{"id":"21de1127-4763-4537-b2ea-d86d14c60f31","arxiv_id":"1908.10937","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"The proposed MBTTBF-SCFB network achieves lower average counting error than several prior methods on ShanghaiTech, UCF_CC_50, and UCF-QNRF, though it trails CAN on two benchmarks.","lead":"This paper describes a neural network for counting people in dense crowds that merges fine-grained image details with global context through multi-level feature fusion. The authors report lower average counting errors than several prior systems on three congested crowd benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Part-A SOTA claim hinges on the 60.2 vs 62.5 scale-supervision delta, but the MRF head-size estimates are unvalidated and no control rules out generic auxiliary-loss gains.","rationale":"The reader's weakest assumption correctly identifies the MRF head-size estimates as unvalidated and the scale-aware supervision as potentially fitting to noise. My stress-test goes one step further and notes that the Part-A state-of-the-art claim specifically depends on the 60.2 vs 62.5 difference; without the scale-aware supervision the method is behind CAN on that benchmark. The ablation as designed cannot distinguish 'correct scale decomposition' from 'extra auxiliary losses.' This is a genuine experimental-design gap, not just a missing quantitative figure. The architecture itself is plausible, and the multi-level fusion and SCFB contributions may well be sound; the UCF-QNRF result (102.1 without scale supervision) provides independent support for the fusion architecture. The paper also includes a detailed ablation study, which is credit-worthy, but the missing control is load-bearing for the headline Part-A comparison and for the claimed 'principled' scale-aware supervision. The appropriate verdict remains CONDITIONAL: the authors should either run the permutation control, validate the MRF estimates, or soften the claim that scale-aware supervision is the source of the improvement. I do not see grounds to reject outright, since the architectural contributions appear independently useful on UCF-QNRF.","tokens_in":15225,"tokens_out":4928,"duration_ms":52038,"concrete_test":"Retrain experiment ix (MBTTBF-SCFB) on ShanghaiTech-A with the four scale-aware ground-truth maps replaced by randomly permuted bin assignments that preserve the per-bin head counts, keeping all auxiliary losses and the rest of the pipeline identical. If MAE stays near 60.2 (rather than reverting toward the 62.5 of SCFB-NS), the reported gain is not attributable to genuine scale-awareness, and the Part-A result should be reported as a deep-supervision effect; if MAE reverts to about 62.5, the scale estimates are doing real work. This is a single, decisive ablation that does not require external annotations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that MBTTBF-SCFB reduces counting error on congested benchmarks, and on ShanghaiTech-A the claim of superiority over CAN (60.2 vs 61.3) depends entirely on the scale-aware supervision: without it (SCFB-NS) the model scores 62.5, which is behind CAN. On UCF-QNRF the architecture alone already beats CAN (102.1 vs 107.0), but Part A is a flagship result. The mechanism for the scale-aware gain is the MRF-based head-size estimation of Sec. 3.3, which assigns each head a Gaussian variance from superpixel/watershed segment sizes and bins heads into four scale maps used as intermediate supervision (Sec. 3.2). The accuracy of these estimates is never quantified; Fig. 5 is qualitative only. If the MRF segment sizes do not track true head size in dense or occluded regions, the four scale maps are effectively arbitrary partitions of heads. In that case the improvement from adding them could be produced by the extra auxiliary losses (deep supervision) rather than by correct scale decomposition. The ablation (viii vs ix) changes both the presence of extra supervision and the validity of the scale labels simultaneously, so it cannot separate these explanations. Since the Part-A SOTA result and the 'principled scale-aware ground truth' contribution both depend on this separation, the central claim is not yet secured.