{"id":"e7dae6e7-9d95-45a4-8292-81e5696dde1f","arxiv_id":"2606.05708","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":2.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"YOLOv8 trained on 7,959 images (public guns/knives plus custom blunt objects) for real-time detection of three threat classes in Indian surveillance scenarios.","lead":"The paper describes training a YOLOv8 model on a combined dataset of guns, knives, and 336 custom blunt-object images to detect threats in surveillance video. A smart generalist might read it to understand how existing computer vision tools are being adapted for real-time public safety monitoring in specific regional contexts.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Small custom blunt-object dataset (336 mobile-camera images) and resulting class imbalance/domain shift prevent support for real-world surveillance generalization claim.","rationale":"Reader's weakest_assumption directly identifies the same dataset representativeness issue as the load-bearing point. The applied nature of the work (standard YOLOv8, no novel method) makes data quality the decisive factor for the deployment claim; the proposed test would falsify or support it concretely. No other internal inconsistency (e.g., math error) exists to elevate.","tokens_in":1735,"tokens_out":364,"duration_ms":15585,"concrete_test":"Collect or obtain 200+ held-out images from actual fixed Indian surveillance cameras (different locations/times from the mobile set); run the final model and compute per-class mAP drop for blunt objects relative to the paper's reported validation numbers. A drop >15% confirms the domain-shift concern lands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that a YOLOv8 model trained on the merged 7,959-image set produces accuracy/efficiency suitable for deployment on Indian CCTV footage. The load-bearing assumption is that the 336-image custom set (iron/wooden/plastic rods) is representative and sufficient. This fails because: (1) severe imbalance (~4% blunt vs. ~48% each gun/knife), (2) collection via mobile camera introduces domain shift in viewpoint, resolution, lighting, and motion blur relative to fixed surveillance cameras, and (3) the abstract only reports metric gains on (presumably in-distribution) validation without target-domain testing or imbalance mitigation details. No evidence is given that longer training avoids overfitting to the tiny blunt class rather than learning spurious mobile-camera cues.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes a YOLOv8-based real-time object detection framework for identifying guns, knives, and blunt objects (iron/wooden/plastic rods) in surveillance videos, targeting Indian scenarios. It describes constructing a custom dataset of 336 mobile-camera images of blunt objects, merging it with a public 7,623-image gun/knife set to yield 7,959 total images, training YOLOv8 on the combined data, and asserts that longer training improves recall and average precision for the blunt class without overfitting. The central claim is that the resulting framework achieves an effective accuracy-efficiency balance suitable for deployment in real-world settings such as campuses, public spaces, and transportation areas.","tokens_in":1881,"tokens_out":587,"duration_ms":37576,"significance":"If the empirical claims were supported by quantitative metrics, validation protocols, and target-domain testing, the work would offer modest practical value by extending standard weapon-detection datasets with region-specific blunt objects. The dataset-merging strategy is a simple, reproducible step that could be useful for practitioners. However, the manuscript provides no such evidence, so its significance remains limited to an initial exploration rather than a validated contribution.","major_comments":[{"comment":"Abstract: The claim that 'increasing the training duration significantly improves recall and average precision for the blunt object class without signs of overfitting' is unsupported by any reported metrics, training curves, validation splits, or error analysis. This directly undermines the central performance and deployment claims.","section":"Abstract"},{"comment":"Dataset construction (Abstract): The 336-image custom blunt-object set constitutes only ~4% of the merged 7,959-image collection, producing severe class imbalance (~48% gun, ~48% knife). Collection via mobile camera further introduces unaddressed domain shift in viewpoint, resolution, lighting, and motion blur relative to fixed CCTV; no target-domain evaluation on Indian surveillance footage is described, rendering the generalization claim to 'real-world surveillance environments' unsupported.","section":"Abstract"},{"comment":"Experimental evaluation (Abstract): No baseline comparisons, ablation studies on training duration, or quantitative results (e.g., per-class mAP, precision-recall values) are supplied to substantiate the accuracy-efficiency balance or the absence of overfitting to mobile-camera cues.","section":"Abstract"}],"minor_comments":[{"comment":"The abstract refers to 'region-specific blunt objects commonly involved in violent activities in Indian surveillance scenarios' without citing supporting statistics or references.