{"id":"693e10fd-d1a1-4c6c-9112-cda22fc866ad","arxiv_id":"1908.03716","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A crowd-counting network with spatial and channel attention modules reports lower mean absolute error than several prior models on four public datasets.","lead":"Crowd counting models estimate the number of people in an image by predicting a density map. This paper adds two attention modules, one over spatial positions and one over feature channels, to a standard CNN and reports lower counting errors than several prior methods on four crowd datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation is confounded: the FCN baseline in Table 5 is not CSRNet's dilation head (six layers ending at 64 channels vs CSRNet's four layers ending at 256), so reported SOTA gains may come from the redesigned head rather than from SAM/CAM.","rationale":"The reader's weakest assumption about cross-paper comparability of published numbers is real, but the stress-test found a sharper internal problem: the paper's own controlled ablation does not isolate the proposed attention modules. Table 1's dilation module ends at 64 channels and has six layers, whereas CSRNet's ends at 256 channels with four layers; Section 3.1 confirms the 64-channel design. The FCN baseline in Table 5 is therefore not CSRNet, and its 13.2 MAE is notably worse than CSRNet's published 10.6. This makes the 'each module helps' conclusion an artifact of the chosen baseline unless the attention modules are shown to improve a correctly configured CSRNet. The same confound affects the headline SOTA comparison: Table 2 compares SCAR (new head plus attention) to CSRNet (old head), so the 1.9 MAE gap on Part A cannot be attributed to SAM/CAM. The proposed concrete test would settle this by comparing exact CSRNet, CSRNet plus attention, and the full SCAR under one training pipeline. I also note the duplicated SFCN/SCAR row in Table 4's cross-location section, which adds a reporting-integrity concern but is secondary to the confound. Since the paper can be fixed with a proper ablation and the architecture idea remains plausible, the reader's CONDITIONAL verdict stays appropriate.","tokens_in":9970,"tokens_out":8812,"duration_ms":91957,"concrete_test":"Train on ShanghaiTech Part B with identical settings (same resizing, optimizer, epoch count, and at least three seeds) three configurations: (i) exact CSRNet with four dilated layers c512/c512/c512/c256 and a 1x1 output layer; (ii) that same CSRNet plus SAM and CAM applied to the 256-channel feature map (using 1x1 projections to form the attention streams); and (iii) SCAR's full architecture as in Table 1. If configuration (ii) beats (i) by roughly the same margin that SCAR beats its current FCN baseline, the attention modules are validated. If (iii) beats (i) but (ii) does not, the reported gain is due to the redesigned dilation head rather than to SAM/CAM. Report MAE and MSE means with standard deviations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.1's ablation is meant to show that SAM and CAM each help, but the 'FCN' baseline does not match the strongest published baseline. Table 1 specifies the proposed dilation module as six layers (c512, c512, c512, c256, c128, c64), whereas CSRNet, which the text says the dilation conv is borrowed from, uses four dilated conv layers ending at 256 channels. Section 3.1 explicitly says the module outputs 1/8-size feature maps with 64 channels, so the FCN baseline is a different network from CSRNet. Consistent with this, Table 5 reports FCN Part B MAE of 13.2, while CSRNet's published Part B MAE is 10.6. Therefore the improvement from FCN (13.2) to SCAR (9.5) is measured against a weaker variant, and the 2.6-point gap between the baseline and CSRNet is as large as the 3.7-point gain attributed to attention. The SOTA claim in Table 2 compares the full model (new dilation head plus attention) to CSRNet, so any observed win confounds the attention modules with the redesigned head. The paper should demonstrate that attention helps on top of a correctly configured CSRNet baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes SCAR, a crowd-counting network that appends a Spatial-wise Attention Module (SAM) and a Channel-wise Attention Module (CAM) to a VGG-16 backbone with a dilation module, fuses the two attention streams by concatenation, and regresses a density map. The paper reports MAE/MSE and PSNR/SSIM on ShanghaiTech Part A/B, UCF_CC_50, and GCC, and claims state-of-the-art results. An ablation on ShanghaiTech Part B reports that both SAM and CAM improve the baseline FCN, with the full model reaching 9.5 MAE.","tokens_in":10279,"tokens_out":5207,"duration_ms":54450,"significance":"The application of non-local attention to crowd counting is plausible, and the two-module design is simple and clearly specified at the architectural level. If the empirical claims are correct, the paper provides a useful baseline and evidence that spatial and channel attention transfer to density regression. The use of density-map quality metrics (PSNR/SSIM) in addition to MAE/MSE is a strength. However, the central SOTA and ablation claims rest on a baseline configuration that does not match the strongest published comparator, and no code or training repeatability information is provided.","major_comments":[{"comment":"The ablation