{"id":"12f5f8f3-b164-48ac-9090-76e1ba9f8b2c","arxiv_id":"2504.13405","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ProgRoCC trains a CLIP model with approximate count labels, predicts counts digit by digit (hundreds, tens, units), and beats prior weakly and semi-supervised crowd counters on SHA, QNRF, and JHU++.","lead":"This paper proposes a new crowd-counting setup where each training image needs only a rough estimate of the number of people, instead of expensive point annotations. A CLIP-based progressive digit-prediction model, ProgRoCC, is shown to outperform existing weakly and semi-supervised counting methods on three standard benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central empirical claim rests on an unvalidated equivalence: synthetic ±5% uniform rough labels (Section 4.2) stand in for real human estimates, yet no experiment uses real rough labels. If real estimates are biased or density-dependent, the reported gains over SSL-FT and SEEM could shrink.","rationale":"The reader's weakest_assumption identifies exactly the concern that I find most load-bearing: the paper's contribution is a new annotation paradigm, and its headline numbers are obtained with synthetic labels that mimic expert estimates but are never compared to real human estimates. All experiments in Tables 1, 4, and 7 depend on this proxy. The ablation in Table 3 varies the width of the uniform error range but never changes the error distribution, so it cannot rule out bias or density-dependent noise. I considered whether the 999-count cap or the lack of code and error bars should be the primary concern, but those are secondary: the cap is acknowledged and only affects very dense scenes, and missing code does not directly threaten the logic of the method. The central idea remains plausible and the method is coherent, but until the rough-label simulation is validated against real annotation behavior, the claim that rough labels are an easier and sufficient supervision signal is conditional. The reader's CONDITIONAL verdict is appropriate, and my stress-test pass does not require moving it.","tokens_in":11865,"tokens_out":3605,"duration_ms":35764,"concrete_test":"Collect real rough count estimates from human annotators (e.g., 5–10 estimates per image) on a held-out subset of 100 SHA and 100 QNRF training images, using an interface that shows only the image and asks for an approximate count with no point-level tools. Train ProgRoCC with the exact pipeline of Section 4.2 but with these real estimates as the rough labels, and evaluate on the standard SHA and QNRF test splits. Compare MAE and MSE against the synthetic-label model, SSL-FT, and SEEM. If the real-label model degrades by more than roughly 10% relative MAE or falls behind SSL-FT or SEEM, the central claim requires qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the unvalidated equivalence between synthetic and real rough labels. Section 4.2 generates training labels by uniformly sampling within ±5% of the true count, drawing n=10 simulated expert estimates and then sampling one training label between the minimum and maximum of those estimates. All headline results in Table 1 are produced by models trained on this synthetic noise, and the paper's central claim is that rough labels are easier to acquire than point annotations. However, no experiment collects estimates from human annotators, and no analysis tests alternative error distributions such as systematic underestimation in dense crowds, heavy-tailed errors, or image-dependent bias. The introduction itself observes that enumeration-based counting underestimates large crowds, so the most plausible real-world error process is not symmetric and uniform. If real human rough labels deviate from the simulated distribution, the digit-wise progressive targets in Section 3.1 become systematically miscalibrated, and the observed margin over SSL-FT and SEEM could shrink or disappear. This is not an internal inconsistency in the optimization, but it is a genuine external-validity gap in the central empirical claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ProgRoCC, a CLIP-based method for 'rough crowd counting,' in which training images are annotated with approximate count labels rather than point-level or exact-count labels. The method predicts a count digit-by-digit (hundreds, tens, units) through a progressive estimation learning strategy, which reduces the number of image-text similarity computations at inference to 30 comparisons for counts up to 999. A visual-language matching adapter is then used to refine visual features using stored key-value pairs. Experiments on SHA, QNRF, and JHU++ report MAEs of 70.0, 114.8, and 92.5, respectively, outperforming the compared semi-supervised and unsupervised methods, and cross-dataset experiments show favorable transferability relative to the same baselines. The central claim is that rough labels, which are claimed to be easier to acquire than point annotations, can provide better accuracy than semi- and weakly-supervised alternatives.","tokens_in":12080,"tokens_out":6698,"duration_ms":69417,"significance":"If validated, the paper would make a useful contribution: it proposes a new annotation paradigm for crowd counting, demonstrates that a CLIP-based progressive digit-classification scheme can learn from noisy count-level labels, and reports consistent gains over several semi-supervised and unsupervised baselines across three datasets. The ablation study isolates the contributions of the progressive strategy and the matching adapter, and the cross-dataset evaluation is a valuable addition. The efficiency gain over direct CLIP matching is also clearly demonstrated. However, the significance