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as a preliminary report lacking any numerical results or experimental details; it may not align with the standards of a serious journal in computer vision without major additions to the experiments and evaluation sections."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed and constructive comments. We agree that several claims in the abstract require supporting quantitative evidence and will revise the manuscript accordingly to strengthen the presentation of results, limitations, and evaluation protocols.","responses":[{"response":"We acknowledge that the abstract statement on training duration effects is not backed by reported metrics or curves in the current version. In the revision we will add per-epoch validation curves, per-class precision-recall values, and an explicit error analysis section demonstrating the observed improvements and lack of overfitting indicators (e.g., validation loss plateau).","revision_made":"yes","referee_comment":"[Abstract] Abstract: The claim that 'increasing the training duration significantly improves recall and average precision for the blunt object class without signs of overfitting' is unsupported by any reported metrics, training curves, validation splits, or error analysis. This directly undermines the central performance and deployment claims."},{"response":"The class distribution is indeed imbalanced; we will report the exact counts and add a discussion of class-weighting or oversampling strategies used during training. On domain shift, we agree the mobile-camera collection differs from fixed CCTV and will revise the generalization language to note this limitation explicitly while clarifying that the public gun/knife images already span varied sources. No target-domain Indian CCTV test set was collected, so we will remove or qualify the deployment claim and list target-domain evaluation as future work.","revision_made":"partial","referee_comment":"[Abstract] Dataset construction (Abstract): The 336-image custom blunt-object set constitutes only ~4% of the merged 7,959-image collection, producing severe class imbalance (~48% gun, ~48% knife). Collection via mobile camera further introduces unaddressed domain shift in viewpoint, resolution, lighting, and motion blur relative to fixed CCTV; no target-domain evaluation on Indian surveillance footage is described, rendering the generalization claim to 'real-world surveillance environments' unsupported."},{"response":"We will expand the experimental section with (i) quantitative tables reporting per-class mAP@0.5, precision, recall, and inference speed on the validation split, (ii) an ablation table comparing different training durations, and (iii) baseline comparisons against YOLOv5 and Faster R-CNN under identical data splits. These additions will directly support the accuracy-efficiency claims.","revision_made":"yes","referee_comment":"[Abstract] Experimental evaluation (Abstract): No baseline comparisons, ablation studies on training duration, or quantitative results (e.g., per-class mAP, precision-recall values) are supplied to substantiate the accuracy-efficiency balance or the absence of overfitting to mobile-camera cues."}],"tokens_in":1493,"tokens_out":578,"duration_ms":25099,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main takeaway is that this is a straightforward application of an existing detector to a new but small custom dataset, with the abstract asserting performance gains and real-world suitability without any numbers or tests to back them up.\n\nThe new element is the 336-image collection of iron, wooden, and plastic rods gathered on a mobile camera and merged with a public gun/knife set to reach 7,959 total images. The authors note that longer training raised recall and average precision on the blunt class without apparent overfitting. That is a practical step for covering region-specific threats in an Indian context.\n\nThe work is honest about focusing on local surveillance needs rather than claiming a new algorithm. It correctly identifies that manual monitoring has limits and that blunt objects matter alongside firearms.\n\nThe soft spots are substantial and central. The abstract claims an effective accuracy-efficiency balance for campuses and public spaces but reports no precision, recall, mAP values, baseline comparisons, or validation protocol. The blunt class is only about 4% of the data, creating severe imbalance with no mention of mitigation. Mobile-camera collection also introduces clear domain shift in viewpoint, resolution, and blur compared with fixed CCTV. No target-domain testing on actual surveillance footage is described, so the generalization claim rests on in-distribution results alone.