{"id":"e10b9fb4-63ab-464f-ae70-a821d3bcf3ae","arxiv_id":"1908.03978","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An adaptive region-division algorithm combines YOLOv3 detection for nearby pedestrians with an inception dilated CNN for distant heads, lowering mean absolute error on a five-scene subway dataset.","lead":"Pedestrian counting method that splits each video frame into a nearby region counted by object detection and a distant region counted by a density-estimation network, with the split line adjusted frame by frame to avoid cutting heads. The method reports a small average improvement over three baselines on five subway scenes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (4) normalizes the ground-truth density map to integrate to 1, not to the pedestrian count, so the reported MAE cannot follow from the method as written; this is a more basic gap than the YOLOv3 dependency.","rationale":"The reader's weakest assumption was the dependence of the region mask and perspective map on YOLOv3 detections. That is a real correctness risk: if the detector misses or mislocates heads, the mask and sigma are wrong. However, Eq. (4) is more load-bearing because it makes the described method internally inconsistent: a density map that integrates to 1 cannot be turned into a pedestrian count without an unstated scaling factor. This concern applies to every frame in the reported experiment, not only to cases where YOLOv3 fails. I therefore keep the reader's CONDITIONAL verdict and recommend that acceptance be contingent on the authors correcting Eq. (4) (or confirming a typo), specifying the count-extraction operation, and ideally releasing code and annotations. Secondary issues, such as lack of error bars and worse performance in two of five scenes, do not by themselves change the verdict. I do not see a reason to move to REJECT, because the inconsistency is plausibly a presentation error, but the conditional is essential.","tokens_in":7121,"tokens_out":7432,"duration_ms":79923,"concrete_test":"Recompute the spatial integral of Di(p) from Eq. (4) for one annotated Subway frame; it will equal 1 by construction. Then reimplement the IDCNN training and the count-extraction used for Eq. (5) exactly as the paper describes. If the predicted density map is summed to obtain z_i, the test MAE will be near the mean count rather than 2.30. Ask the authors to state the exact count-extraction formula or release code; if a missing normalization or division by ||Pi|| at inference is what yields 2.30, Eq. (4) must be corrected before the claim can be reproduced.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 defines the distant-region ground-truth density map as Di(p) = sum_{P in Pi} (1/||Pi||) N(p;P,sigma), with N a normalized Gaussian kernel. Because each Gaussian integrates to 1, the spatial integral of Di(p) is exactly 1 for every frame, independent of the number of pedestrians. The paper gives no count-extraction step that recovers z_i from this normalized map, yet Eq. (5) evaluates MAE on z_i and Section 5.3 reports counts such as GT 117 versus ET 116.1. If z_i were the sum of the predicted density map, the output would saturate near 1 per frame and the MAE would be roughly the mean count (about 28), not the reported 2.30. If some rescaling is applied instead, that rescaling needs the test-time count and is either unstated or leaks annotation information into the estimate. This is an internal inconsistency in the training/evaluation pipeline, and it weakens the central claim more directly than the acknowledged dependence on YOLOv3 detections, because every distant-region training and test result depends on the definition of the target density map.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid pedestrian-counting pipeline for subway surveillance frames. A dynamic boundary, computed from YOLOv3 detections and an expected head-height line H, splits each frame into a nearby region counted by YOLOv3 and a distant region counted by a newly designed inception dilated CNN (IDCNN); the two counts are fused to produce a per-frame total. Experiments on a five-scene subway dataset report an average MAE of 2.30, below the compared baselines He et al. [7], MCNN, and CSRNet. The central claim is that the dynamic region division plus IDCNN yields a lower average MAE than straight-line double-region division and the external baselines.","tokens_in":7393,"tokens_out":7499,"duration_ms":75065,"significance":"The underlying design idea is reasonable and of practical interest: instead of fitting a single density estimator to a scene with highly variable head scale, the paper uses detection where pedestrians are large and density regression where they are small, and it attempts to make the boundary respect head integrity. The multi-dilation inception module that lets the network learn the effective dilation rate is a simple and sensible contribution. The comparison includes external baselines (MCNN, CSRNet) in addition to the authors' prior method [7], which is a useful feature of the experiments. If the empirical results were reproducible, the method would be a credible baseline for scene-adaptive counting. However, the evaluation pipeline is not internally consistent as written, the mask-generation pseudocode is not reproducible, and no ablation isolates the claimed sources of improvement.","major_comments":[{"comment":"As written, Eq. (4) defines D_i(p) = Σ_{P∈P_i} (1/‖P_i‖) N(p;P,σ), where N is a normalized Gaussian. Because each term integrates to unity, the