is currently tempered by the fact that all rough labels are generated synthetically with a specific uniform noise model, so the paper's motivating claim that rough labels are easier to acquire and behave like human estimates is not directly tested.","major_comments":[{"comment":"The central claim that rough labels are 'easier to acquire' and yield better accuracy rests on an unvalidated equivalence between synthetic and real rough labels. All experiments use labels drawn uniformly from a ±5% interval around the true count, with n=10 simulated 'expert' estimates and then a training label sampled between the min and max of those estimates. No experiment uses labels produced by human annotators, and the paper's own motivation in Section 1 states that enumeration-based estimates in dense crowds are biased toward underestimation. A realistic error process is therefore likely to be biased, heavy-tailed, or density-dependent rather than symmetric and uniform. If real human rough labels deviate from this model, the progressive digit targets in Section 3.1 become miscalibrated and the reported margins over SSL-FT and SEEM could shrink. Please validate the noise model with a human-annotation study on a subset of images, or at minimum stress-test the method under alternative error models such as multiplicative log-normal noise, systematic undercounting, and density-dependent bias, and report the resulting MAE/MSE.","section":"Section 4.2 and Table 3"},{"comment":"The visual-language matching adapter is a key component of the method (Table 2 shows a 2.1 MAE improvement on SHA), but the manuscript does not specify how the M key-value pairs are initialized, when the adapter is updated relative to the contrastive training in Eq. (5), whether the update rules in Eqs. (6)-(8) are applied per batch or per epoch, or how the text features u_i are obtained for the value update in Eq. (8). The adapter appears to be updated by non-differentiable nearest-neighbor rules, yet it is used to refine visual features during inference, so the training procedure must be described precisely. Without these details or released code, the adapter component is not reproducible and the reported gains cannot be independently verified.","section":"Section 3.2 and Section 4.2"},{"comment":"The method hard-limits the counting range to 0-999, and Table 5 shows that ProgRoCC's MSE degrades sharply in the [800,∞) interval (161.1 MSE versus 141.7 for CCTrans), which is consistent with the model being unable to predict counts above 999. It is unclear whether the model operates on whole images resized to 224×224 or on image patches; if whole images are used, high-density images are systematically underestimated by construction. Since QNRF contains many images with more than 1,000 people, the headline QNRF and JHU++ numbers in Table 1 partially reflect this truncation. Please clarify the whole-image versus patch protocol and quantify the effect of the 999 cap, or qualify the claim of robustness on very dense crowds.","section":"Section 3.1 and Table 5"}],"minor_comments":[{"comment":"There are several typos, including 'nosiy lables' and 'highly nosiy lables', and 'board knowledge' in Section 1 should be 'broad knowledge'.","section":"Section 4.4.2"},{"comment":"In the text of Section 4.4, TransCrowd is cited as reference [27], but TransCrowd is reference [12] in the bibliography; CCTrans is [27]. Please correct the citations.","section":"Section 4.4"},{"comment":"The text prompt template used for the digit labels is never stated. Please provide the exact template, for example 'There are approximately {number} people in the photo', since the prompt wording directly affects CLIP matching.","section":"Section 3.1"},{"comment":"The figure caption says that green straight arrows and black straight arrows denote the training stages of PEL and VlMA, but the figure does not appear to include a legend; please clarify the arrow colors in the figure itself.","section":"Section 3.2 and Figure 2"},{"comment":"The statement that performance is 'relatively stable' at error sampling rates between ±15% and ±50% is only weakly supported by Table 3, since the MAE varies between 119.8 and 133.6 in that range; please either quantify this variability or soften the wording.","section":"Section 4.4.2"}],"recommendation":"major_revision","confidential_remarks":"The technical core of the paper is plausible and the experiments are reasonably broad, but the absence of any validation with real human rough labels is a substantive gap for a paper whose central claim is about annotation cost and human estimation behavior. The reproducibility gaps around the matching adapter are also important. I would be willing to accept a revision that adds a human-label validation study or convincingly stress-tests the method under realistic noise models, and that provides a complete training description for the adapter."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. It proposes rough crowd counting: train with approximate count labels instead of point annotations, using CLIP and digit-by-digit progressive prediction (hundreds, tens, units) plus a key-value adapter. The task framing is new relative to the cited literature, the three-dataset evaluation is broad, and the efficiency point is real—30 text matches instead of 1,000. The ablation showing TransCrowd and CCTrans degrade badly on rough labels while ProgRoCC does not is genuinely informative. The method is coherent, the experiments are not obviously cherry-picked, and the citation pattern looks relevant rather than padded.