{"id":"40505077-ee4c-48e1-9109-7f66c06dfeeb","arxiv_id":"2505.08834","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper describes a 5-column M-CNN with rotation self-supervision and Sinkhorn distribution matching for crowd counting and a VGG19-LSTM with dense residual blocks for violence detection, but the evidence does not support the claimed state-of-the-art results.","lead":"A master's thesis proposes two deep learning pipelines, one for crowd counting and one for violence detection, but its central claim of annotation-free counting is contradicted by its own training details, and the reported results are sample images without benchmark totals. It is worth a quick read to see common pitfalls in self-supervised method claims.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zero-annotation counting claim is contradicted by the paper's own supervised pre-training and ground-truth density-map loss in §4.2.","rationale":"I agree with the reader's REJECT verdict. The reader's weakest_assumption focuses on the undefined Sinkhorn prior in §3.2.4; I see that as a genuine reproducibility issue, but the more load-bearing defect is the direct contradiction between the zero-annotation contribution and the training protocol in §4.1.2, §4.2.1, and §4.2.2(B), which explicitly uses annotated density maps for pre-training and distribution matching. Fixing the prior definition would not repair the central claim if the loss still targets human-annotated density maps. The anomaly-detection claim also lacks baselines and an SCVD description, and no reproducible code is supplied, but the counting contradiction alone is sufficient to reject the central contribution as stated. Hence the reader's verdict stands unchanged.","tokens_in":42229,"tokens_out":5618,"duration_ms":55844,"concrete_test":"Trace the supervision signal through §4.2.1–§4.2.2: if the stage-2 loss or pre-training objective references ground-truth density maps, as §4.2.2(B) states, the zero-annotation claim is invalid. For a positive test, retrain the full pipeline with no annotated density maps, using only rotation pretext and Sinkhorn matching to a defined unlabeled prior, and report aggregate MAE/MSE on ShanghaiTech Part A/B; if no such numbers or prior definition are supplied, the claim remains unverified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central crowd-counting claim is that no annotated image is used. For that to hold, every training loss must be computable without human labels. The paper's own training section says otherwise. §4.1.2 states that annotated and unannotated data are used and that annotated density maps are used for pre-training. §4.2.1 describes a supervised pre-training mechanism using images and their ground-truth maps. §4.2.2(B) says the distribution-matching objective is to minimize the difference between ground-truth density maps and predicted density maps. Any distance or Sinkhorn match to manually annotated density maps, in pre-training or fine-tuning, contradicts the zero-annotation contribution. The undefined prior distribution in §3.2.4 is a serious reproducibility gap, but it is secondary: specifying that prior would not make a pipeline that regresses to annotated density maps annotation-free. Moreover, no aggregate MAE/MSE on ShanghaiTech Part A/B or UCF-QNRF is reported, so the claimed quantitative evaluation is absent.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript, based on an M.Sc. thesis, proposes two systems for crowd scene analysis: a self-supervised multi-column CNN for crowd counting and a VGG19-LSTM model for crowd anomaly detection. The counting pipeline is described as a two-stage method in which stage one uses rotation classification as a pretext task and stage two uses Sinkhorn distribution matching for density estimation; the abstract and contribution list claim that this method is completely self-supervised and does not use a single annotated image. Results are reported on ShanghaiTech and UCF-QNRF, while the anomaly detection model is evaluated on the Hockey Fight and SCVD datasets with precision, recall, and F1-score metrics.","tokens_in":42383,"tokens_out":2985,"duration_ms":32349,"significance":"If the central claim were correct, a crowd counting method that requires no annotated images would be a substantial contribution because it would remove the dominant annotation bottleneck in the field. The paper also proposes a five-branch M-CNN and a VGG19-LSTM architecture with wide dense residual blocks, and it reports frame-level anomaly detection rather than clip-level prediction. However, the central zero-annotation claim is directly contradicted by the manuscript's own training description, the quantitative evaluation is missing the promised aggregate MAE/MSE numbers, and the anomaly detection comparisons with prior state-of-the-art methods are not presented. The paper does not ship code, trained models, or a derivation of the loss functions, so the strengths are mainly the clear problem motivation and the architectural description rather than verified empirical results.","major_comments":[{"comment":"The paper's central claim, stated in the abstract and in Section 1.9.1, is that crowd counting is performed by 'completely self-supervised learning