{"id":"0c349675-8496-4af6-a51e-4520fcc5d1f2","arxiv_id":"2412.05331","paper_version":3,"verdict":"REJECT","confidence":"LOW","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A thesis-style report combines standard object detection and tracking tools into a surveillance pipeline and reports precision and recall numbers without revealing the dataset or evaluation protocol.","lead":"This preprint describes a proposed surveillance system that combines known computer vision tools, such as YOLO and Gaussian Mixture Models, to detect activity and record only interesting video clips. It reports detection and tracking metrics, but does not provide the dataset, code, or evaluation details needed to verify them.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed precision, recall, F1, and processing-time results in Chapter 4 are asserted without any dataset, ground truth, detection threshold, or evaluation protocol, leaving the central performance claim unverifiable and unfalsifiable.","rationale":"The reader's weakest assumption is exactly the load-bearing concern here: the reported performance numbers in Chapter 4 are detached from any experimental protocol. I read the paper in good faith as a project report that assembles standard components (GMM, YOLO/SSD/Faster R-CNN, DeepSORT, Kalman/Siamese, RNN/LSTM) and claims high performance. For that central claim to be credible, one would need to know what data the numbers came from and how they were computed. The only evaluation-related content is a set of figures and a table with no supporting methodology: no dataset name, no ground-truth definition, no detection threshold, no split, no occlusion-generation procedure, no hardware/software configuration for the timing results, and no baseline comparison. This is not a case of 'outside current consensus' or a theoretical disagreement; it is a missing-evidence problem. The paper also contains no code, no reproducibility artifacts, and no formal verification, so there is no independent support to compensate for the absent protocol. A REJECT verdict is appropriate because, as written, the paper provides no verifiable evidence for its stated contribution. I am not alleging fabrication or misconduct; I am noting that the claims are unfalsifiable without the missing evaluation artifacts. If the authors supplied the dataset and evaluation script and the numbers reproduced, the manuscript could become a legitimate empirical report. Until then, the correct verdict is rejection.","tokens_in":5306,"tokens_out":2024,"duration_ms":23156,"concrete_test":"Ask the authors to release the exact dataset, ground-truth annotations, evaluation script, and model configuration used for Chapter 4. If these are unavailable, the performance claim is unsupported. If they are provided, rerun the detection and tracking pipelines on a fixed validation split using a stated IoU threshold and confidence threshold, and compare the resulting precision, recall, and F1 to the reported values; a material discrepancy would confirm that the original numbers were not obtained under a reproducible protocol.","verdict_should_be":"REJECT","load_bearing_attack":"For the central claim to hold, the system must actually have been evaluated under a defined, reproducible protocol. Chapter 4 presents Figure 5, Table 2, and Figure 7 as evidence: YOLO precision 0.90 and recall 0.85, SSD 0.88/0.83, Faster R-CNN 0.92/0.87, tracking F1 0.85 under occlusion, and reduced processing time after optimization. However, the paper never states which video dataset was used, how ground truth was obtained, how detections were matched to ground truth (e.g., IoU threshold), what confidence threshold was applied, what the train/test or video split was, or how the 'with occlusions' condition was generated. Without these details, the reported numbers are not empirical measurements but assertions. The same problem applies to Figure 6 (GMM comparison) and Figure 7 (processing time): no axis labels, no units, no dataset, and no configuration are provided. The surrounding text only describes a pipeline of standard tools (YOLO, DeepSORT, GMM, Kalman filters) without specifying any novel components or implementation choices that could be checked. Because the evaluation protocol is entirely absent, even a faithful reimplementation of the described architecture would not tell us whether the reported numbers are correct; the evidence needed to evaluate the central claim is simply missing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a video surveillance system, called the Integrated Detection and Tracking System (IDTS), that segments continuous video into activity-only clips using Gaussian Mixture Model background subtraction, CNN detectors (YOLO, SSD, Faster R-CNN), Kalman/Siamese tracking, and RNN/LSTM action recognition. The paper reports high precision and recall values for the three detectors, tracking F1-scores with and without occlusion, and processing-time improvements after optimization, along with high-level