\n\nThe soft spots are mostly about external validity, not internal execution. Rough labels are generated by uniform sampling within ±5% of the true count (Section 4.2) to simulate estimates from different experts, but no experiment actually collects rough labels from human annotators. The paper's own introduction says enumeration-based estimates underestimate dense crowds, so the most plausible human error is biased and density-dependent, not symmetric uniform. If real rough labels deviate from that model, the reported gains over SSL-FT and SEEM could shrink. That matters because \"easier to acquire\" is the central premise. On top of that, the ±5% default is chosen from the Table 3 ablation on QNRF, and the paper does not say whether that is a held-out split; there are no error bars and no code. The 999-count cap is acknowledged and visibly hurts MSE in dense scenes, which is an honest but real limitation.\n\nNone of this makes the method incoherent. The training is standard supervised learning on rough labels, there is no circularity in the objective, and the paper is honest about its own limitations. Who this is for: anyone working on reducing counting annotation cost, or on adapting CLIP to regression via digit decomposition. I would bring it to a reading group to discuss the synthetic-noise equivalence, and I would send it to peer review with a request for human-annotation experiments or at least a sensitivity analysis over realistic error models. The task framing alone is worth citing.","headline":"A genuinely new task framing and a clever CLIP digit-wise counting scheme, but the core claim rests on synthetic ±5% labels standing in for human rough estimates—worth reviewing, not yet settled.","tokens_in":12641,"tokens_out":2507,"would_cite":true,"duration_ms":24809,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Rough count labels—approximate per-image totals instead of point annotations—suffice to train crowd counting that beats semi-supervised methods and rivals fully supervised ones.","keywords":["crowd counting","rough labels","weakly supervised learning","vision-language models","progressive estimation","digit-wise classification","approximate annotation"],"falsifier":"Collect rough count labels from human annotators on SHA, QNRF, and JHU++ without revealing true counts, train ProgRoCC on these real rough labels, and compare its MAE/MSE against the results obtained with synthetic ±5% labels. If performance degrades toward or below the semi-supervised baselines (for instance, SSL-FT), the equivalence between simulated and real rough labels—and with it the central claim—is not supported.","tokens_in":11639,"feed_emoji":"👥","tokens_out":3972,"duration_ms":37743,"temperature":0.7,"pith_summary":"The paper introduces rough crowd counting, a supervision setting where each training image carries only an approximate total count rather than expensive point-level or exact-count annotations. It argues that this cheaper supervision is enough: a CLIP-based model, trained to predict counts digit by digit with a progressive estimation strategy, reports MAEs of 70.0 on SHA, 114.8 on QNRF, and 92.5 on JHU++. These results outperform semi-supervised baseline SSL-FT and unsupervised SEEM by wide margins and come close to fully supervised CSRNet on SHA. The claim matters because rough labels are far easier to collect, which could make large-scale crowd counting practical in new domains.","feed_headline":"Approximate labels beat semi-supervised crowd counting","feed_subtitle":"A CLIP model that predicts counts digit by digit needs only rough per-image totals and rivals fully supervised methods.","key_machinery":"The central mechanism is progressive estimation learning (PEL), which turns count prediction into three classification steps—hundreds, tens, and units digits—each over ten text prompts. This reduces inference to 30 contrastive matches, down from up to 1,000 for direct count-matching, and lets the model leverage CLIP's pretrained visual-text alignment. A second component, the visual-language matching adapter (VlMA), maintains $M$ key-value pairs; during training it updates keys and values based on the distance between queried values and text embeddings, and during inference it blends the stored value with the visual embedding before matching. Together, PEL provides the coarse-to-fine supervision and VlMA refines the visual features for more reliable matching.","core_discovery":"On the paper's own terms, the central discovery is that a CLIP model can learn to count crowds from approximate per-image totals if the count is decomposed into digit-wise classification tasks. Instead of matching an image against prompts for every possible count, ProgRoCC predicts the hundreds, tens, and units digits sequentially, using only 30 image-text matches to estimate counts up to 999. A visual-language matching adapter stores effective key-value alignments between visual and textual features and refines the visual embedding at inference. The paper reports that this approach outperforms current semi-supervised and unsupervised counting methods on three benchmarks, and that it transfers across datasets better than several fully supervised methods.","pith_inferences":["Editorial inference: the paper's simulation of rough labels as uniform random samples within ±5% of the true count has not been validated against real human rough estimates; human errors are likely biased toward underestimation in dense crowds and may be scene-dependent, so real-world performance could differ.","Editorial inference: the digit-wise decomposition is a general strategy that could be applied to other vision-language quantity estimation tasks, such as object counting in open-vocabulary settings or estimating lengths, areas, and other scalar attributes from images.","Editorial inference: if rough labels prove sufficient