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a crowd-counting network built on VGG16 that combines multi-level bottom-top and top-bottom fusion branches, scale-complementary feature extraction blocks (SCFB), and an MRF-based head-size estimation framework that generates four scale-aware density maps used as intermediate supervision. Experiments on ShanghaiTech Part A/B, UCF_CC_50, and UCF-QNRF report MAE/MSE values and show improvements over several recent methods, with headline results of 60.2 on ShanghaiTech Part A and 97.5 on UCF-QNRF.","tokens_in":15563,"tokens_out":4309,"duration_ms":42118,"significance":"If the results hold, the fusion scheme is a useful contribution to congested-scene counting: the architecture-only MBTTB-SCFB variant already outperforms CAN on UCF-QNRF (102.1 vs. 107.0), and the ablation study systematically compares multiple fusion topologies. However, the paper provides no code or repeated-run statistics, and the central scale-aware ground-truth contribution depends on an unvalidated MRF head-size estimator. The paper's own tables also contradict the introduction's claim of outperforming all existing methods, since CAN is better on ShanghaiTech Part B and UCF_CC_50. These issues make the strong comparative claims premature.","major_comments":[{"comment":"The only evidence that the MRF-based scale-aware ground truth is responsible for the improvement is the contrast between the SCFB-without-scale-supervision row (62.5 MAE on ShanghaiTech Part A) and the full SCFB row (60.2). This contrast changes two variables at once: the presence of additional auxiliary supervision and the correctness of the inferred head-size labels. Because Section 3.3 never quantitatively validates the MRF estimates against actual head sizes, and Figure 5 is qualitative only, the gain could come from a generic auxiliary-loss or deep-supervision effect rather than from genuine scale decomposition. I recommend adding a control that uses the same four-way auxiliary supervision with constant-variance labels, nearest-neighbor labels, or permuted scale-bin assignments; without such a control, the paper's principal contribution is not secured.","section":"Section 3.3 and Table 1, rows (viii)-(ix)"},{"comment":"The introduction states that the proposed method 'outperforms all existing methods by a significant margin,' but the paper's own tables show otherwise: on ShanghaiTech Part B the proposed method scores 8.0 MAE versus CAN's 7.8, and on UCF_CC_50 it scores 233.1 versus CAN's 212.2. Only the abstract's narrower claim of outperforming 'several recent methods' is consistent with the data. This overclaim should be corrected, and the comparative discussion should acknowledge the specific datasets where CAN is ahead.","section":"Section 1 and Section 5.3, Tables 2-4"},{"comment":"All reported comparisons are single MAE and MSE values with no error bars, multiple random seeds, repeated trials, or significance testing. Several headline differences are small: 60.2 versus 61.3 on ShanghaiTech Part A and 8.0 versus 7.8 on Part B are within typical run-to-run variation for crowd-counting models trained with different initializations. To support the state-of-the-art claim, the authors should report the mean and standard deviation over at least a few runs, or perform a paired test, and state the exact training/evaluation protocol used for each table.","section":"Section 5.2-5.3, Tables 1-4"}],"minor_comments":[{"comment":"The dataset name is inconsistent: the abstract and Section 5.1 use 'UCF CROWD 50,' while Table 3 uses 'UCF CROWD 50' and the text elsewhere uses 'UCF CC 50' and 'UCF_CC_50.' Please standardize.","section":"Throughout"},{"comment":"The sentence 'simple fusion scheme of addition/concatenation (experiments (i) and (ii))' refers to the wrong rows; it should say experiments (ii) and (iii), since experiment (i) is the baseline without fusion.","section":"Section 5.2, ablation study text"},{"comment":"The paragraph beginning 'The second level of bottom-top fusion path' describes the top-bottom path; the phrase should read 'The second level of top-bottom fusion path' to avoid confusion.","section":"Section 3.1, top-bottom fusion description"},{"comment":"The attention equation writes A = sigmoid(cat(F^1_56, F^2_456, F^1_43, F^2_543)), but the subsequent formula uses A1 through A4 as if they are per-branch maps; please clarify how the four-channel attention map is split and how spatial dimensions of the fused features are aligned.","section":"Section 3.1, attention module"},{"comment":"Adam is described as having a 'momentum of 0.9'; this is imprecise because Adam uses beta1 and beta2 hyperparameters. Please specify the exact beta values and any weight decay or batch size.","section":"Section 4, training details"},{"comment":"Equation (1) uses a single sigma for all heads, but Section 3.3 later assigns per-head variances