{"id":"f25d8d45-e06d-4348-be3d-6c290c8b035b","arxiv_id":"2412.05252","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":1.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"This is a review of classical and CNN-based object detection algorithms with no new experimental findings.","lead":"This paper surveys object detection methods, from classical feature-based techniques to convolutional neural network detectors. It serves as an introductory compendium of known methods, but contains no new experiments or results.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table III's unsourced performance numbers and the Faster R-CNN/YOLOv8 factual errors undermine the review's central claim to be a reliable comprehensive reference.","rationale":"The reader's weakest assumption identifies the same issue: Table III's performance numbers and the technical descriptions must be accurate for the review to serve as a reference. I examined the full text and confirmed that Table III has no source attribution and contains entries that are not standard published results, such as R-CNN COCO mAP 54% and Open Images mAP 55%. The prose errors in Section IV.D (Faster R-CNN attributed to Girshick et al. rather than Ren et al.) and Section IV.F (YOLOv8 described as transformer-based) reinforce that the compilation is unreliable. Because the paper's contribution is explicitly a comprehensive review, these errors are load-bearing rather than cosmetic. I recommend no change to the reader's REJECT verdict; the concern strengthens it but does not redirect it.","tokens_in":8783,"tokens_out":4804,"duration_ms":45844,"concrete_test":"Trace every cell of Table III to a published evaluation with matching dataset, split, and mAP/IoU protocol. Concretely, check whether the original R-CNN and Fast R-CNN papers report COCO or Open Images mAP and whether any cited source reports ImageNet \"mAP\" for these detectors. If the majority of Table III cells cannot be sourced from the cited literature, the quantitative comparison is unsupported and the REJECT verdict stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is to provide a comprehensive, reliable overview, so the accuracy of its compiled facts is load-bearing. Table III reports mAP for R-CNN, Fast R-CNN, Faster R-CNN, Mask R-CNN, YOLO, and SSD on Pascal VOC, COCO, ImageNet, and Open Images with no citation, no dataset split, no IoU threshold, and no metric definition. Several entries do not correspond to published evaluations: R-CNN and Fast R-CNN were not evaluated on COCO or Open Images in their original papers, and ImageNet detection results are not reported as mAP in the cited references. Section IV.D misattributes Faster R-CNN to Girshick et al., citing [18] (the Fast R-CNN paper), whereas the correct attribution is Ren et al. [19]. Section IV.F states that YOLOv8 uses a transformer-based backbone; the cited Ultralytics repository and the YOLOv8 architecture use a convolutional CSPDarknet-style backbone. These are factual errors, not interpretive disagreements, and they occur in exactly the parts of a survey that give it its value.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a survey of object detection, organized into classical computer vision techniques (SIFT, HOG, Viola-Jones, DPM), generic region proposal methods, and convolution-based detectors (R-CNN, SPP-Net, Fast R-CNN, Faster R-CNN, Mask R-CNN, YOLO, SSD). It also covers applications, popular datasets, and evaluation metrics, and it offers comparative tables and a list of future research directions. The authors state as their key contribution a comprehensive overview that spans both classical and CNN-based approaches.","tokens_in":8951,"tokens_out":7667,"duration_ms":73061,"significance":"If the paper were factually accurate, it would provide a compact introductory survey with a clear taxonomy and useful summary tables. It contains no original experiments or derivations, so its entire value rests on the correctness of the compiled information. The paper cannot currently serve as a reliable reference because of an unsourced quantitative comparison table and several factual errors in the core technical descriptions. These problems are not peripheral; they occur precisely in the parts of a survey that give it its value.","major_comments":[{"comment":"The quantitative performance comparison in Table III is unsourced and contains entries that do not correspond to published evaluations. No citation, dataset split, IoU threshold, or metric definition is given, and the original R-CNN and Fast R-CNN papers do not report detection results on COCO or Open Images. Because the paper's stated contribution is to be a comprehensive and reliable overview, this table is load-bearing and cannot be verified.","section":"Section VI, Table III"},{"comment":"Faster R-CNN is misattributed to Girshick et al. with citation [18], which is the Fast R-CNN paper. The correct attribution is Ren et al. [19]. This is a factual error in a core model description and directly undermines the review's reliability.","section":"Section IV-D"},{"comment":"Several YOLO version descriptions are inaccurate. YOLOv8 is described as having a transformer-based backbone, but the cited Ultralytics repository and the actual YOLOv8 architecture use a convolutional CSPDarknet-style backbone. YOLOv9 is described as using adversarial training, whereas its cited paper proposes programmable gradient information; YOLOv10 is described as using real-time feedback loops, whereas it is an NMS-free end-to-end detector. These misstatements are factual errors, not interpretive differences.","section":"Section IV-F"},{"comment":"The reported R-CNN accuracy figures (44%, 54%, 58%, 66%) are inconsistent with published results and are presented without any protocol. The original R-CNN paper reports PASCAL VOC mAP values that differ from these numbers. The authors should either correct the values with proper citations or remove them, because the current numbers are unverifiable.","section":"Section IV-A"},{"comment":"The definition of Average Precision is incorrect. The formula sums P(k) times Precision at Recall(k) and divides by n, which is not the standard average precision or the 11-point interpolation used by the cited datasets. Since mAP is the primary metric in Table III, this misdefinition compounds the reliability problem.","section":"Section VII-B"}],"minor_comments":[{"comment":"The cross-reference to datasets says '(see Section 5)', but the datasets are described in Section VI, not Section 5.","section":"Section II"},{"comment":"The model years in Table I are inconsistent with the cited publications. For example, R-CNN is listed as 2013 but the cited paper [17] is from 2014, and SPPNet is listed as 2015 while the original publication [23] first appeared in 