spatial integral of D_i is exactly 1 for every frame, independent of the pedestrian count. The paper never explains how the scalar count z_i in Eq. (5) is recovered from this normalized map. If z_i is the integral or sum of the network output, the estimate is constrained near 1 per frame while the dataset average is 28.78, making the reported MAE of 2.30 impossible. If an additional rescaling is applied, it is not described and would require the test-time count to be known, leaking annotation information. This is a load-bearing inconsistency in the training/evaluation definition, and Table 1 cannot be interpreted until it is resolved.","section":"§4.1, Eq. (4)"},{"comment":"The pseudocode for the dynamic mask contains apparent mistakes that prevent reproduction of the division. In the non-overlap branch, line 14 re-assigns the same submatrix as line 11, and line 15 unconditionally marks columns from br x_{i-1} to Iwidth as distant, which includes the current bounding box and all boxes to its right. This contradicts the stated head-completeness guarantee and the illustration in Figure 3. The overlap branch is also unexplained: it uses tf y_{i-1} rather than H for the upper boundary. Please correct the pseudocode and, ideally, release the mask-generation code so the proposed division rule can be checked.","section":"§3.2, Algorithm 1"},{"comment":"The paper attributes the performance gain to dynamic region division and the IDCNN, but the comparison is not controlled. There is no ablation that replaces the dynamic division with a straight-line division while keeping the same YOLOv3/IDCNN subsystems, and no ablation that removes the IDCNN or the inception module. The reported average improvement over MCNN is 2.30 versus 2.69, yet in Jing'an Temple and South Railway Station the proposed method is worse than MCNN (1.59 versus 1.44 and 2.81 versus 2.35, respectively), and no variance estimates or significance tests are provided. As presented, the specific contribution of the dynamic division is not established.","section":"§5.3, Table 1"},{"comment":"The Gaussian width σ = 0.15 M(p) is obtained by linear regression from YOLOv3 bounding boxes, and the dynamic mask itself is built from those boxes. If YOLOv3 misses or mislocates pedestrians, the distant-region ground truth, the region boundary, and the head-completeness guarantee are all affected. The paper does not report YOLOv3 recall or precision on the nearby region, nor does it analyze the sensitivity of the final counts to detector errors. Because the training target and the region division both depend on the same detector, this is a load-bearing dependency that needs explicit validation.","section":"§4.1 and §3.2"}],"minor_comments":[{"comment":"The MAE formula should be written with an explicit parenthesis: MAE = (1/N) Σ_{i=1}^N |z_i − z_i^gt|; as printed, the markup is ambiguous.","section":"§5.2, Eq. (5)"},{"comment":"The linear-regression step for the perspective map M(p) is described in one sentence; please specify the input features, the number of training boxes, and how M(p) is extrapolated to pixels without detected boxes.","section":"§4.1"},{"comment":"The word 'Ulteriorly' appears to be a word-choice error; 'subsequently' or 'furthermore' would be clearer.","section":"Abstract and Introduction"},{"comment":"The phrase 'The third row column' is garbled; it should say 'the third column'.","section":"Figure 5 caption"},{"comment":"The discussion of occlusion in the Jing'an Temple and South Railway Station scenes is qualitative; provide quantitative occlusion statistics or at least example frames if this explanation is retained.","section":"§5.3"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the corrected evaluation pipeline may change the numbers in Table 1. I recommend requiring the authors to state explicitly whether Eq. (4) reflects the actually implemented density target and to report the count-recovery step; if the normalization is only a typographical error, the paper should be re-evaluated after the text is corrected, and if it is not, the empirical claims should be rejected. The lack of code and the absence of ablation make independent verification difficult, and the editor should weigh that in the revision requirements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nI read the He et al. paper on dynamic region division for pedestrian counting. The core idea is a modest but real step: use YOLOv3 detections to shape the boundary between a nearby region (counted by detection) and a distant region (counted by a density network), so heads don't get bisected by a straight line. The inception-style dilated convolution with parallel dilation rates is a reasonable workaround to the dilation-rate selection problem. The reported average MAE (2.30) is better than the three baselines on the five-scene subway dataset, and the comparison includes external methods (MCNN, CSRNet), not just their own prior work. So far so good.\n\nThe problem is in the details. Equation (4) defines the ground-truth density map as a sum of normalized Gaussians weighted by 1/||Pi||. Each Gaussian integrates to 1, so the integral of Di(p) is exactly 1 for every frame, no matter how many pedestrians are in the scene. The paper never explains how the pedestrian count is recovered from that map. The MAE in Eq. (5) compares counts, and their figures show counts like 117 vs 116.1. If the prediction is the sum of the density map, it would saturate near 1 and the MAE would be about 28, not 2.30. If some rescaling is applied, it is unstated and would need the true count at test time. That is a load-bearing inconsistency, not a typo: every training and test result depends on the definition of the target density.