{"id":"fa566876-c6a9-446a-abfe-5221e01287db","arxiv_id":"1908.08652","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MTCNet improves crowd counting accuracy by adding an auxiliary count-group classification task to a shared VGG-16 and dilated-convolution density estimator.","lead":"This paper presents MTCNet, a neural network that counts people in crowds by combining a density-map prediction task with a count-range classification task. It reports lower average counting errors than earlier methods on three public crowd datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed MTL gain is not isolated: Section 4.5 lacks a capacity/control ablation and λ is tuned on test data, so the auxiliary task is not established as the cause of the reported improvement.","rationale":"The paper's central claim is causal: adding crowd-count group classification as an auxiliary task reduces counting error. The only internal evidence is the Section 4.5 ablation, which compares joint training (63.7 MAE) with main-task-only training (68.2 MAE). That comparison is the load-bearing point: if the gain is actually due to the extra classifier's parameters, the auxiliary loss, or test-set-tuned λ, then the paper has not shown what it claims. The reader's weakest_assumption identifies exactly this missing control, and the concrete random-label test would settle it. Other issues, such as absent code and missing error bars, weaken confidence but are secondary; they do not directly test the causal mechanism. Because the reader already issued a CONDITIONAL verdict for these reasons, the stress-test does not change the verdict. The paper should be accepted only if the control ablation supports the MTL attribution, or revised to soften the causal claim.","tokens_in":6870,"tokens_out":3417,"duration_ms":40824,"concrete_test":"Train the joint MTCNet exactly as in the paper but replace the auxiliary count-group labels with random permutations (or a fixed dummy label), keeping architecture, optimizer, λ, and epoch count identical, and report main-task MAE on ShanghaiTech Part A and Part B averaged over at least three seeds; if MAE stays near 63.7/8.6, the improvement over Section 4.5's standalone baseline is not due to meaningful count-group supervision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central assertion is that the count-group classifier improves density estimation. The only direct evidence is Section 4.5: main-task-only MAE 68.2 vs joint MAE 63.7. This comparison does not hold the model class fixed: the joint model adds a classifier head (three dense layers) and a second loss, changing parameter count and optimization landscape. The standalone number 68.2 appears to reproduce CSRNet, so the comparison may be between two different networks rather than an ablation of one architecture with and without auxiliary supervision. Additionally, Section 4.4 selects λ=1e-3 by trying values on ShanghaiTech Part A and reporting the best MAE; no validation split is described, so the headline 63.7 may be a test-set-selected optimum. With no error bars or repeated runs, the 4.5-point gap cannot be confidently attributed to the MTL mechanism. If the gain survives a control with uninformative auxiliary labels, the concern is resolved; if not, the paper's central claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MTCNet, a multi-task learning architecture for crowd counting. The main task is density-map regression with a VGG-16 front-end and a dilated convolutional backend; the auxiliary task is a ten-class crowd-count-group classifier that shares the front-end. The combined loss is MSE plus a weighted cross-entropy term (Eq. 6). The authors report lower MAE and MSE than prior state-of-the-art methods on ShanghaiTech Part A/B and UCF_CC_50, and they include a weight-factor sensitivity study (Section 4.4) and an ablation showing that joint training improves the main-task MAE from 68.2 to 63.7 on ShanghaiTech Part A (Section 4.5).","tokens_in":7088,"tokens_out":6241,"duration_ms":62904,"significance":"If the reported results hold, the paper demonstrates a simple and effective use of multi-task learning for crowd counting, with claimed MAE improvements of 5.8%–14.9% over competing methods on standard benchmarks. The architectural idea is clearly presented, and the paper includes an ablation, a sensitivity analysis, and comparisons on two widely used datasets. The main strengths are the straightforward integration of an auxiliary count-group classifier and the consistency of the reported gains across three benchmark settings. However, the current experimental design does not yet isolate the effect of the auxiliary task: the ablation in Section 4.5 compares models that differ in architecture and supervision, and the hyperparameter λ is selected using test-set MAE in Section 4.4. Additional control experiments and a proper validation procedure are needed before the central claim can be considered established.","major_comments":[{"comment":"The ablation compares a main-task-only MAE of 68.2 with a joint-training MAE of 63.7, but the model is not