from the MRF estimates. The notation should distinguish the constant-variance baseline from the scale-aware formulation.","section":"Section 3.1, Eq. (1)"},{"comment":"The definition of the final conv layers c3i and c3j as '{C32,1,1-R}' seems to indicate a filter size of 1x1 with 32 input channels, but the output channel count is unclear. Please specify the exact layer parameters and how intermediate density-map predictions are produced from these features.","section":"Section 3.2, SCFB description"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The punchline: this is a solid, incremental engineering paper that deserves peer review, but the headline numbers on ShanghaiTech-A are thinner than the abstract implies, and the scale-aware supervision story needs more evidence.\n\nWhat is actually new: the MBTTBF architecture with two-level bottom-top and top-bottom fusion, and the SCFB blocks with cross-scale residual connections. The ablation is well-designed and shows each piece adds something: fusion direction matters, multi-level helps, SCFB helps, and adding the scale-aware ground truth helps. On UCF-QNRF the improvement over CAN (97.5 vs 107) is real and impressive. On ShanghaiTech-A the paper reports the best MAE (60.2 vs CAN's 61.3), but by a margin that could easily be noise.\n\nSoft spots are proportionate. First, no code, no error bars, no repeated runs. For a paper claiming state of the art on several benchmarks, that matters; a 1-point MAE difference is not interpretable without variance. Second, the introduction says the method outperforms all existing methods by a significant margin, but their own tables show CAN is better on ShanghaiTech-B and much better on UCF_CC_50 (212.2 vs 233.1). The abstract's \"outperform several recent methods\" is accurate; the intro overclaims. Third, the MRF head-size estimates are never validated quantitatively—Fig. 5 is qualitative. The scale-aware intermediate maps depend on those estimates being correct. The ablation shows adding scale supervision improves ShanghaiTech-A from 62.5 to 60.2, but this change also adds auxiliary losses; the paper does not disentangle whether the gain comes from correct scale decomposition or from extra deep supervision. That is the main weakness: the \"principled\" part of the ground truth is asserted, not demonstrated.\n\nThe citation pattern looks fine—the paper builds on recent fusion work and cites relevant datasets and baselines. The ablation is genuinely informative.\n\nWho this is for: people working on crowd counting or multi-scale fusion will find the architecture and ablation useful. The UCF-QNRF result is worth knowing even if the Part-A SOTA is fragile. This paper deserves a serious referee rather than a desk reject; the issues are fixable with more experiments and more careful claims.\n\nMy recommendation: send it to peer review, and ask the authors for variance estimates, a quantitative validation of the MRF scale estimates, and a control that isolates scale supervision from plain auxiliary loss.","headline":"Solid incremental architecture paper; the ShanghaiTech-A SOTA claim rests on a small delta and the scale-supervision gain is not cleanly attributed.","tokens_in":16025,"tokens_out":2375,"would_cite":false,"duration_ms":24159,"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":"A bidirectional, multi-level fusion scheme with scale-aware supervision reduces counting error in congested crowds.","keywords":["crowd counting","density map estimation","multi-scale feature fusion","bottom-top fusion","top-bottom fusion","scale-aware supervision","Markov random field","congested scenes"],"falsifier":"Compare the MRF segment-size estimates with manually measured head widths on a sample of ShanghaiTech-A or UCF-QNRF images, and retrain the SCFB network with density maps built from the measured sizes; if agreement is weak in dense regions, or if the measured-size maps do not preserve the reported MAE 60.2 and 97.5, the scale-aware supervision claim is falsified.","tokens_in":15056,"feed_emoji":"👥","tokens_out":7685,"duration_ms":68469,"temperature":0.7,"pith_summary":"Congested crowd counting is dominated by scale variation, but the paper argues that straightforward multi-scale fusion fails because simple addition or concatenation merges correlated features without learning what each scale contributes. It proposes a VGG16-based network with two hierarchical fusion paths, one carrying spatial detail upward and one carrying context downward, joined by scale-complementary residual blocks that