{"id":"068f5a19-23bf-484c-befb-2bd45a83b043","arxiv_id":"1908.02797","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A three-part CNN with a count attention mechanism routes dense and sparse image regions to networks of different capacities and reports state-of-the-art counting errors on five benchmarks.","lead":"This paper introduces ACM-CNN, a crowd counting network that uses a coarse prediction map to pick dense regions and sends those regions to a deeper fine network while keeping sparse regions in a shallow network. The method reports lower counting errors than prior methods on five benchmark datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Count attention's localization assumption is unsupported and not directly supervised; test-set tuning of N/S further weakens the claimed adaptive-capacity advantage.","rationale":"The reader's weakest assumption -- that coarse-map peaks mark true dense regions -- is indeed the load-bearing point. The architecture is coherent and the ablations in Table 8 indicate each component contributes, which is real evidence that the overall design works. The generality experiment in Table 9 also suggests the count-attention combination has some effect beyond the particular coarse/fine networks chosen. However, no experiment isolates whether the selected patches are the right patches. A comparison against random or oracle selection would directly test whether the adaptive-capacity story is responsible for the gains. The test-set tuning of N and S is a separate but compounding issue: it undermines the confidence in the absolute numbers, though it does not by itself invalidate the architecture. Given these concerns, the reader's CONDITIONAL verdict remains appropriate; no change is needed.","tokens_in":12338,"tokens_out":4216,"duration_ms":49434,"concrete_test":"Train ACM-CNN on a held-out split of ShanghaiTech Part A and UCSD, and measure the overlap (e.g., density-mass intersection over union) between the top-N patches selected from C0 and the top-N density regions of the ground-truth density map, across training epochs. Then compare final MAE against two controls using identical training setups: (i) random patch selection with the same N and S, and (ii) oracle selection from ground-truth density peaks. If random-selection MAE is close to ACM-CNN, or oracle selection does not improve, then the count-attention localization assumption is not load-bearing for the reported accuracy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that ACM-CNN improves accuracy via adaptive capacity rests on count attention selecting the true dense regions. Section 3.1 selects the N highest-value pixels of C0 and crops regions around them (Eqs. 3-4), but C0 is never directly supervised: the intermediate loss Lb (Eq. 8) supervises only the merged map Mb, and Eq. 5 strips the selected C0 regions out of Mb. Consequently, gradients do not encourage C0 to produce accurate peaks inside the selected areas, and the hard, non-differentiable selection cannot be corrected by the final loss within a training step. The paper provides no analysis or visualization validating that the selected patches coincide with genuinely dense regions. Additionally, Section 5.1 tunes N and S on the test sets (Figure 9, Table 7) without a proper validation split or error bars, so the reported margins may reflect test-set selection rather than the mechanism. If C0's peaks are misplaced, or if the patch is nearly the whole image (e.g., N=1, S=0.9h x 0.9w on ShanghaiTech Part A), the fine network may refine a large or incorrect region and the adaptive-capacity explanation loses its foundation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes ACM-CNN, a crowd counting architecture with three modules: a shallow multi-column coarse network, a VGG-16 based fine network, and a smooth fusion network. A 'count attention' mechanism selects the N highest-valued pixels in the coarse density map C0, extracts the corresponding patches from the input for the fine network, and strips those patches from C0 before the two feature maps are summed and passed to the smooth network. The authors report state-of-the-art or near-state-of-the-art MAE/MSE on ShanghaiTech Part A/B, WorldExpo'10, UCSD, Mall, and UCF CC 50, and provide ablations showing each module contributes, plus a generality experiment using CrowdNet sub-networks.","tokens_in":12691,"tokens_out":6448,"duration_ms":70676,"significance":"The adaptive-capacity idea is reasonable and the three-module decomposition is clearly described. The paper includes useful ablations (Table 8) and a generality experiment (Table 9). If the reported numbers are reproducible, the method would be competitive. However, the significance is currently limited by evaluation concerns: hyperparameters N and S appear to be selected on test sets, no error bars are reported, and no code is released, so the state-of-the-art claim rests on weaker evidence than the tables suggest.","major_comments":[{"comment":"The count attention mechanism assumes that the highest-valued pixels in the coarse map C0 identify dense crowd regions, but this property is never supervised or validated. The intermediate loss Lb in Eq. (8) supervises the merged map Mb = CN+1 + F, and Eq. (5) strips the