without using a single annotated image.' This is contradicted by the manuscript's own training description. Section 4.1.2 states that 'annotated and unannotated data is used for training purposes' and that the annotated dataset, consisting of crowd photos with manually annotated density maps, is used for pre-training. Section 4.2.1 then describes a supervised pre-training step in which the model is fed images and their ground-truth maps. Section 4.2.2(B) states that the distribution matching objective is to minimize the difference between ground-truth density maps and predicted density maps. Since the pipeline regresses toward manually annotated density maps, either in pre-training or in distribution matching, the zero-annotation claim is not supported. This is a load-bearing point because the paper's main novelty rests on eliminating the annotation bottleneck.","section":"Sections 4.1.2, 4.2.1, and 4.2.2(B)"},{"comment":"Stage 2 is described as matching the predicted density distribution to 'a prior distribution that approximates the statistical characteristics of natural crowds' using Sinkhorn matching, but the prior is never defined. No equation specifies the prior's parametric form, how it is estimated, or what data is used to derive it. Likewise, the Sinkhorn objective is not written out, and no algorithm or implementation detail is given. Because the entire density estimation stage is built on matching to this prior, the unspecified prior makes the method non-reproducible and prevents the reader from verifying whether the claimed self-supervised density estimation is even well posed.","section":"Section 3.2.4"},{"comment":"Although Section 4.4 defines MAE and MSE, and the abstract claims evaluation on ShanghaiTech and UCF-QNRF, no aggregate MAE or MSE values are reported for any test set. The experimental section only presents per-image ground-truth versus predicted counts for a few selected images, and those numbers include substantial errors, e.g., Table 4.3 Image 3 reports 707 ground truth versus 431.64 predicted, and Table 4.5 reports 975 versus 577. Without aggregate errors or a comparison against existing crowd counting methods on the full test splits, the quantitative evaluation claims in the abstract and in Section 7 are unsupported.","section":"Section 4.5 and Tables 4.1-4.5"},{"comment":"The claim that the anomaly detection model outperforms other state-of-the-art approaches is not supported by any comparative table. Section 6.4 reports only precision (0.91), recall (0.82), and F1-score (0.86) on the Hockey Fight dataset without comparing to prior published results, and Section 6.5 reports one result on SCVD. Additionally, Table 6.2 is mislabeled: the section heading refers to SCVD, the table's first column says 'Hockey Fight Dataset,' and the row also says 'Hockey Fight Dataset,' making the reported SCVD result ambiguous. Without a comparison to existing methods and a clear reporting of the experimental setup, the 'outperform state-of-the-art' statement is not verifiable.","section":"Sections 6.4 and 6.5"}],"minor_comments":[{"comment":"The notation in the MAE and MSE formulas is inconsistent: y_c is defined as 'number of peoples counted' and the denominator symbols are garbled, while the predicted count is not clearly distinguished from the ground-truth count. Please define y_pred and y_GT explicitly and rewrite the equations with standard notation.","section":"Equations 1 and 2 (Section 4.4)"},{"comment":"Table 4.4 and Table 4.5 have the same caption, 'Ground Truth vs. Predicted Count for Images with Scale In-variation,' but Table 4.5 appears in the complex-background subsection. The captions should be corrected.","section":"Tables 4.4 and 4.5"},{"comment":"The row label in Table 6.2 says 'Hockey Fight Dataset' while the section heading is 'Experimental Results on SCVD.' This typo should be fixed so the reported 0.93/0.92/0.90 result is attributed to the correct dataset.","section":"Table 6.2"},{"comment":"The subsection numbering in Chapter 3 is inconsistent: '3.1 Theoretical Background' is followed by '3. Data Augmentation' instead of '3.1.1 Data Augmentation.' This should be corrected throughout the chapter.","section":"Section 3.1.1"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an unmodified M.Sc. thesis, including front matter such as declarations, dedications, and acknowledgements. Beyond the editorial fit, the central technical claim is contradicted by the authors' own description of supervised pre-training and ground-truth density-map loss, and the missing aggregate results prevent any meaningful evaluation. The paper is not suitable for publication in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the paper's headline contribution—crowd counting with completely self-supervised learning, no annotated images—is contradicted by its own methods section. The stress-test note is correct on this. It's not a nitpick; it's the center of the paper.\n\nWhat's good: the paper tackles a real problem (annotation cost in crowd counting) and the proposed pipeline (rotation self-supervision, multi-column CNN, Sinkhorn distribution matching) is a plausible idea worth exploring. The literature review is broad and the thesis is clearly written at an expository level. The anomaly detection model (VGG19 + LSTM + dense residual block) is a standard but sensible architecture.