are supervised by density maps binned by estimated head size. On the three congested benchmarks tested, the method lowers counting error, reporting MAE 97.5 on UCF-QNRF, below the 107.0 of the context-aware baseline in the same table, and MAE 60.2 on ShanghaiTech-A. A sympathetic reading of the contribution is a recipe for making scale explicit in both the architecture and the training signal when only head locations are annotated.","feed_headline":"Two-way feature fusion lowers crowd-count error in congested scenes","feed_subtitle":"The network posts its lowest mean absolute errors on UCF-QNRF and ShanghaiTech-A where scale variation is hardest.","key_machinery":"The load-bearing object is the MBTTBF-SCFB network built on a VGG16 backbone. It contains a bottom-top branch that propagates spatial detail upward through two levels of fusion and a top-bottom branch that propagates semantic context downward through two levels, with a learned self-attention module combining the four resulting feature maps. The scale complementary feature extraction block (SCFB) is a cross-scale residual unit: for adjacent-layer features $F_i$ and $F_j$ it computes residuals and exchanges them ($\\hat{F}_i = F_i + F_j^r$, $\\hat{F}_j = F_j + F_i^r$) before convolution and loss. The scale-aware ground truth is generated by a Markov random field that combines SLIC superpixel segmentation and watershed segmentation of the head-annotation distance transform to estimate each head's Gaussian variance, after which heads are binned into four size classes to form maps $Y^s_3,\\dots,Y^s_6$ that supervise the intermediate blocks. This machinery is what carries the claim that the network learns complementary rather than redundant scale features.","core_discovery":"The central claim is that multi-level bidirectional fusion, rather than one-directional propagation or late concatenation, is what lets a counting network combine spatial detail from shallow layers with semantic context from deep layers. The ablation supports this: on ShanghaiTech-A, mean absolute error falls from 78.3 for the VGG16 baseline to 63.2 for multi-level bottom-top/top-bottom fusion with concatenation, and then to 60.2 when the fusion blocks are replaced by scale-complementary blocks supervised by scale-aware ground-truth maps. The authors attribute the gain to cross-scale residual connections that make each branch encode what the adjacent scale lacks, and to four-bin density maps that push intermediate layers toward scale-specialized features. They report results on ShanghaiTech, UCF_CC_50, and UCF-QNRF and state that the proposed method outperforms several recent methods on all three datasets.","pith_inferences":["The scale-estimation step is never directly validated against measured head sizes, so part of the 2.3 MAE gain attributed to scale-aware supervision could be the network adapting to the MRF's estimate noise rather than to true head scale; a direct validation would settle this.","The same bidirectional fusion and cross-scale residual recipe could transfer to other dense-object counting problems without scale labels, such as cell or animal counting, using image segmentation as the scale proxy.","A clean test would retrain the SCFB blocks with scale maps computed from manually measured head sizes on a subset of UCF-QNRF; if MAE does not improve or degrades, the fusion architecture rather than the MRF scale estimates would be carrying the result.","Because the network is built on VGG16 with many auxiliary losses, some of the improvement could be tied to the backbone's capacity and the extra supervision signals rather than to the fusion topology itself; the paper does not isolate that."],"forward_implications":["On UCF-QNRF, the reported MAE of 97.5 places the method ahead of the 107.0 reported for the context-aware baseline in the same comparison, on the largest of the three datasets.","On ShanghaiTech-A, the reported MAE of 60.2 is the best in the paper's comparison table, and removing the scale-aware supervision raises it to 62.5 in the ablation.","On ShanghaiTech-B and UCF_CC_50, the reported errors (MAE 8.0 and 233.1) trail the context-aware baseline, so the paper's 'several recent methods' claim is accurate but does not extend to every listed method.","Because each ablation step improves MAE, the design's components contribute additively: bidirectionality, a second fusion level, cross-scale residuals, and scale-aware supervision each add a measurable gain."],"supporting_citations":[{"why":"supplies the VGG16 backbone whose