selected C0 regions out before this sum; the hard top-N selection is non-differentiable, so no gradient encourages C0 to place its peaks in truly dense areas. The manuscript provides no visualization or quantitative check that the selected patches coincide with dense regions (e.g., an IoU measure against the ground-truth density or a comparison of the selected patches with the true head locations). Without such evidence, the improvement of C+F over C or F in Table 8 cannot be attributed to correct localization rather than simply to the added capacity of a second network.","section":"Section 3.1-3.2, Eqs. (5)-(6), (8)"},{"comment":"The hyperparameters N and S appear to be selected on the test sets. The dataset splits described in Sections 4.2 and 4.4 contain no validation portion, and Figure 9 plots MAE against N and S for Shanghaitech Part A and UCSD without distinguishing train, validation, or test. Choosing N=1, S=[0.9*height,0.9*width] for dense sets and N=5, S=[70,70] for sparse sets from these curves and then reporting the resulting test numbers as state-of-the-art risks test-set overfitting. The authors should tune on a held-out validation split (or fix N and S before evaluation) and report the mean and standard deviation over multiple runs, especially for the 5-fold cross-validation on UCF CC 50 in Table 5.","section":"Section 5.1, Tables 1-5"},{"comment":"For all dense datasets, the chosen configuration is N=1 with a patch covering 0.9*height by 0.9*width, i.e., 81% of the image area. Under this configuration, the fine network processes nearly the entire image and the coarse network contributes only the remaining border regions; this weakens the paper's central claim of assigning 'different capacities to different portions' adaptively. The authors should either justify this configuration as a limiting case of adaptive capacity or provide results for smaller patches to demonstrate that spatial selection is what drives the gain.","section":"Section 5.1, Table 7"}],"minor_comments":[{"comment":"There is a typo: 'convolutional nerual network' should be 'convolutional neural network'.","section":"Section 2.2"},{"comment":"There is a typo: 'the mooth network' should be 'the smooth network'.","section":"Section 3.3"},{"comment":"The label 'CrowdNet contact' should be 'CrowdNet concatenation'.","section":"Table 9"},{"comment":"The notation in the MAE and MSE equations uses a norm symbol where an absolute value is intended; please replace with |ci - ci_hat| for clarity.","section":"Section 4.1, Eqs. (11)-(12)"},{"comment":"The method 'LBP + RR' is cited as reference [5], but [5] is MCNN; please correct the citation.","section":"Table 2"},{"comment":"The architecture diagram is very dense; the count attention submodule, in particular, would benefit from a separate, more detailed illustration.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a competent engineering contribution, but the empirical claims need to be made more robust. The main concern is the apparent test-set tuning of N and S and the lack of validation-based parameter selection; the authors should also directly evaluate whether the count attention localization is correct. If the authors address these points, the paper could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the quick take on arXiv:1908.02797. The paper proposes ACM-CNN, which uses a coarse network to produce a density map, picks the highest-value pixels as 'dense region' centers, crops patches around them, and runs a fine VGG-based network on those patches; a smooth network then fuses coarse and fine maps. That count attention mechanism and the coarse/fine/smooth split are genuinely new as a combination, even though each ingredient (multi-column CNNs, attention, VGG backbones) is known. This is a legitimate incremental contribution, not a new framework.\n\nWhat it does well: the architecture is clearly described, the ablations in Table 8 are consistent with the design (each component helps), and the generality test in Table 9 using CrowdNet subnets suggests the attention mechanism transfers beyond the specific coarse/fine choices. The writing is straightforward and the method is reproducible in principle.\n\nSoft spots, in order of importance. First, the evaluation is weaker than the claims. Section 5.1 tunes N and S directly on the test sets (Figure 9, Table 7) with no validation split and no error bars. For ShanghaiTech Part A the chosen setting is N=1, patch size 0.9x0.9 of the image, meaning the fine network sees almost the whole image; that makes the 'adaptive capacity' story weaker for the dense datasets. Second, the count attention's localization assumption is not directly validated. The coarse map C0 is never supervised on its own; the intermediate loss Lb supervises Mb after the selected regions have been stripped out (Eqs. 5 and 8), and the hard selection is non-differentiable, so nothing in the training signal enforces that the selected peaks actually coincide with dense crowd regions. The paper shows no visualization or quantitative check of where the selected patches fall. Third, reported gains over prior art are marginal (e.g., 72.2 vs 73.6 MAE on ShanghaiTech Part A) and no code is released, so independent confirmation is not possible.