changing the method, I would be willing to reconsider positively. The paper's scope and writing are otherwise suitable for this journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central claim holds up: replacing pixel-wise density regression with a per-annotation expected-count loss is a genuinely useful idea, and the paper shows it with careful experiments and a clear derivation. The math in Eqs. (3)–(7) is sound, and the stress-test observation is right: for L_Bayes+, the loss is an upper bound on per-image absolute count error, so even if the fixed Gaussian likelihood in Eq. (3) is misspecified for some head sizes or densities, minimizing the loss is still a sensible surrogate for the metric that actually matters. The dummy background point in Eq. (13) is a reasonable fix for pixels far from any head, and the ablation over d shows the method is not brittle to that choice.\n\nThe empirical work is the strongest part. Four datasets, ablations over sigma and d, robustness to annotation noise, cross-dataset generalization, and a check that the benefit is not tied to one backbone. The paper also releases code, which makes the result reproducible in principle. The comparison against a same-backbone baseline is the right way to isolate the loss's contribution, and the gains—especially the 15% MAE improvement on UCF-QNRF—are large enough to matter.\n\nSoft spots are real but minor. First, Table 1 reports single runs with no variance or repeated-seed statistics. Given the claimed improvements are 8–15%, I cannot tell from the paper how much is beyond run-to-run noise. That is a reporting gap, not a correctness defect, but for a loss-function paper it should be fixed. Second, sigma and d are tuned on a UCF-QNRF validation split and then applied to all datasets; the ablation shows the loss is robust over a wide range of sigma, which mitigates this, but the paper should say clearly how sensitive the headline numbers are to those choices. Third, the “Bayesian” name is generous—this is a likelihood-based soft assignment with a uniform prior, not a full Bayesian treatment—but that is a naming issue, not a scientific one.\n\nThe citation pattern is clean; no self-citation loop, and the related work fairly covers detection, direct regression, density estimation, and hybrid losses. The writing is clear enough that I could follow the method without going back to earlier papers.\n\nWho is this for? Anyone working on crowd counting or point-supervised dense prediction. It deserves a serious referee and, after a revision that adds variance statistics and tightens the hyperparameter discussion, it would be a solid accept. I would bring it to a reading group and would cite it if I were doing counting work.","headline":"A clean, well-tested loss for crowd counting that holds up on inspection; the main weakness is single-run reporting, not the method.","tokens_in":13491,"tokens_out":1077,"would_cite":true,"duration_ms":13548,"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":"Bayesian loss replaces imperfect density targets with count-expectation supervision, improving crowd counting across all tested benchmarks.","keywords":["crowd counting","point supervision","Bayesian loss","density map estimation","soft assignment","background modelling","dense crowd scenes","deep learning"],"falsifier":"Take a crowd image and train one model with every person annotated and a second model with a random half of the annotations removed, holding the network and all other settings fixed: since the Bayesian loss forces the total expected count to equal the number of annotated points, the half-annotated model should underestimate the true count by a factor of roughly two; if it instead recovers the full count, the loss is not enforcing the count supervision the paper claims.","tokens_in":12592,"feed_emoji":"👥","tokens_out":8144,"duration_ms":75860,"temperature":0.7,"pith_summary":"This paper proposes replacing the standard crowd-counting training target—a density map made by smoothing point annotations with a Gaussian kernel—with a Bayesian loss that supervises the expected count at each annotated point. For every pixel, the loss allocates the predicted density among nearby head annotations according to a posterior probability computed from Gaussian likelihoods, and requires each head to account for one unit of count. A background label, built from a dummy point placed a margin away from each head, absorbs density in non-head regions. The paper reports that this single loss, on a standard VGG-19 backbone with no extra detectors or multi-scale modules, improves accuracy over the pixel-wise baseline on UCF-QNRF, ShanghaiTech A/B, and UCF CC 50, and sets a new state of the art on the largest, hardest dataset.","feed_headline":"Bayesian loss cuts crowd-counting error on UCF-QNRF to 88.7","feed_subtitle":"Instead of fitting Gaussian-blurred density maps, it supervises the count at each head point and wins on four benchmarks.","key_machinery":"The load-bearing object is the density contribution probability model, a soft partition of each pixel's estimated density among the annotated heads and a background label, built from Bayes' theorem with equal priors and an isotropic Gaussian likelihood of width $\\sigma$. Its expected-count identity $E[c_n] = \\sum_m p(y_n|x_m) D_{\\mathrm{est}}(x_m)$ converts point annotations into per-head count constraints, and the dummy background point $z_0^m = z_n^m + d (x_m - z_n^m)/\\|x_m - z_n^m\\|_2$ creates a background likelihood $\\mathcal{N}(x_m; z_0^m, \\sigma^2 I)$ so that pixels far from heads are assigned to the background and penalized for producing density. The machinery lets the loss supervise aggregate count expectations rather than pixel values, which is what makes the training target robust to the imperfections of Gaussian-kernel density maps.","core_discovery":"On the paper's own terms, the central discovery is that point annotations are better treated as weak labels for a count-expectation constraint than as sources of a pseudo ground-truth density map. Defining the posterior label probability $p(y_n|x_m) = \\mathcal{N}(x_m; z_n, \\sigma^2 I) / (\\sum_{n'=1}^N \\mathcal{N}(x_m; z_{n'}, \\sigma^2 I) + p(x_m|y_0))$, the expected count associated with annotated head $n$ is $E[c_n] = \\sum_m p(y_n|x_m) D_{\\mathrm{est}}(x_m)$, and the loss $\\mathcal{L}_{\\mathrm{Bayes+}} = \\sum_n