baseline is confounded with the dilation-head design. Table 1 specifies a dilation module of six dilated conv layers (c512, c512, c512, c256, c128, c64), while the CSRNet architecture cited in Section 3.1 as the source of the dilation conv uses four dilated layers ending at 256 channels; Section 3.1 confirms that the proposed module outputs 64-channel 1/8-size maps. The published CSRNet Part B MAE is 10.6, while the FCN baseline in Table 5 is 13.2, so the 3.7-point improvement attributed to the attention modules is measured against a weaker network. Please rerun the ablation with a faithful CSRNet configuration, or separately ablate the dilation-head design from the attention modules, so the contribution of SAM and CAM is isolated.","section":"Section 5.1 (Table 5) and Table 1"},{"comment":"The cross-location GCC results are not clearly attributed. In Table 4 the row with MAE 87.2 in the cross-location section is labeled 'SFCN' rather than SCAR, while the text states that SCAR achieves 87.2 for cross-location. As printed, the table does not actually show SCAR's cross-location MAE, which invalidates the 'best MAE' claim for that protocol unless the row is relabeled. Please correct the label and verify the reported values.","section":"Section 4.5 (Table 4)"},{"comment":"All reported results appear to come from a single training run with no standard deviation. Tables 2 through 5 report only point estimates; on ShanghaiTech Part A the gap between SCAR (66.3 MAE) and CSRNet (68.2 MAE) is 1.9, which is within the typical run-to-run variation of crowd-counting models. Please report mean plus/minus standard deviation over at least three independent runs for the main tables and at least for the Part B ablation, and state whether test-time augmentation or multi-scale evaluation was used.","section":"Sections 4.2, 4.3, and 5.1 (Tables 2-5)"},{"comment":"The construction of the ground-truth density maps is not specified. Section 4.2 states that images are resized to 576 by 768 and density maps are generated under the same size, but it does not describe how the head annotations are converted to density maps, for example fixed Gaussian sigma versus geometry-adaptive kernels, normalization, or how resizing affects the density values. Without this information, the quantitative comparisons are not reproducible and may not be comparable to the cited baselines. Please specify the generation procedure explicitly.","section":"Section 4.2"}],"minor_comments":[{"comment":"The conv3 row of the VGG-16 backbone lists c128, but VGG-16 conv3 has 256 channels; this is likely a typo and should be corrected for reproducibility.","section":"Table 1"},{"comment":"The caption repeats 'Column 3 and 4' and does not correctly map the five displayed columns to the spatial and channel attention maps; please update the column labels.","section":"Figure 5 caption"},{"comment":"The sentence claiming '8 best places from the 12 metrics' is hard to verify because Table 4 lists four metrics for each of three protocols; please state precisely which cells are being counted.","section":"Section 4.5"},{"comment":"The phrase 'different different fusion strategies' contains a duplicated word and should be reworded.","section":"Section 5.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a computer vision journal, but the incremental contribution depends on the ablation and comparison being trustworthy. I would encourage the editor to request code release and a corrected baseline experiment before considering it further."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zhi,\n\nQuick take: this is a competent incremental paper that bolts the non-local block (Wang et al.) and its channel-wise counterpart onto a dilated VGG-16 and shows improved MAE/MSE on four crowd benchmarks. The modules themselves are not new — SAM is the non-local block, CAM is the channel variant from SCA-CNN. The new bit is the combination, the concatenation fusion, and the application to counting. That's real but modest novelty.\n\nWhat's good: the ablation on ShanghaiTech Part B shows both modules help when added to their FCN baseline, with MAE dropping from 13.2 to 9.5. The density-map quality numbers (PSNR/SSIM) also improve. The paper is clearly written and the equations are consistent with the non-local formulation.\n\nWhere it's soft: the stress-test is right. The 'FCN' baseline in Table 5 is not CSRNet's dilation head. Table 1 shows a six-layer head ending at 64 channels; CSRNet uses four layers ending at 256. The reported baseline MAE on Part B is 13.2 versus CSRNet's published 10.6. So the attention gain is measured from a weaker starting point, and the claim that SCAR beats CSRNet confounds the new head with the attention modules. To support the contribution, the authors need an ablation that adds SAM/CAM to a properly configured CSRNet. Also: no standard deviations, single runs, no code release, and the conv3 channel count in Table 1 (128) looks like a typo (VGG-16 has 256).\n\nWorth noting: the state-of-the-art claim depends on comparing to published numbers from other papers without shared evaluation code. The margins are small (1.9 on Part A, 1.1 on Part B), so they may be within run-to-run noise.