held fixed between the two settings. The joint model adds a classifier head consisting of three dense layers and a second loss term, and the paper does not specify whether the standalone model retains the X2 feature pathway and the X1-X2 concatenation when the auxiliary branch is removed. A capacity-matched control—for example, training the same two-branch architecture with the auxiliary loss applied to randomly shuffled or fixed class labels, or with the auxiliary branch present but its loss detached—is necessary to determine whether the 4.5-point MAE improvement is due to the count-group supervision itself or to the extra parameters and changed optimization landscape. Without such a control, the paper's central claim that the auxiliary task improves density estimation is not established.","section":"Section 4.5, 'Effect of MTL setting'"},{"comment":"Table 3 reports MAE and MSE on ShanghaiTech Part A for five values of λ and selects λ = 1e-3 as optimal because it gives the lowest MAE, but the paper does not describe any validation split or nested cross-validation procedure. The test set is therefore used for model selection, which makes the reported 63.7 MAE an optimistic, test-set-selected number. The claim of a 5.8% improvement over CSRNet may be inflated by this selection effect. Please use a held-out validation split to choose λ and report the final test performance only once, together with an estimate of variability across runs.","section":"Section 4.4, 'Effect of the weight factor λ'"},{"comment":"Table 2 reports only a single MAE and MSE for the 5-fold cross-validation on UCF_CC_50, which contains only 50 images. Given the small dataset and the known sensitivity of crowd-counting results to the fold split and to random initialization, a single aggregate number does not support the claim of a 10.5% improvement over the second-best method. Please report per-fold results and the mean ± standard deviation across folds, and preferably across multiple random seeds as well.","section":"Section 4.3, Evaluation on UCF_CC_50"}],"minor_comments":[{"comment":"Equation (3) is rendered illegibly; the count-group label formula should be typeset properly, for example as class_label = min(round(9 * (C_i^GT - C_min)/(C_max - C_min)), 9). Please also clarify whether C_min and C_max are computed on the training split only and how test counts outside that range are handled.","section":"Section 2.3, Eq. (3)"},{"comment":"The flow of X1 and X2 into the density branch and the placement of the AdaptiveMaxPool2d layer are ambiguous. A short textual description of the tensor shapes at each concatenation and pooling step would help readers reproduce the architecture.","section":"Figure 2 and Section 2"},{"comment":"The training details are incomplete: the paper states a fixed learning rate of 1e-7 but does not report batch size, number of epochs, weight decay, or whether all hyperparameters were held fixed across the λ values in Table 3. These details are needed for independent verification.","section":"Section 2.4, Training details"},{"comment":"The paper reports the auxiliary-task classification accuracy only for the selected λ (56.6% standalone vs. 79.7% in MTL). Showing the auxiliary-task accuracy for each λ value would help confirm that the auxiliary branch actually learns the count-group structure and would strengthen the interpretation of the λ sensitivity results.","section":"Section 4.4, Table 3"},{"comment":"The comparison with prior methods does not state whether the competing methods used data augmentation. Since the paper emphasizes that its own results are obtained without data augmentation, a note on the augmentation settings of the baselines would make the comparison fairer.","section":"Section 4.2, Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the proposed idea is plausible, but the experimental validation as presented is not yet sufficient to support the central attribution claim. The required control experiments and proper validation-split methodology are feasible within the paper's scope, so I recommend a major revision rather than rejection. I would also encourage the authors to release code and trained models, since the missing training details currently hinder reproduction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I largely agree with the reader's conditional verdict. MTCNet is a straightforward and clearly presented extension of CSRNet: add a ten-bin crowd-count classifier as an auxiliary task sharing the VGG-16 front-end, train with a weighted sum of MSE and cross-entropy. The reported MAE/MSE beats the cited SOTA on ShanghaiTech A/B and UCF_CC_50, and the ablation shows both tasks improve in joint training. That is a genuine new result in the sense that these numbers and this specific combination are not in the literature.\n\nWhat the paper does well: the writing is clear, the architecture is specified well enough to re-implement, and the comparison uses standard benchmarks and metrics. The λ sweep is shown in full, which at least lets readers see sensitivity. The ablation direction is right: main-task-only MAE 68.2 vs joint 63.7, and auxiliary classification accuracy 56.6% vs 79.7%. Those are meaningful improvements if they hold.