\n\nNone of these flaws sink the central idea; they just mean the reported numbers are probably optimistic and the mechanism is under-evidenced. The paper would benefit from a proper validation split for hyperparameters, error bars across multiple runs, and at least a qualitative analysis of count attention patches.\n\nFor whom: crowd counting researchers will find this a useful point of comparison, especially the attention-based patch selection idea. It deserves a serious refereeing, though not a desk reject; with the evaluation fixed it could be a solid conference paper.\n\nRecommendation: engage with it, but ask for the hyperparameter selection to be moved off the test set and for some attention localization analysis before trusting the numbers.","headline":"Incrementally novel attention-based crowd counter with a plausible design, but the headline results rest on test-set tuning and an unvalidated assumption that coarse peaks mark dense regions.","tokens_in":13093,"tokens_out":2114,"would_cite":false,"duration_ms":20390,"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":"ACM-CNN assigns different network capacity to dense and sparse crowd regions, cutting count error on five benchmarks.","keywords":["crowd counting","density map estimation","count attention","adaptive capacity","multi-scale CNN","coarse-to-fine network","surveillance analysis","attention mechanism"],"falsifier":"On a held-out set of crowd images, compare the top-$N$ patch centers picked by the count attention mechanism against the actual positions of the highest-density head clusters in the ground-truth density map. If the selected centers miss a substantial fraction of true dense clusters while the final MAE stays low, the attention mechanism is not the cause of the accuracy; conversely, if errors concentrate in missed clusters, the premise is confirmed.","tokens_in":12139,"feed_emoji":"👥","tokens_out":6491,"duration_ms":65301,"temperature":0.7,"pith_summary":"The paper tries to show that crowd counting improves when a model does not process every image region with the same fixed network. It proposes a three-part architecture: a coarse network first produces a rough density map, a count attention mechanism selects the most densely populated patches from that map, and a fine network processes only those patches; a smooth network then fuses the coarse and fine outputs into the final density map. On five public datasets the authors report lower mean absolute error than the baselines they compare against, on both dense and sparse scenes. The point of the work is that capacity allocation, not just scale diversity, is what a counting network should learn.","feed_headline":"Adaptive-capacity network cuts crowd-count error on five benchmarks","feed_subtitle":"Processing only the crowded patches with a deeper network lifts accuracy on dense and sparse scenes alike.","key_machinery":"The load-bearing object is the count attention mechanism. It traverses the coarse density map $C_0$, selects the $N$ positions with the highest pixel values as centers $\\text{pos}_i$, crops the corresponding spatial patches $\\text{CSM}_i$ from the input, and accumulates them into the high-density map $S_{N+1} = S_1 + \\sum_i \\text{CSM}_i \\otimes S_0$, while the complement $C_{N+1} = C_1 - \\sum_i \\text{CSM}_i \\otimes C_0$ keeps the low-density regions. This pair of operations is what lets one network spend deep capacity exactly where the coarse estimate says people are dense. The secondary machinery is the smooth network, a small stack of convolution-ReLU layers that fuses the fine feature map with the low-density coarse map so the final density map does not show seams at the attention patch boundaries.","core_discovery":"The central claim is that adaptive capacity allocation beats a single fixed-capacity structure for crowd counting: a shallow multi-column network is good at finding where crowds are, a deep network is good at resolving dense regions, and combining them through an attention-guided split yields accurate density maps. The count attention mechanism reads the coarse density map, takes the highest-valued pixel positions as attention centers, crops patches around them, and iteratively builds a high-density input $S_{N+1}$ for the fine network while stripping those same regions from the coarse feature map to form the low-density map $C_{N+1}$. The fine and coarse features are added, then passed through a small smoothing network to remove the rigid boundary between attended and unattended regions. The authors report MAE reductions to 72.2 on ShanghaiTech Part A, 1.01 on UCSD, and 291.6 on UCF CC 50, and ablation results attributing the gain to the attention mechanism and the smoothing fusion.","pith_inferences":["A natural extension is to make the attention selection differentiable or soft instead of a hard top-$N$ crop; this could let the model learn how many patches and what sizes are needed per image and remove the dataset-dependent threshold parameter.","If the coarse network's peaks are unreliable, the error propagates: a missed dense cluster never reaches the fine network. A testable prediction of the paper's mechanism is that counting error concentrates in clusters the coarse map ranks below the top $N$.","The same