\n\nWhere it falls apart: §4.1.2 says annotated and unannotated data are used; §4.2.1 describes supervised pre-training on images with ground-truth density maps; §4.2.2(B) says distribution matching minimizes the difference between ground-truth and predicted density maps. So the training is supervised, full stop. The undefined prior distribution in §3.2.4 is a separate, serious reproducibility gap, but even a well-specified prior wouldn't make a pipeline that regresses to annotated density maps annotation-free.\n\nThe evaluation is also missing. No aggregate MAE/MSE is reported for ShanghaiTech or UCF-QNRF; the experimental results are a few sample images with predicted counts, some quite far off (e.g., 962 vs 799, 975 vs 577). The anomaly detection results have no baselines, and the SCVD dataset is not described or cited.\n\nThis is not ready for peer review. The honest path is to either drop the annotation-free claim, specify the prior, and report full benchmark numbers, or present the work as a thesis/technical report. As is, it deserves a desk reject, not referee time.","headline":"The paper's central zero-annotation counting claim is contradicted by its own supervised pre-training and ground-truth density loss; without aggregate results or a defined prior, it is not ready for review.","tokens_in":42941,"tokens_out":2396,"would_cite":false,"duration_ms":24334,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This thesis claims crowd counting works without any annotated images, using rotation classification plus Sinkhorn density matching in a five-column CNN, and that a VGG19-LSTM flags violent frames one at a time.","keywords":["Crowd scene analysis","Crowd counting","Crowd anomaly detection","Multi-column convolutional network","Self-supervised learning","Convolutional LSTM","Density map estimation","Violence detection"],"falsifier":"Reproduce stage 2 on the ShanghaiTech Part A test set: write down the prior distribution the Sinkhorn step must match to, train the density regressor against that target only, and compute aggregate MAE and MSE over the whole test set. If no concrete prior can be stated from the thesis, or if the predicted counts are no better than a constant-density baseline, the self-supervised counting claim is not settled.","tokens_in":41958,"feed_emoji":"👥","tokens_out":17504,"duration_ms":148921,"temperature":0.7,"pith_summary":"This thesis sets out to remove two bottlenecks in crowd analysis: the need for annotated head locations in crowd counting, and the clip-level granularity of crowd violence detection. For counting, it claims a completely self-supervised pipeline — a five-column multi-scale CNN learns crowd features by predicting the rotation angle of image crops, and a second stage freezes those features and trains a density regressor by matching its predicted density map, through Sinkhorn distance, to a prior distribution of natural crowd densities, with no annotated density maps involved; the model is claimed to generalize across the ShanghaiTech and UCF-QNRF datasets. If that claim holds, dense-scene counting becomes trainable from raw surveillance footage at essentially zero annotation cost. For anomaly detection, the thesis claims a spatio-temporal VGG-19-LSTM, with fully connected layers replaced by wide dense residual blocks, that classifies violence at the level of individual frames and outperforms state-of-the-art approaches on the Hockey Fight and SCVD datasets.","feed_headline":"Counting crowds with zero annotated images","feed_subtitle":"Rotation pretext plus Sinkhorn density matching aims to drop the annotation bottleneck in counting.","key_machinery":"The load-bearing object for counting is the Sinkhorn matching step in stage 2: the predicted density distribution is aligned, via optimal transport, to a 'prior distribution that approximates the statistical characteristics of natural crowds,' and this matching is what is supposed to make the density regressor trainable without annotation. Around it sit the five-column M-CNN — parallel convolutional branches with filters of sizes 3×3, 5×5, 7×7, 9×9, and 11×11 that capture local and global scale — and the rotation-classification head (0°, 90°, 180°, 270°) that supplies the stage-1 pseudo-labels. For anomaly detection the machinery is the modified VGG-19, whose fully connected layers are replaced by wide dense residual blocks, feeding an LSTM that captures motion between frames at times t and t−1, ending in a binary normal-versus-abnormal classification.","core_discovery":"The central discovery the thesis argues for is that the supervision signal for crowd counting does not have to come from annotated heads. Stage 1 trains the feature extractor on a rotation-classification pretext task — predicting whether a 112×112 crop was rotated by 0°, 90°, 180°, or 270° — which forces the network to learn crowd-relevant structure from raw images alone. Stage 2 freezes those features, appends two further convolutional blocks, and trains the density regressor by minimizing the Sinkhorn distance (an optimal-transport measure) between the predicted density map and a prior distribution