architecture diagrams and a short literature survey. The writing is organized as a project report rather than a research paper, with the experimental chapter consisting almost entirely of figures and tables without supporting protocol details.","tokens_in":5615,"tokens_out":3180,"duration_ms":28906,"significance":"If the reported performance were substantiated, the system would address a practical surveillance need: reducing storage by recording only activity-rich segments and enabling searchable retrieval. The pipeline is composed of standard, mature components, so the potential contribution lies in the integration and evaluation rather than in novel algorithmic ideas. The paper provides no code, no dataset, no ground-truth definition, and no reproducible evaluation protocol, so the central performance claims are currently unsupported. The practical motivation is valid, but the manuscript as written does not permit an independent assessment of whether the system works as claimed.","major_comments":[{"comment":"The central performance claims (YOLO Precision 0.90, Recall 0.85; SSD Precision 0.88, Recall 0.83; Faster R-CNN Precision 0.92, Recall 0.87; tracking F1 0.85 with occlusions) are presented without any evaluation protocol. The paper never names a dataset, defines ground-truth annotations, specifies detection confidence thresholds or IoU matching criteria, or describes train/test or video splits. Without these details, the numbers are assertions rather than empirical measurements, and the central claim of the paper is unfalsifiable.","section":"Chapter 4, Figure 5 and Table 2"},{"comment":"The claims about GMM handling illumination changes and about reduced processing time after optimization are accompanied by figures lacking axis labels, units, dataset descriptions, and experimental configurations. It is impossible to determine what is being compared, under what conditions, or whether the reported improvement is meaningful. These figures therefore do not support the conclusions drawn from them.","section":"Chapter 4, Figures 6 and 7"},{"comment":"No implementation details are provided for the detection, tracking, or action-recognition modules: there are no model versions, hyperparameters, training procedures, feature extraction specifics, or code availability statements. Because the entire system depends on these components, the manuscript is not reproducible even in principle, and the reported results cannot be traced to a concrete implementation.","section":"Section 3.3 and Chapter 4"},{"comment":"The conclusion claims the system uses NVIDIA GPUs and PostgreSQL for storage, but these components do not appear in the methodology or in the experimental setup described in Chapter 4. This inconsistency leaves the reader unable to determine which hardware and storage configuration produced the reported results, further undermining the reproducibility of the performance claims.","section":"Sections 3 and 5.1"}],"minor_comments":[{"comment":"The chapter ends mid-sentence with the fragment 'using IDTS meth'; the sentence should be completed.","section":"Section 1.2"},{"comment":"The sentence 'Detection is improved on various-sized objects for multiple scenarios by multi-scale and contextual analysis' appears twice in the abstract; one copy should be removed.","section":"Abstract"},{"comment":"The heading '5.1 Summary' appears before the 'Chapter 5: Conclusions' heading; the section numbering and ordering should be corrected.","section":"Chapter 5 layout"},{"comment":"The table is titled 'Hardware and/or Software Components' but lists only software; either add the hardware components or rename the table.","section":"Table 1"},{"comment":"References [6] and [8] are self-citations to IoT intrusion detection papers that are unrelated to video surveillance and do not support the claims in the text; they should be removed or replaced with relevant citations.","section":"References [6] and [8]"},{"comment":"The claim that frame differencing 'has an accuracy of less than 40%' is attributed to reference [3], which is a human activity recognition paper and does not appear to contain this result; the citation should be verified and corrected.","section":"Section 2.1"},{"comment":"Figure 1 is labeled 'System Requirements' but is not described in the text, and its diagram content is not self-explanatory; a brief explanation should be added.","section":"Figure 1"}],"recommendation":"reject","confidential_remarks":"This manuscript is not yet at the standard of a research publication: the central experimental claims are unsupported by any dataset, protocol, or implementation details, and several citations do not support the statements they are attached to. Because the missing evidence is not a matter of local editing but requires the author to actually perform and document a proper evaluation, I recommend rejection rather than major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a well-structured project report, not a research paper. The central performance claims in Chapter 4 (YOLO precision 0.90, etc.) are asserted with no dataset, ground truth, thresholds, or evaluation protocol. I agree with the reader's rejection, but I want to give the author credit for laying out a coherent end-to-end surveillance pipeline.