attend-then-refine loop could transfer to other dense-object counting tasks, such as cell counting in microscopy or vehicle counting in aerial imagery, where a cheap density estimate can gate a deeper refinement network."],"forward_implications":["On dense-scene benchmarks, the full three-network model lowers MAE from 74.5 to 72.2 on ShanghaiTech Part A when the smooth network is added, so the smoothing stage contributes accuracy, not just visual smoothness.","On sparse-scene benchmarks, the same architecture reaches MAE 1.01 on UCSD and 2.3 on Mall, showing the adaptive split does not hurt when crowds occupy only a small part of the frame.","The count attention mechanism is not tied to the specific coarse and fine networks: swapping in the deep and shallow networks from CrowdNet still improves UCSD MAE from 1.5 with simple concatenation to 1.2 with count attention, indicating it is a general fusion strategy.","Because the method is fully convolutional and trained end-to-end with an $L_2$ loss plus an intermediate supervision term, it can be inserted into existing counting pipelines without changing the training paradigm."],"supporting_citations":[{"why":"Supplies the multi-column coarse architecture and the geometry-adaptive Gaussian kernels used to build ground-truth density maps.","marker":"[5]"},{"why":"Provides the patch-switching baseline and the motivation that fixed-capacity columns underperform across density levels.","marker":"[7]"},{"why":"Provides the VGG deep backbone from which the fine network is built.","marker":"[8]"},{"why":"Context-pyramid baseline compared on ShanghaiTech and WorldExpo, and prior work on encoding global context into density estimation.","marker":"[17]"},{"why":"Introduces the WorldExpo'10 dataset and the cross-scene counting baseline used in Table 2.","marker":"[35]"},{"why":"Introduces the UCSD dataset and the Gaussian-process regression baseline and ROI evaluation protocol.","marker":"[37]"},{"why":"Provides the deep and shallow sub-networks used in the generality experiment showing count attention beats simple concatenation.","marker":"[41]"},{"why":"Introduces the Mall dataset and the regression baselines used in Table 4.","marker":"[4]"}],"fun_headline_variants":["Adaptive capacity CNN lowers crowd-count error on five sets","Attention-guided capacity split sharpens crowd density maps","Coarse-to-fine adaptive network counts dense crowds better","Crowd counting gets adaptive: capacity allocation improves results","Smooth fusion plus attention key to adaptive crowd counting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The mechanism assumes the highest-valued pixels in the coarse density map are the true dense crowd centers, so if the coarse network peaks in the wrong places, the fine network refines irrelevant patches and the claimed benefit of adaptive capacity does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive capacity CNN lowers crowd-count error on five sets","Attention-guided capacity split sharpens crowd density maps","Coarse-to-fine adaptive network counts dense crowds better","Crowd counting gets adaptive: capacity allocation improves results","Smooth fusion plus attention key to adaptive crowd counting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000257,"raw_usage":{"total_tokens":1573,"prompt_tokens":936,"completion_tokens":637,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":559}},"tokens_in":552,"tokens_out":637,"duration_ms":7625,"temperature":1.0,"reasoning_tokens":559,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:33:38.686803+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out set of crowd images, compare the top-$N$ patch centers picked by the count attention mechanism against the actual positions of the highest-density head clusters in the ground-truth density map. If the selected centers miss a substantial fraction of true dense clusters while the final MAE stays low, the attention mechanism is not the cause of the accuracy; conversely, if errors concentrate in missed clusters, the premise is confirmed.","supporting_citations":[{"cited_title":"Zhang, D","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-column coarse architecture and the geometry-adaptive Gaussian kernels used to build ground-truth density maps."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the patch-switching baseline and the motivation that fixed-capacity columns underperform across density levels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Context-pyramid baseline compared on ShanghaiTech and WorldExpo, and prior work on encoding global context into density estimation."},{"cited_title":"Zhang, H","cited_arxiv_id":null,"evidence_quote":"Introduces the WorldExpo'10 dataset and the cross-scene counting baseline used in Table 2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the UCSD dataset and the Gaussian-process regression baseline and ROI evaluation protocol."},{"cited_title":"Boominathan, S","cited_arxiv_id":null,"evidence_quote":"Provides the deep and shallow sub-networks used in the generality experiment showing count attention beats simple concatenation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Mall dataset and the regression baselines used in Table 4."}],"review_version":1}