\n\nWhere it falls short: the central attribution problem is not resolved. Section 4.5 compares a network trained with only the density loss against the same base network plus a classifier head and a second loss. The classifier head adds parameters and changes the optimization landscape, so the 4.5-point MAE gap cannot be confidently assigned to the auxiliary task's semantic signal. A control with shuffled or uninformative auxiliary labels would settle this; without it, the paper's main claim is under-identified. Second, λ=1e-3 is selected directly from Part A test-set MAE. Five values, so the risk is modest, but it still means the headline number is a test-selected optimum and comparisons to published numbers that didn't tune on test data are not apples-to-apples. Third, there are no error bars or per-fold results on UCF_CC_50, so we can't judge stability. These are all fixable with additional experiments, not theoretical issues.\n\nBottom line: this is a credible, well-scoped empirical paper. It deserves a serious referee. With a proper control ablation and a validation-based λ selection, the MTL claim would be convincing. As it stands, I'd treat the SOTA numbers as promising but not fully substantiated.","headline":"Decent MTL-for-crowd-counting paper with new SOTA numbers, but the ablation doesn't fully isolate the auxiliary task and λ is chosen on the test set.","tokens_in":7564,"tokens_out":3542,"would_cite":false,"duration_ms":35246,"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":"The paper's central claim is that adding a crowd-count group classification task to a dilated-convolution density estimator lowers counting error, with the best reported MAE on ShanghaiTech and UCF_CC_50.","keywords":["crowd counting","multi-task learning","density estimation","dilated convolution","crowd-count group classification","ShanghaiTech dataset","UCF_CC_50 dataset","VGG-16"],"falsifier":"Train a single-task density estimator with the same number of parameters and the same optimization schedule as MTCNet—for instance, keep the classifier branch architecture but set its loss weight to zero or freeze it—and compare MAE on ShanghaiTech Part A. If this control reaches the same 63.7 MAE, the auxiliary task is not the cause of the improvement.","tokens_in":6711,"feed_emoji":"👥","tokens_out":6808,"duration_ms":62938,"temperature":0.7,"pith_summary":"This paper tries to show that a crowd-density counting network gets better when it is trained at the same time on a second, simpler task: sorting each image's crowd count into one of ten ranges. The joint model, MTCNet, shares a VGG-16 feature extractor between the density-regression branch and the classification branch. On ShanghaiTech Part A, Part B, and UCF_CC_50, it reports mean absolute errors below the best previous methods—5.8%, 14.9%, and 10.5% lower, respectively—without data augmentation. If true, this matters because scale variation is a main cause of error in crowd counting, and the auxiliary classifier is a cheap way to make shared features sensitive to crowd-density scale.","feed_headline":"Side-task classifier cuts crowd-counting error by up to 14.9%","feed_subtitle":"Training a density network with a ten-group count classifier beats prior methods on three benchmarks.","key_machinery":"The load-bearing mechanism is the shared VGG-16 front-end followed by a two-branch split: the density branch uses dilated convolutions to regress the density map, while the auxiliary branch classifies the shared features into ten equal-width crowd-count groups. The auxiliary branch's ground-truth labels are obtained by binning each image's count into one of ten equal intervals between the minimum and maximum training counts. This classifier is what carries the argument, because it forces the shared features to separate images by density scale, and the combined MSE plus weighted cross-entropy loss propagates that scale information back into the feature extractor.","core_discovery":"MTCNet is a multi-task network whose main branch is a CSRNet-style dilated-convolution density estimator and whose auxiliary branch is a ten-class crowd-count group classifier. Both branches share the first ten layers of VGG-16. The auxiliary branch takes the 128-feature-map block X2, passes it through convolutional and fully connected layers, and outputs a softmax over ten count ranges; the main branch concatenates X1 (512 maps) with X2 and regresses a density map using dilation rate two. The combined loss is the density MSE plus λ times the classification cross-entropy, with λ=1e-3 giving the best result. In the paper's ablation, the joint model reaches 63.7 MAE on ShanghaiTech Part A while the main task alone reaches 68.2, and the auxiliary classifier's accuracy rises from 56.6% standalone to 79.7% in joint training.","pith_inferences":["The paper leaves implicit whether the benefit comes from the classification labels per se or from the architectural regularization of