conv3-conv6 features are the inputs to both fusion branches.","marker":"[53]"},{"why":"provides the ShanghaiTech dataset and the common nearest-neighbor scale-estimation baseline that the MRF approach is designed to improve.","marker":"[74]"},{"why":"provides the UCF-QNRF dataset, its composition-loss result, and the large-scale congested benchmark where the method reports MAE 97.5.","marker":"[19]"},{"why":"is the context-aware baseline whose UCF-QNRF and ShanghaiTech-A numbers the proposed method beats and whose Part B and UCF_CC_50 numbers it trails.","marker":"[33]"},{"why":"is the prior top-down feedback fusion that motivates the top-bottom path of the proposed architecture.","marker":"[47]"},{"why":"supplies the Markov random field framework used to combine annotation and segmentation cues for head-size estimation.","marker":"[25]"},{"why":"supplies the SLIC superpixel algorithm used to over-segment the input image in the scale-estimation pipeline.","marker":"[1]"},{"why":"supplies the watershed transformation applied to the distance transform of head locations in the same pipeline.","marker":"[4]"}],"fun_headline_variants":["Two-way scale fusion shrinks crowd-count errors in congested scenes","Bidirectional feature flow improves crowd counting in dense crowds","Cross-scale residual blocks boost crowd-count accuracy in packed scenes","Bottom-top and top-bottom fusion beats single-pass counting nets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the image segment around each annotated head, as measured by the Markov random field combining superpixels and watershed regions, matches the head's actual size in dense and occluded areas; if it does not, the scale-aware supervision is teaching the network to match noisy labels.","fun_headline_variants_meta":{"raw":{"variants":["Two-way scale fusion shrinks crowd-count errors in congested scenes","Bidirectional feature flow improves crowd counting in dense crowds","Cross-scale residual blocks boost crowd-count accuracy in packed scenes","Bottom-top and top-bottom fusion beats single-pass counting nets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000765,"raw_usage":{"total_tokens":3407,"prompt_tokens":972,"completion_tokens":2435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":2367}},"tokens_in":588,"tokens_out":2435,"duration_ms":16490,"temperature":1.0,"reasoning_tokens":2367,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:29:16.880986+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the MRF segment-size estimates with manually measured head widths on a sample of ShanghaiTech-A or UCF-QNRF images, and retrain the SCFB network with density maps built from the measured sizes; if agreement is weak in dense regions, or if the measured-size maps do not preserve the reported MAE 60.2 and 97.5, the scale-aware supervision claim is falsified.","supporting_citations":[{"cited_title":"Single-image crowd counting via multi-column convolutional neural network","cited_arxiv_id":null,"evidence_quote":"provides the ShanghaiTech dataset and the common nearest-neighbor scale-estimation baseline that the MRF approach is designed to improve."},{"cited_title":"Composition loss for counting, density map estimation and localization in dense crowds","cited_arxiv_id":null,"evidence_quote":"provides the UCF-QNRF dataset, its composition-loss result, and the large-scale congested benchmark where the method reports MAE 97.5."},{"cited_title":"Context- aware crowd counting","cited_arxiv_id":null,"evidence_quote":"is the context-aware baseline whose UCF-QNRF and ShanghaiTech-A numbers the proposed method beats and whose Part B and UCF_CC_50 numbers it trails."},{"cited_title":"Top-down feed- back for crowd counting convolutional neural network","cited_arxiv_id":null,"evidence_quote":"is the prior top-down feedback fusion that motivates the top-bottom path of the proposed architecture."},{"cited_title":"Markov random field models in computer vision","cited_arxiv_id":null,"evidence_quote":"supplies the Markov random field framework used to combine annotation and segmentation cues for head-size estimation."},{"cited_title":"Slic superpixels","cited_arxiv_id":null,"evidence_quote":"supplies the SLIC superpixel algorithm used to over-segment the input image in the scale-estimation pipeline."},{"cited_title":"The watershed transformation ap- plied to image segmentation","cited_arxiv_id":null,"evidence_quote":"supplies the watershed transformation applied to the distance transform of head locations in the same pipeline."}],"review_version":1}