F(1 - E[c_n]) + F(0 - E[c_0])$ drives each head's expected count to one and the background's expected count to zero. At inference the posterior probabilities sum to one over all labels, so the total estimate is just the sum of the density map. This formulation removes the need to specify a 'correct' Gaussian-blurred density map, and the paper shows the resulting model is less sensitive to the Gaussian width and more tolerant of annotation noise.","pith_inferences":["Because the loss only needs point locations and a Gaussian likelihood, the same expected-count supervision could transfer to other point-supervised counting problems, such as nuclei or vehicle counting; a direct test would swap the dataset while keeping the network fixed.","The reported robustness to annotation noise implies point labels need not mark exact head centers, which could make the method viable with cheaper, noisier crowd-sourced annotations.","The fixed Gaussian likelihood is the main hand-set component; learning $\\sigma$ per head or conditioning it on scale and occlusion is a natural extension the paper mentions but does not explore, and could close part of the remaining MAE gap.","If annotation coverage is incomplete—some people in the scene are simply not labeled—the loss's count constraint would push the total estimate toward the number of labels rather than the true count, so applying the method to partially annotated data would require an explicit missing-label term."],"forward_implications":["A standard VGG-19 trained with Bayesian loss outperforms specialized multi-scale and detector-assisted counting models on UCF-QNRF, ShanghaiTech A/B, and UCF CC 50.","On UCF-QNRF, MAE drops to 88.7 from the previous best 132, a 43.3-point reduction, without external detectors or multi-scale architectures.","The loss is insensitive to the Gaussian width $\\sigma$: Bayesian MAE stays below 98 for $\\sigma$ from 0.1 to 32, whereas the baseline varies from 118.4 to 136.2.","Training with Bayesian loss transfers across datasets better than the pixel-wise baseline when trained on UCF-QNRF and tested on the other three benchmarks.","The loss improves both VGG-19 and AlexNet backbones by similar margins, suggesting the benefit comes from the supervision scheme rather than a particular network."],"supporting_citations":[{"why":"Supplies the UCF-QNRF dataset and the previous best result (CL-CNN) that the proposed loss outperforms by a large margin.","marker":"[16]"},{"why":"Supplies the ShanghaiTech A/B datasets and the geometry-adaptive Gaussian-kernel density-map convention that serves as the baseline target.","marker":"[57]"},{"why":"Supplies the UCF CC 50 dataset used to evaluate the method on extremely dense, small-data scenes.","marker":"[15]"},{"why":"Introduced the Gaussian-kernel point-to-density conversion and least-square objective that the Bayesian loss is designed to replace.","marker":"[20]"},{"why":"Represents the pixel-wise density-map regression training approach used as the baseline comparator.","marker":"[53]"},{"why":"Provides the VGG-19 backbone architecture used in all main experiments and comparisons.","marker":"[39]"}],"fun_headline_variants":["Bayesian loss improves crowd counting without blurry density maps","Bayesian loss cuts crowd-count error to 88.7 on UCF-QNRF","Point supervision via Bayesian loss beats density maps for counting","Crowd counting with Bayesian loss: no blurry ground truth needed","Bayesian loss removes need for hand-tuned Gaussian kernels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The loss assumes that all predicted density is explainable by spatial proximity to annotated heads (or to a dummy background point), so unannotated people, head sizes that do not match the single fixed Gaussian spread, and background density near heads can silently bias the counts.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian loss improves crowd counting without blurry density maps","Bayesian loss cuts crowd-count error to 88.7 on UCF-QNRF","Point supervision via Bayesian loss beats density maps for counting","Crowd counting with Bayesian loss: no blurry ground truth needed","Bayesian loss removes need for hand-tuned Gaussian kernels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001908,"raw_usage":{"total_tokens":7532,"prompt_tokens":1057,"completion_tokens":6475,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":673,"completion_tokens_details":{"reasoning_tokens":6385}},"tokens_in":673,"tokens_out":6475,"duration_ms":44285,"temperature":1.0,"reasoning_tokens":6385,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:06:11.287181+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a crowd image and train one model with every person annotated and a second model with a random half of the annotations removed, holding the network and all other settings fixed: since the Bayesian loss forces the total expected count to equal the number of annotated points, the half-annotated model should underestimate the true count by a factor of roughly two; if it instead recovers the full count, the loss is not enforcing the count supervision the paper claims.","supporting_citations":[{"cited_title":"Composition loss for counting, density map estima- tion and localization in dense crowds","cited_arxiv_id":null,"evidence_quote":"Supplies the UCF-QNRF dataset and the previous best result (CL-CNN) that the proposed loss outperforms by a large margin."},{"cited_title":"Single-image crowd counting via multi-column convolutional neural network","cited_arxiv_id":null,"evidence_quote":"Supplies the ShanghaiTech A/B datasets and the geometry-adaptive Gaussian-kernel density-map convention that serves as the baseline target."},{"cited_title":"Multi-source multi-scale counting in extremely dense crowd images","cited_arxiv_id":null,"evidence_quote":"Supplies the UCF CC 50 dataset used to evaluate the method on extremely dense, small-data scenes."},{"cited_title":"Learning to count objects in images","cited_arxiv_id":null,"evidence_quote":"Introduced the Gaussian-kernel point-to-density conversion and least-square objective that the Bayesian loss is designed to replace."},{"cited_title":"Cross-scene crowd counting via deep convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Represents the pixel-wise density-map regression training approach used as the baseline comparator."}],"review_version":1}