\n\nWho's this for? Someone tracking crowd-counting progress or wanting a quick baseline for attention-based counting. It's not a conceptual breakthrough. With the baseline confound fixed and a few robustness runs, it could be a solid workshop/short-paper contribution. I'd send it to peer review because the core question — does non-local attention help counting — is legitimate and the ablation, once corrected, answers it.\n\nBest,","headline":"Known attention blocks applied to crowd counting with an honest but confounded ablation; worth refereeing if the baseline is fixed.","tokens_in":10809,"tokens_out":1806,"would_cite":false,"duration_ms":18239,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding non-local spatial and channel attention to a regression CNN improves crowd-counting accuracy.","keywords":["crowd counting","density map estimation","spatial attention","channel attention","self-attention","non-local networks","VGG-16","congested scenes"],"falsifier":"Re-run SCAR and the closest baselines (especially CSRNet and ic-CNN) from their public implementations on ShanghaiTech Part A and Part B using the same image sizes, optimizer settings, and multiple random seeds, then check whether SCAR's reported MAE/MSE advantages persist across seeds; if the gaps shrink below the seed-to-seed variation or reverse, the claimed state-of-the-art result would not survive a controlled comparison.","tokens_in":9746,"feed_emoji":"👥","tokens_out":3847,"duration_ms":43615,"temperature":0.7,"pith_summary":"This paper tries to show that crowd counting improves when the regression network looks beyond local image patches and explicitly models two kinds of attention: spatial attention, which relates every pixel to every other pixel to capture large-range context, and channel attention, which highlights the feature channels that respond to heads rather than backgrounds. The proposed network, SCAR, attaches these two attention modules to a VGG-16-based dilated convolutional backbone and merges their outputs by concatenation before regressing the density map. If the claim is right, the same recipe could be reused in other pixel-wise prediction tasks that suffer from background confusion and limited receptive fields. The paper reports lower mean absolute error on ShanghaiTech Part A/B, GCC, and UCF_CC_50 than the compared baselines, with ablation experiments on ShanghaiTech Part B showing that each attention module alone helps and that combining them helps most.","feed_headline":"Attention modules cut crowd-counting error on four datasets","feed_subtitle":"SCAR adds spatial and channel attention to a VGG-16 regression net, lowering MAE on ShanghaiTech, GCC, and UCF_CC_50.","key_machinery":"The central machinery is a pair of non-local attention modules built on the dilated convolutional features: SAM computes an HW-by-HW spatial attention map that encodes pixel-wise context over the whole image, and CAM computes a C-by-C channel attention map that models dependencies between feature channels, each with a learnable scaling parameter (lambda and mu). The two attention outputs are concatenated with the original feature map and regressed to a one-channel density map by a 1x1 convolution and 8x upsampling, trained with mean squared error. This mechanism is what lets the network use large-range context and head-region channel selectivity rather than relying only on local appearance.","core_discovery":"SCAR claims that placing two non-local self-attention modules on top of a dilated VGG-16 feature extractor reduces crowd-counting error. The Spatial-wise Attention Model computes a softmax-normalized HW-by-HW map relating every spatial position to every other, then reweights the feature map with a learnable scale; the Channel-wise Attention Model computes a C-by-C map relating every channel pair and reweights channels, so the network learns to trust head-related channels and suppress background response. The outputs of both modules are concatenated with the backbone features and passed through a 1x1 convolutional regressor followed by 8x upsampling to produce the density map. On ShanghaiTech Part A this yields MAE 66.3 versus CSRNet's 68.2, on Part B 9.5 versus 10.6, and on UCF_CC_50 259.0, the best MAE in the reported comparison. Ablation on Part B shows FCN baseline MAE 13.2, FCN+SAM 11.0, FCN+CAM 11.5, and the full SCAR 9.5, which the paper reads as evidence that both attention streams contribute and that their combination is additive.","pith_inferences":["Beyond the paper: if the attention modules are truly task-agnostic, they could be dropped into saliency detection or semantic segmentation networks to test whether the same spatial-context and channel-selectivity gains transfer; the paper names these applications as future work but does not test them.","Beyond the paper: the Part B fusion experiment compares concatenation with element-wise sum, but a direct extension would test weighted fusion or multi-scale fusion of SAM and CAM to see whether the two streams carry complementary information that a learned weighting could exploit further.","Beyond the paper: because the paper reports single-run results, a direct extension would add multiple random seeds and report means and standard deviations; this would reveal whether the reported gaps, such as the 1.9 MAE difference on Part A, persist beyond training noise.","Beyond the paper: the cross-location GCC results still show a large absolute MAE gap relative to random splitting, which