adding a small output branch; a testable extension is to replace the ten-class classifier with an auxiliary branch predicting a scalar count via regression and compare MAE.","Because the auxiliary branch consumes only X2, the count-group classifier may be blind to the deeper X1 features; attaching the classifier to the concatenated features might sharpen scale separation further.","The ten-way binning is arbitrary; using ordinal regression or fewer bins might retain most of the gain with fewer parameters, which would help distinguish the task-supervision effect from capacity effects."],"forward_implications":["A coarse count-range classifier can be added to existing dilated-convolution density estimators as a drop-in auxiliary head and reduce MAE on both dense and sparse scenes.","Joint training with the auxiliary task works without data augmentation, so the gain is not dependent on extra training data.","The optimal weight of the auxiliary loss is small (1e-3), so the classification loss helps most as a weak regularizer rather than as a dominating objective.","The auxiliary classifier's standalone accuracy of 56.6% vs 79.7% under joint training implies density regression supplies useful signal back to the classification task, making the two tasks mutually reinforcing."],"supporting_citations":[{"why":"Supplies the ShanghaiTech dataset, its train-test split, and the MCNN baseline that MTCNet compares against.","marker":"[9]"},{"why":"Supplies the CSRNet dilated-convolution density estimator that the main task is built on, including hyperparameters and dilation rate two.","marker":"[11]"},{"why":"Supplies the UCF_CC_50 dataset and the five-fold cross-validation protocol used for evaluation.","marker":"[16]"},{"why":"Supplies the multitask-learning rationale of shared representations, inductive bias, and regularization.","marker":"[12]"},{"why":"Supplies the pretrained VGG-16 layers used as the shared front-end.","marker":"[13]"},{"why":"Supplies dilated convolution, the mechanism that enlarges receptive field without pooling.","marker":"[14]"}],"fun_headline_variants":["Side-task classifier cuts crowd-count error by 14.9%","Multi-task network improves crowd counting by 14.9%","Auxiliary count groups sharpen density estimates","Sharing features with count classifier beats single-task","MTCNet: auxiliary group classifier boosts accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central assumption is that the counting improvement comes from the auxiliary classification supervision itself; the ablation does not control for the extra parameters of the classifier branch, so the gain could in principle come from a larger or differently trained network.","fun_headline_variants_meta":{"raw":{"variants":["Side-task classifier cuts crowd-count error by 14.9%","Multi-task network improves crowd counting by 14.9%","Auxiliary count groups sharpen density estimates","Sharing features with count classifier beats single-task","MTCNet: auxiliary group classifier boosts accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000146,"raw_usage":{"total_tokens":1169,"prompt_tokens":920,"completion_tokens":249,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":172}},"tokens_in":536,"tokens_out":249,"duration_ms":3049,"temperature":1.0,"reasoning_tokens":172,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:32:59.050755+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a single-task density estimator with the same number of parameters and the same optimization schedule as MTCNet—for instance, keep the classifier branch architecture but set its loss weight to zero or freeze it—and compare MAE on ShanghaiTech Part A. If this control reaches the same 63.7 MAE, the auxiliary task is not the cause of the improvement.","supporting_citations":[{"cited_title":"Learning to count objects in images,","cited_arxiv_id":null,"evidence_quote":"Supplies the ShanghaiTech dataset, its train-test split, and the MCNN baseline that MTCNet compares against."},{"cited_title":"Deep people counting in extremely dense crowds,","cited_arxiv_id":null,"evidence_quote":"Supplies the CSRNet dilated-convolution density estimator that the main task is built on, including hyperparameters and dilation rate two."},{"cited_title":"CSRNet: Dilated Convolutional Neural Networks for Understanding the Highly Congested Scenes,","cited_arxiv_id":null,"evidence_quote":"Supplies the UCF_CC_50 dataset and the five-fold cross-validation protocol used for evaluation."},{"cited_title":"Learning to count with cnn boosting,","cited_arxiv_id":null,"evidence_quote":"Supplies the multitask-learning rationale of shared representations, inductive bias, and regularization."},{"cited_title":"Cross-scene crowd counting via deep co nvolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained VGG-16 layers used as the shared front-end."},{"cited_title":"Single - Image Crowd Counting via Multi -Column Convolutional Neural Network,","cited_arxiv_id":null,"evidence_quote":"Supplies dilated convolution, the mechanism that enlarges receptive field without pooling."}],"review_version":1}