\n\nThis paper would interest someone assembling a quick local prototype but offers little for readers working on detection methods or scalable systems. I would not bring it to a reading group or cite it. It lacks the evidence needed for serious peer review.","headline":"This applies YOLOv8 to a merged gun/knife dataset plus a 336-image mobile-camera blunt-object set but gives no metrics to support the deployment claims.","tokens_in":2352,"tokens_out":385,"would_cite":false,"duration_ms":22075,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A YOLOv8 model trained on guns, knives and a custom set of 336 blunt-object images supports real-time threat detection in surveillance video.","keywords":["object detection","YOLOv8","surveillance","threat detection","gun detection","knife detection","blunt objects","real-time"],"falsifier":"A separate test set of real Indian surveillance videos on which the model shows a clear drop in accuracy or frame rate below usable levels would falsify the suitability claim.","tokens_in":2634,"feed_emoji":"📹","tokens_out":681,"duration_ms":44601,"temperature":0.7,"pith_summary":"The paper builds an automated system to spot guns, knives and blunt objects such as rods or sticks inside live camera feeds. It gathers 336 new labeled images of blunt objects with a mobile camera and adds them to an existing public collection of 7,623 gun and knife pictures. The combined set of 7,959 images trains a YOLOv8 detector whose recall and precision on the new class keep rising with longer training runs and show no overfitting. This setup targets the limits of human watchers who tire and miss events across many cameras. The authors position the result as ready for use in places like campuses, streets and transport hubs because it runs fast enough for continuous video while maintaining usable detection rates.","feed_headline":"YOLOv8 detects guns, knives and blunt objects in live video","feed_subtitle":"Custom 336-image blunt set merged with public weapons data trains efficient model for campus and street surveillance.","key_machinery":"YOLOv8 object detection model trained on the consolidated 7,959-image dataset covering guns, knives and blunt objects.","core_discovery":"The paper claims that training a YOLOv8 object detector on the merged dataset of public gun and knife images plus the new 336-image blunt-object collection produces a system that balances accuracy and speed for real-time surveillance. Extended training improves metrics on the blunt-object class without signs of overfitting. The resulting framework is presented as deployable in real-world settings such as campuses, public spaces and transportation areas.","pith_inferences":["The same custom-data method could fill gaps for other locally common objects not covered by public weapon sets.","Running the detector on low-power edge hardware would test whether it fits existing camera installations without extra servers.","Connecting the output directly to alert systems could shorten the time between detection and human response.","The gains on the blunt-object class show that public datasets alone leave out important regional variations in threat appearance."],"forward_implications":["Longer training duration raises recall and average precision for the blunt object class.","No overfitting appears even after extended training on the merged dataset.","The trained model supports real-time performance needed for live surveillance feeds.","Detection covers the three threat classes relevant to violent activity in the target environments."],"fun_headline_variants":["YOLOv8 on merged blunt gun and knife image dataset","Extended training improves YOLOv8 blunt object recall","7959 images train YOLOv8 for real time surveillance","YOLOv8 balances accuracy and efficiency in surveillance"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The 336 custom blunt-object images, once merged with the public gun and knife set, form a training collection that matches the variety and camera conditions of actual Indian surveillance footage closely enough for the model to generalize.","fun_headline_variants_meta":{"raw":{"variants":["YOLOv8 on merged blunt gun and knife image dataset","Extended training improves YOLOv8 blunt object recall","7959 images train YOLOv8 for real time surveillance","YOLOv8 balances accuracy and efficiency in surveillance"]},"model":"grok-4.3","cost_usd":0.008225,"raw_usage":{"total_tokens":3738,"prompt_tokens":681,"num_sources_used":0,"completion_tokens":65,"cost_in_usd_ticks":82249500,"prompt_tokens_details":{"text_tokens":681,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2992,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":681,"tokens_out":65,"duration_ms":25432,"temperature":1.0,"reasoning_tokens":2992,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-28T01:59:45.454698+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A separate test set of real Indian surveillance videos on which the model shows a clear drop in accuracy or frame rate below usable levels would falsify the suitability claim.","supporting_citations":[],"review_version":1}