assumed to capture the statistical character of natural crowds, so the prediction is pulled toward a statistical target rather than compared with ground-truth annotations. For anomaly detection, the thesis claims that replacing VGG-19's fully connected layers with wide dense residual blocks and passing the extracted spatial features through an LSTM yields a frame-level binary normal/abnormal classifier that reports F1 scores of 0.86 on the Hockey Fight dataset and 0.90 on the SCVD dataset and claims to outperform other state-of-the-art approaches.","pith_inferences":["The same two-stage recipe — rotation pretext followed by matching a prediction to a statistical prior — could transfer to other counting problems whose objects have predictable statistics (vehicles, animals, cells), because the prior plays the role that ground-truth density maps play in supervised training.","The frame-level anomaly formulation, if it holds up, points toward temporal action segmentation: frame scores from the VGG19-LSTM could be pooled to label intervals of a surveillance clip as normal or violent without clip-level annotations.","A careful reader should not assume the strongest reading of 'completely self-supervised': Sections 4.1.2 and 4.2.1 describe supervised pre-training on annotated density maps before the self-supervised stage, so the no-annotation claim is best understood as applying to the counting stage proper unless the author states otherwise.","The counting results chapter reports per-image predicted counts and rotation-accuracy curves but no aggregate MAE/MSE over the full ShanghaiTech or UCF-QNRF test sets, despite the abstract promising those metrics; an independent replication would need to supply the aggregate numbers to confirm the headline claim."],"forward_implications":["Crowd-counting models could be trained from raw, unannotated surveillance footage, eliminating the manual head-annotation step that currently restricts deployment to new scenes.","The five-column design asserts that widening a counting network with more parallel scale-specific filters improves accuracy on occluded scenes, non-uniform density, and scale variation without making the network deeper.","A density regressor trained purely by distribution matching would transfer across datasets such as ShanghaiTech and UCF-QNRF without re-annotating for each new scene.","Frame-level violence classification lets an operator act on short-duration incidents that clip-level or video-level detectors would smooth away."],"supporting_citations":[{"why":"Supplies the multi-column CNN architecture (three parallel branches) that this work extends to five columns for scale-aware crowd counting.","marker":"[13]"},{"why":"Rotation-angle prediction as a self-supervision pretext task; the idea that stage 1 of the counting pipeline is built on.","marker":"[56]"},{"why":"Companion reference for rotation prediction as a self-supervision pretext task for learning visual features.","marker":"[57]"},{"why":"Convolutional LSTM for violence detection; the lineage the anomaly model extends with VGG-19 spatial features.","marker":"[14]"},{"why":"Bidirectional convolutional LSTM violence detector; the state-of-the-art approach this work claims to outperform.","marker":"[15]"},{"why":"VGG-19, the architecture whose convolutional features (with pre-trained weights) carry the spatial side of the anomaly detection model.","marker":"[120]"},{"why":"Wide dense residual networks; the block design that replaces VGG-19's fully connected layers in the proposed anomaly detector.","marker":"[125]"}],"fun_headline_variants":["Crowd counting without a single labeled head","Rotation pretext unlocks crowd counting","Sinkhorn distance skips density map labels","Dense residual LSTM improves anomaly detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire counting pipeline rests on the 'prior distribution' of natural crowd densities that stage 2's Sinkhorn matching targets, but the thesis never defines that distribution's form, its parameters, or how it is estimated (Section 3.2.4); if that target is wrong or arbitrary, the density estimation has nothing principled to match to.","fun_headline_variants_meta":{"raw":{"variants":["Crowd counting without a single labeled head","Rotation pretext unlocks crowd counting","Sinkhorn distance skips density map labels","Dense residual LSTM improves anomaly detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000815,"raw_usage":{"total_tokens":3605,"prompt_tokens":1015,"completion_tokens":2590,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":2536}},"tokens_in":631,"tokens_out":2590,"duration_ms":17472,"temperature":1.0,"reasoning_tokens":2536,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:59:49.289278+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce stage 2 on the ShanghaiTech Part A test set: write down the prior distribution the Sinkhorn step must match to, train the density regressor against that target only, and compute aggregate MAE and MSE over the whole test set. If no concrete prior can be stated from the thesis, or if the predicted counts are no better than a constant-density baseline, the self-supervised counting claim is not settled.","supporting_citations":[],"review_version":1}