suggests that even with attention, synthetic-to-real or cross-camera generalization remains a bottleneck; a natural extension is to measure how much of that gap attention modules close compared with a no-attention baseline."],"forward_implications":["On ShanghaiTech Part A and Part B, SCAR reports lower MAE/MSE than the listed baselines including CSRNet and ic-CNN, with MAE 66.3/9.5 versus CSRNet's 68.2/10.6.","On the synthetic GCC dataset, SCAR reports the best MAE among the compared methods under all three evaluation protocols: 31.7 for random splitting, 55.8 for cross-camera, and 87.2 for cross-location.","On UCF_CC_50, SCAR reports the best MAE of 259.0 among the compared methods, improving on ic-CNN's 260.9 and CSRNet's 266.1.","The ablation on ShanghaiTech Part B shows each attention module reduces error on its own and that SAM outperforms CAM, while the full model improves density map quality as measured by PSNR and SSIM.","The paper's architecture is backbone-agnostic in principle, so the attention modules could be attached to other feature extractors or other pixel-wise regression heads wherever long-range context and background suppression matter."],"supporting_citations":[{"why":"Provides the ShanghaiTech dataset, its train/test split, and the MCNN baseline that SCAR compares against.","marker":"[17]"},{"why":"Provides the dilated convolutional baseline CSRNet, the VGG-16 style backbone and dilation module design that SCAR builds on, and the PSNR/SSIM density-map quality comparison.","marker":"[6]"},{"why":"Supplies the non-local self-attention mechanism that both SAM and CAM adapt for spatial and channel dimensions.","marker":"[25]"},{"why":"Supplies the VGG-16 network whose first 10 convolutional layers form the local feature extractor of SCAR.","marker":"[23]"},{"why":"Provides the GCC synthetic dataset and the FCN/SFCN baselines, as well as the evaluation protocols used in the GCC experiments.","marker":"[16]"},{"why":"Motivates the self-attention formulation used in the attention modules, particularly the scaled dot-product attention style of correlation.","marker":"[24]"},{"why":"Provides the CP-CNN contextual pyramid baseline that SCAR compares against on ShanghaiTech and UCF_CC_50.","marker":"[21]"},{"why":"Provides the Switching-CNN baseline, another VGG-16-based counting method that SCAR compares against on ShanghaiTech and UCF_CC_50.","marker":"[18]"}],"fun_headline_variants":["SCAR attention modules improve crowd counting on four datasets","Spatial and channel attention cut crowd-counting MAE on multiple sets","Self-attention boosts density map accuracy in SCAR crowd counter","Attention to heads and context lowers crowd-counting error","SCAR adds spatial and channel attention, trimming MAE on 4 datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art claim assumes that the published MAE/MSE numbers from other papers were obtained under directly comparable training and evaluation conditions, so that small differences, like the 1.9 MAE gap over CSRNet on Part A, are meaningful rather than protocol or training noise.","fun_headline_variants_meta":{"raw":{"variants":["SCAR attention modules improve crowd counting on four datasets","Spatial and channel attention cut crowd-counting MAE on multiple sets","Self-attention boosts density map accuracy in SCAR crowd counter","Attention to heads and context lowers crowd-counting error","SCAR adds spatial and channel attention, trimming MAE on 4 datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000459,"raw_usage":{"total_tokens":2336,"prompt_tokens":1016,"completion_tokens":1320,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":1242}},"tokens_in":632,"tokens_out":1320,"duration_ms":9530,"temperature":1.0,"reasoning_tokens":1242,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:04:20.107748+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run SCAR and the closest baselines (especially CSRNet and ic-CNN) from their public implementations on ShanghaiTech Part A and Part B using the same image sizes, optimizer settings, and multiple random seeds, then check whether SCAR's reported MAE/MSE advantages persist across seeds; if the gaps shrink below the seed-to-seed variation or reverse, the claimed state-of-the-art result would not survive a controlled comparison.","supporting_citations":[{"cited_title":"Zhang, D","cited_arxiv_id":null,"evidence_quote":"Provides the ShanghaiTech dataset, its train/test split, and the MCNN baseline that SCAR compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the dilated convolutional baseline CSRNet, the VGG-16 style backbone and dilation module design that SCAR builds on, and the PSNR/SSIM density-map quality comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the non-local self-attention mechanism that both SAM and CAM adapt for spatial and channel dimensions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GCC synthetic dataset and the FCN/SFCN baselines, as well as the evaluation protocols used in the GCC experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the CP-CNN contextual pyramid baseline that SCAR compares against on ShanghaiTech and UCF_CC_50."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Switching-CNN baseline, another VGG-16-based counting method that SCAR compares against on ShanghaiTech and UCF_CC_50."}],"review_version":1}