\n\nThe YOLOv3 dependency is a softer concern. The mask and the perspective map M(p) are both fitted from detector boxes, so missed or mislocated detections corrupt the region boundary and the Gaussian sigmas. That is an engineering trade-off, and a proper ablation would show how sensitive the result is. There are no error bars, no significance tests, and no code or annotations released. The method also loses to MCNN in two of the five scenes, so the average gain is fragile.\n\nThe citation pattern is fine; self-citing the prior straight-line method as baseline is appropriate.\n\nWho should read this? Someone building hybrid detection/counting systems for fixed cameras would find the dynamic mask idea worth considering. But the unexplained normalization means the central claim is not supported as written. I would want a revision that derives the count extraction step and provides per-scene variance before trusting the 2.30 number.\n\nRecommendation: send it to peer review — the idea is worth exploring and the flaw is fixable — but reviewers should push hard on the density normalization before accepting any of the quantitative claims.","headline":"A hybrid detection/density counting pipeline with a genuinely dynamic mask, but the normalized density target in Eq. (4) means the reported MAE does not follow from the method as written.","tokens_in":7902,"tokens_out":3943,"would_cite":false,"duration_ms":39950,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Dynamic region division that never cuts a head in half improves pedestrian counting on a five-scene subway benchmark.","keywords":["pedestrian counting","crowd counting","dynamic region division","density map","YOLOv3","dilated convolution","inception module","surveillance video"],"falsifier":"Re-run the pipeline on the same five subway scenes after disabling the detector's bounding boxes and drawing the boundary as a straight line at the same expectation height; if the average MAE does not rise materially above 2.30, the dynamic division itself is not what is carrying the improvement. Alternatively, in a scene where YOLOv3 is known to miss a cluster of nearby pedestrians, the mask will fail to include those heads, and counting that frame will show whether the missed detections propagate into the distant-region density estimate.","tokens_in":6895,"feed_emoji":"🚇","tokens_out":4133,"duration_ms":39413,"temperature":0.7,"pith_summary":"This paper argues that pedestrian counting in perspective-distorted surveillance video improves when the frame is not split by a fixed straight line but by a scene-adaptive boundary that keeps every detected head whole. It claims that a dynamic mask, built from YOLOv3 bounding boxes and an expected division height, lets a detector handle the near field and a density network handle the far field, and that fusing these counts yields lower average counting error than the straight-line double-region method and than two deep baselines on a five-scene subway dataset. A reader should care because the failure mode addressed—cutting one head across a region boundary and fitting large heads with Gaussian kernels—is a real obstacle for crowd counting in high-perspective scenes.","feed_headline":"Dynamic region split cuts pedestrian counting error to MAE 2.30","feed_subtitle":"By drawing the near/far boundary around detected heads, the method beats straight-line and density-only baselines on subway video.","key_machinery":"The carrying mechanism is the dynamic division mask $M$: a binary image, produced by Algorithm 1 from YOLOv3 bounding boxes and expectation height $H$, that assigns each pixel to distant (1) or nearby (0) region while preserving head completeness. The mask also supplies the perspective map via linear regression on detected boxes, setting $\\sigma = 0.15 M(p)$ for Gaussian head kernels in the distant region. The other named component, IDCNN, concatenates three dilated convolution branches (rates 1, 2, 3) in inception layers so the network learns which dilation rate to use instead of requiring the rate to be chosen a priori.","core_discovery":"The paper's central discovery is that the region boundary itself is part of the counting problem: drawing it as a straight line through a perspective-distorted scene splits heads and corrupts both the detection and density estimates. Its dynamic region division uses the detected bounding boxes to compute an expectation height $H$, then generates a mask that follows the tops and sides of those boxes so every head remains either entirely in the nearby region or entirely in the distant region. With this mask, the nearby region is counted by a pedestrian detector and the distant region by an inception dilated convolutional network (IDCNN), and the fused result reports average MAE 2.30 over the five subway scenes, the lowest among the compared methods.","pith_inferences":["If the claim holds, the division boundary, not just the counting network, is a major source of error in crowd counting; future systems could learn the boundary jointly with the counting objective rather than deriving it from detections.","A testable extension would replace YOLOv3 boxes with an oracle detector to isolate how much of the MAE gain comes from the mask itself versus the IDCNN, since both depend on the same detections.","The same mask idea could generalize to other