\n\nWhat is actually new: nothing, algorithmically. The system combines standard tools—GMM background subtraction, YOLO/SSD/Faster R-CNN detection, Kalman/Siamese tracking, LSTM action recognition—into a video-segmentation-and-search workflow. That integration is not itself a research contribution, and the paper does not claim to introduce a new architecture. It does, however, describe a sensible systems design, and the storage/search motivation (record only activity-rich segments) is practical.\n\nWhere the soft spots are: the experimental chapter is the core problem. Figure 5 and Table 2 present precision/recall/F1 as if measured, but no dataset is named, no ground truth is described, no matching threshold (e.g., IoU) is given, and the occlusion condition is not specified. Figure 6 and 7 lack axes and units. This is not a nitpick: without a protocol, the numbers cannot be checked or reproduced, so the central claim is unverified. Additionally, references [6] and [8] are self-citations to unrelated IoT security papers; they don't support the text where they appear. The limitations section is candid about computational cost and false positives, which shows the author understands the system's weaknesses.\n\nOne more thing: the paper reads like an undergraduate project report, and that's fine for what it is. But as a research submission it doesn't provide enough evidence for its performance claims.\n\nRecommendation: desk reject. There is no reproducible result here. If the author later provides the dataset, code, and evaluation protocol, the pipeline could be a useful engineering case study. As it stands, sending it to reviewers would waste their time. I would not cite it, and I wouldn't bring it to reading group.","headline":"Coherent project report with unsupported performance numbers; desk reject.","tokens_in":6037,"tokens_out":1878,"would_cite":false,"duration_ms":17757,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An Integrated Detection and Tracking System can turn surveillance video into searchable, event-only clips while sustaining high detection accuracy.","keywords":["Dynamic Scene Analysis","Object Detection","Activity Recognition","Convolutional Neural Networks","YOLO","Real-Time Processing","Video Surveillance","Motion Tracking"],"falsifier":"Apply the described GMM-plus-CNN-plus-Kalman/Siamese pipeline to a public surveillance video with labeled objects and compare detection precision, recall, and tracking F1 to the paper's reported numbers; because the paper does not name its dataset or thresholds, this externally defined run is the settling test.","tokens_in":5145,"feed_emoji":"🎥","tokens_out":6054,"duration_ms":50033,"temperature":0.7,"pith_summary":"This paper tries to establish that an Integrated Detection and Tracking System can make surveillance video intelligent by recording only the segments where activity occurs. It combines adaptive background subtraction, convolutional object detectors, and motion trackers so that storage is reduced and events become searchable. The reported measurements—precision and recall near 0.9 for YOLO, SSD, and Faster R-CNN, and tracking F1 of 0.85 under occlusion—are meant to show the hybrid pipeline works under realistic conditions. If true, the system would shift surveillance from passive recording to active, queryable event capture.","feed_headline":"Surveillance system records only activity-rich video moments","feed_subtitle":"Detector-tracker pipeline reports high accuracy while storing only event segments, easing storage and search.","key_machinery":"The load-bearing object is the Integrated Detection and Tracking System, a staged pipeline that starts with Gaussian Mixture Model background subtraction to separate changed pixels from the static scene. Those changed regions are passed to CNN detectors such as YOLO, SSD, and Faster R-CNN for localization and classification; Kalman filters predict motion and Siamese networks match appearances so identities persist through occlusions; and RNN/LSTM layers capture temporal patterns for action analysis. The machinery decides which video frames are worth keeping, which objects are in them, and whether two detections across frames are the same object.","core_discovery":"The central claim is that a single hybrid pipeline, called the Integrated Detection and Tracking System, can detect activity in a video stream, segment only the active parts, and track objects across frames well enough to keep identities during occlusions. In the paper's own terms, GMM-based background subtraction decides when to record, CNN detectors identify what is present, Kalman filters and Siamese networks keep track of who is where, and RNN/LSTM networks read what is happening. The headline numbers are