in practice, crowd counting datasets could be created at web scale by crowdsourcing approximate totals, enabling rapid retraining for new camera geometries, weather conditions, or object categories without per-target annotation.","Editorial inference: a testable extension is to combine rough count labels with a small number of point annotations, which might close the remaining gap to fully supervised methods while keeping annotation cost low."],"forward_implications":["Rough labels could replace point annotations for building large crowd-counting datasets, because annotators only need to give an approximate total per image instead of locating every person.","Digit-wise progressive prediction lets a CLIP-based counter scale to larger counts with a linear increase in the number of matches: roughly 30 matches for counts under 1,000 and about 300 for counts under 10,000.","The reported performance on SHA (70.0 MAE) is within 1.8 MAE of fully supervised CSRNet, suggesting that cheap approximate supervision can substitute for exact supervision in some density regimes.","Cross-dataset experiments show that ProgRoCC trained on SHA transfers to QNRF and JHU++ with lower MAE than fully supervised CSRNet and HMoDE+REL, indicating that rough-label training may generalize better to unseen scenes.","The method's robustness to label noise—performance stays relatively stable across error sampling ranges from ±15% to ±50%—implies it could tolerate imprecise human estimates in practice."],"supporting_citations":[{"why":"Provides the CLIP vision-language backbone whose pretrained alignment ProgRoCC adapts for counting.","marker":"[22]"},{"why":"Supplies the SHA dataset and the MCNN baseline used for comparison.","marker":"[31]"},{"why":"Supplies the dense QNRF dataset that stresses counting in highly congested scenes.","marker":"[7]"},{"why":"Supplies the JHU++ dataset covering varied weather conditions and large-scale evaluation.","marker":"[26]"},{"why":"SSL-FT is the semi-supervised baseline ProgRoCC claims to outperform by 14.7% MAE on SHA and 24.0% MAE on QNRF.","marker":"[30]"},{"why":"SEEM is the unsupervised baseline ProgRoCC outperforms by large margins, demonstrating the value of rough labels.","marker":"[28]"},{"why":"CSRNet is the fully supervised baseline whose SHA performance ProgRoCC approaches, anchoring the comparison.","marker":"[11]"},{"why":"CrowdCLIP is a prior CLIP-based unsupervised counting method that ProgRoCC extends and outperforms.","marker":"[14]"}],"fun_headline_variants":["Rough crowd totals beat semi-supervised counting","CLIP counts crowds digit-by-digit from rough labels","Progressive digit estimation for rough crowd counting","Vision-language adapter improves rough counting","Counting crowds without per-target annotations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's reported advantage depends on the assumption that real human rough estimates vary like uniform random samples within ±5% of the true count; if human estimates are biased, heavy-tailed, or image-dependent, the training signal would differ and the observed gains over semi-supervised methods could shrink.","fun_headline_variants_meta":{"raw":{"variants":["Rough crowd totals beat semi-supervised counting","CLIP counts crowds digit-by-digit from rough labels","Progressive digit estimation for rough crowd counting","Vision-language adapter improves rough counting","Counting crowds without per-target annotations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1193,"prompt_tokens":849,"completion_tokens":344,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":465,"completion_tokens_details":{"reasoning_tokens":279}},"tokens_in":465,"tokens_out":344,"duration_ms":3641,"temperature":1.0,"reasoning_tokens":279,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:08:45.277158+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect rough count labels from human annotators on SHA, QNRF, and JHU++ without revealing true counts, train ProgRoCC on these real rough labels, and compare its MAE/MSE against the results obtained with synthetic ±5% labels. If performance degrades toward or below the semi-supervised baselines (for instance, SSL-FT), the equivalence between simulated and real rough labels—and with it the central claim—is not supported.","supporting_citations":[{"cited_title":"Single-image crowd counting via multi-column convolutional neural network","cited_arxiv_id":null,"evidence_quote":"Supplies the SHA dataset and the MCNN baseline used for comparison."},{"cited_title":"Composition loss for counting, density map estima- tion and localization in dense crowds","cited_arxiv_id":null,"evidence_quote":"Supplies the dense QNRF dataset that stresses counting in highly congested scenes."},{"cited_title":"Jhu-crowd++: Large-scale crowd counting dataset and a benchmark method","cited_arxiv_id":null,"evidence_quote":"Supplies the JHU++ dataset covering varied weather conditions and large-scale evaluation."},{"cited_title":"Self-supervised learning with data-efficient su- pervised fine-tuning for crowd counting","cited_arxiv_id":null,"evidence_quote":"SSL-FT is the semi-supervised baseline ProgRoCC claims to outperform by 14.7% MAE on SHA and 24.0% MAE on QNRF."},{"cited_title":"Csrnet: Di- lated convolutional neural networks for understanding the highly congested scenes","cited_arxiv_id":null,"evidence_quote":"CSRNet is the fully supervised baseline whose SHA performance ProgRoCC approaches, anchoring the comparison."},{"cited_title":"Crowdclip: Unsupervised crowd counting via vision-language model","cited_arxiv_id":null,"evidence_quote":"CrowdCLIP is a prior CLIP-based unsupervised counting method that ProgRoCC extends and outperforms."}],"review_version":1}