regression-plus-detection hybrids, such as vehicle counting or cell counting in images with large scale variation.","The paper's per-scene failure in occlusion-heavy scenes suggests that dynamic division could be extended to incorporate occlusion cues, not just head completeness, when choosing the boundary."],"forward_implications":["On the subway station dataset, the dynamic-division pipeline reports average MAE 2.30, lower than the straight-line double-region method (2.75), MCNN (2.69), and CSRNet (2.82), with the best per-scene result on three of the five scenes.","Fusing detection for the nearby region and density estimation for the distant region avoids using Gaussian kernels on large heads and avoids detector failures in dense distant crowds, so the method extends to scenes where neither approach alone is reliable.","The dynamic mask can be recomputed per frame, so the region division adapts to pedestrian layout rather than being fixed by camera geometry.","IDCNN removes the need to tune dilation rate by letting the network combine multiple rates, which can be applied to other counting or segmentation tasks that use dilated convolutions."],"supporting_citations":[{"why":"Supplies the straight-line double-region baseline that the dynamic division improves, along with the subway dataset and the MAE evaluation protocol.","marker":"[7]"},{"why":"YOLOv3 provides the bounding boxes used for both the dynamic mask and the nearby-region detection.","marker":"[16]"},{"why":"CSRNet is the dilated-convolution baseline and the source of the dilation-rate selection problem that IDCNN addresses.","marker":"[11]"},{"why":"MCNN is a compared density-based baseline and the origin of geometry-adaptive Gaussian kernels for heads.","marker":"[9]"},{"why":"Introduces the density-map formulation with Gaussian kernels that the distant-region ground truth uses.","marker":"[2]"},{"why":"The inception module inspires concatenating multiple dilated convolution outputs so the network learns the dilation rate.","marker":"[17]"},{"why":"Provides the perspective-map idea and the manual-annotation approach that the paper replaces with regression from detector boxes.","marker":"[8]"}],"fun_headline_variants":["Head-aware region split lowers pedestrian count error to 2.30","Adaptive boundary keeps heads whole, cuts counting MAE to 2.30","Dynamic split by head boxes beats straight-line counting","Counting crowds with a smarter split: MAE 2.30 on subway"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the detector finds every nearby pedestrian's head and places it correctly, because a missed or misplaced detection makes the region mask, the head-completeness guarantee, and the Gaussian kernel widths for distant heads all wrong.","fun_headline_variants_meta":{"raw":{"variants":["Head-aware region split lowers pedestrian count error to 2.30","Adaptive boundary keeps heads whole, cuts counting MAE to 2.30","Dynamic split by head boxes beats straight-line counting","Counting crowds with a smarter split: MAE 2.30 on subway"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000433,"raw_usage":{"total_tokens":2169,"prompt_tokens":867,"completion_tokens":1302,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":1226}},"tokens_in":483,"tokens_out":1302,"duration_ms":10590,"temperature":1.0,"reasoning_tokens":1226,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:55:56.422361+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the pipeline on the same five subway scenes after disabling the detector's bounding boxes and drawing the boundary as a straight line at the same expectation height; if the average MAE does not rise materially above 2.30, the dynamic division itself is not what is carrying the improvement. Alternatively, in a scene where YOLOv3 is known to miss a cluster of nearby pedestrians, the mask will fail to include those heads, and counting that frame will show whether the missed detections propagate into the distant-region density estimate.","supporting_citations":[{"cited_title":"The novel dynamic region division can meet the challenge of perspective distortion and avoid to cut the head into two parts","cited_arxiv_id":null,"evidence_quote":"Supplies the straight-line double-region baseline that the dynamic division improves, along with the subway dataset and the MAE evaluation protocol."},{"cited_title":"Single-image crowd counting via multi-column convolutional neural network,","cited_arxiv_id":null,"evidence_quote":"YOLOv3 provides the bounding boxes used for both the dynamic mask and the nearby-region detection."},{"cited_title":"Learning to count objects in images,","cited_arxiv_id":null,"evidence_quote":"MCNN is a compared density-based baseline and the origin of geometry-adaptive Gaussian kernels for heads."},{"cited_title":"2) Appropriate learn- ing models are applied to count pedestrians in each obtained region","cited_arxiv_id":null,"evidence_quote":"Introduces the density-map formulation with Gaussian kernels that the distant-region ground truth uses."},{"cited_title":"Body structure aware deep crowd counting,","cited_arxiv_id":null,"evidence_quote":"The inception module inspires concatenating multiple dilated convolution outputs so the network learns the dilation rate."},{"cited_title":"Research on the impact of crowd flow on crowd risk in large gathering spots,","cited_arxiv_id":null,"evidence_quote":"Provides the perspective-map idea and the manual-annotation approach that the paper replaces with regression from detector boxes."}],"review_version":1}