detector precision and recall (YOLO 0.90/0.85, SSD 0.88/0.83, Faster R-CNN 0.92/0.87) and tracking F1 falling from 0.92 without occlusion to 0.85 with occlusion. The paper presents these as evidence that the system is accurate enough for real surveillance use while recording far less video.","pith_inferences":["Because the segmentation module runs before detection, any activity missed at the GMM stage would never reach the detectors, so end-to-end recall is capped by segmentation recall; a testable extension is to report clip-level recall jointly.","The reported numbers are component-wise rather than end-to-end, so the storage and search benefits would follow only if the full pipeline is evaluated on complete videos; that evaluation is a natural next step the paper does not provide.","The described RNN/LSTM action module implies the same infrastructure could index clips by action type, allowing queries such as 'person running' or 'person loitering'; the paper describes this capability but does not demonstrate it."],"forward_implications":["Surveillance recordings would shrink to activity-rich clips, lowering storage cost and reducing review time.","Operators could retrieve specific events by timestamp, detected object, or recognized action instead of watching entire footage.","Tracking through occlusion at F1 0.85 would let the system keep object identities when targets are briefly hidden, which matters in crowded scenes.","The claimed real-time optimizations would support near-instant response in security monitoring and similar time-sensitive settings."],"supporting_citations":[{"why":"Supplies the hybrid deep model combining dense optical flow with deep learning for human activity recognition, which the paper adapts for its action analysis module.","marker":"[3]"},{"why":"Supplies the CNN-plus-LSTM structure for detecting behavior from video, grounding the paper's RNN/LSTM action-recognition design.","marker":"[4]"},{"why":"Supplies adaptive tracking used to classify and monitor activities, informing the tracking component's role in activity segmentation.","marker":"[5]"},{"why":"Supplies motion segmentation and qualitative dynamic scene analysis from image sequences, the basis for the paper's scene-change segmentation approach.","marker":"[7]"},{"why":"Supplies statistical regularization for detecting and tracking moving objects in space and time, underlying the GMM and tracking choices.","marker":"[9]"}],"fun_headline_variants":["Hybrid system logs only key moments in surveillance video","Deep learning surveillance that records only active moments","Event-triggered video recording reduces CCTV storage needs","Hybrid model identifies activity to save storage in surveillance","Object detection and tracking with smart clip segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracy numbers are meaningful only if they were measured on a defined, representative dataset with a consistent protocol, but the paper does not specify the dataset, ground truth, or thresholds.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid system logs only key moments in surveillance video","Deep learning surveillance that records only active moments","Event-triggered video recording reduces CCTV storage needs","Hybrid model identifies activity to save storage in surveillance","Object detection and tracking with smart clip segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00045,"raw_usage":{"total_tokens":2289,"prompt_tokens":989,"completion_tokens":1300,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":1227}},"tokens_in":605,"tokens_out":1300,"duration_ms":9610,"temperature":1.0,"reasoning_tokens":1227,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:54:18.332209+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the described GMM-plus-CNN-plus-Kalman/Siamese pipeline to a public surveillance video with labeled objects and compare detection precision, recall, and tracking F1 to the paper's reported numbers; because the paper does not name its dataset or thresholds, this externally defined run is the settling test.","supporting_citations":[{"cited_title":"E., Zhong, B., & Ouyang, X","cited_arxiv_id":null,"evidence_quote":"Supplies the CNN-plus-LSTM structure for detecting behavior from video, grounding the paper's RNN/LSTM action-recognition design."},{"cited_title":"P., Brandt, A., & Gerl, R","cited_arxiv_id":null,"evidence_quote":"Supplies adaptive tracking used to classify and monitor activities, informing the tracking component's role in activity segmentation."},{"cited_title":"Optimized IoT Intrusion Detection using Machine Learning Technique","cited_arxiv_id":"2412.02845","evidence_quote":"Supplies motion segmentation and qualitative dynamic scene analysis from image sequences, the basis for the paper's scene-change segmentation approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies statistical regularization for detecting and tracking moving objects in space and time, underlying the GMM and tracking choices."}],"review_version":1}