2014.","section":"Table I"},{"comment":"The sentence 'recent CNN architectures have significantly improved accuracy to below 5%' is unclear because object detection performance is typically reported as mAP, where higher values are better; the metric and direction need to be stated explicitly.","section":"Section VIII"},{"comment":"The entries for YOLOv5 and YOLOv8 cite only GitHub repositories; peer-reviewed references or official technical reports would be preferable in a survey of this kind.","section":"Section IV-F"}],"recommendation":"reject","confidential_remarks":"The unsourced quantitative table (Table III) and the multiple misattributions in the core technical sections suggest that the manuscript has not been carefully fact-checked. I recommend that the editor treat the review as unreliable until these issues are resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this review is a rough orientation map for newcomers, but the map has wrong labels and the one quantitative table is unsourced. Not citable, not referee-worthy as is.\n\nWhat is new: nothing technically new—it's a review. What it does do: it organizes the material into classical vs CNN-based, gives a readable section on region proposals, and has a reasonable high-level strengths/limitations table (Table I). The applications and datasets sections are fine for a first pass. For a reader who knows nothing, the narrative arc from SIFT/HOG through R-CNN to SSD is not misleading in its general shape.\n\nThe problems are load-bearing. Table III reports mAP for six detectors across four datasets with no source, split, IoU threshold, or metric definition. Several entries cannot be right: R-CNN and Fast R-CNN have no standard COCO or Open Images numbers from their original papers, and ImageNet detection is not typically reported as mAP in those references. The paper attributes Faster R-CNN to Girshick et al. citing [18], the Fast R-CNN paper, whereas the correct reference is Ren et al. [19]. Section IV.F says YOLOv8 uses a transformer-based backbone; the cited Ultralytics repo uses a convolutional CSPDarknet-style backbone. For a survey, whose value is accurate compilation, these are not typos—they undermine the central claim of being a reliable comprehensive reference.\n\nAlso, the claim of comprehensiveness is overstated: coverage stops around 2017–2020 for the core architectures, with newer work mentioned only as one-line YOLO iterations. The discussion's \"accuracy to below 5%\" is vague. These may be fixable, but the paper needs a full fact-check pass and a source for every number before it can serve the intended audience.\n\nWho it's for: an absolute beginner who wants the shape of the field and will not act on the numbers. With corrections, it could be a serviceable short survey; as it stands, the errors in the empirical claims make it risky to hand to a student.\n\nRecommendation: desk reject, or invite a major revision only if the authors are willing to redo Table III with proper citations and fix the factual errors. The core idea of the survey is fine; the execution is not.","headline":"A readable but factually unreliable review; the unsourced Table III and misattributions kill its value as a reference.","tokens_in":9418,"tokens_out":1822,"would_cite":false,"duration_ms":18417,"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":"The paper provides a single reference spanning classical feature-based detection and CNN-based detectors, organized into two-stage and one-stage families.","keywords":["Object detection","Convolutional neural networks","Classical computer vision","Two-stage detectors","One-stage detectors","YOLO","SSD","Mean average precision"],"falsifier":"Look up reference [18] and check whether it actually describes Faster R-CNN rather than Fast R-CNN, since the text credits Faster R-CNN to that same citation; a mismatch would expose an attribution error. Separately, re-run the six detectors on Pascal VOC, COCO, ImageNet, and Open Images under a single evaluation protocol and compare the resulting mAP, FPS, and model size to Table III; any large mismatch would show the table is not a dependable summary.","tokens_in":8594,"feed_emoji":"🔍","tokens_out":8665,"duration_ms":74546,"temperature":0.7,"pith_summary":"This paper is a survey of object detection algorithms rather than a new algorithm. It tries to establish that the field is best understood as two successive families—classical pipelines built from handcrafted features and a learned classifier, and deep convolutional detectors that learn features directly from data—and that a single reference covering both has been missing. The paper organizes modern CNN detectors into two-stage models (region proposals followed by classification) and one-stage models (direct regression), compares their strengths and limitations, and compiles datasets, evaluation metrics, and benchmark figures. If the survey succeeds, a reader can trace the evolution of the field and choose a detector family based on the speed-accuracy trade-off.","feed_headline":"Survey maps object detection from handcrafted features to CNN models","feed_subtitle":"Side-by-side comparison of two-stage and one-stage detectors, with datasets, metrics, and future directions.","key_machinery":"The organizing machinery is a two-way taxonomy of detector architectures. Classical object detection is described as a three-stage pipeline—proposal generation, handcrafted feature extraction, and classification—while CNN detection is described in terms of learned hierarchical features. Within the CNN family the paper distinguishes two-stage detectors, whose separate region-proposal stage (selective search or a learned Region Proposal Network) feeds a classifier, from one-stage detectors, which treat detection as direct regression or classification over a grid of anchors. The taxonomy carries the argument by converting each model's history into a sequence of bottlenecks removed: shared feature computation, ROI pooling, learned proposals, and multi-scale anchor predictions.","core_discovery":"The central claim is that a comprehensive review of object detection must cover both classical computer vision techniques and convolution-based models, and that this paper provides that coverage. Toward this end it categorizes detectors into classical techniques such as SIFT, HOG, and deformable part models, and CNN-based detectors, which it further splits into two-stage detectors (R-CNN, SPP-Net, Fast R-CNN, Faster R-CNN, Mask R-CNN) and one-stage detectors (YOLO, SSD). It presents a quantitative comparison across Pascal VOC, COCO, ImageNet, and Open Images, an evaluation-metric summary centered on IoU and mAP, and a list of future directions including tiny-object detection, 3D detection, multi-modal detection, and few-shot learning. The review's conclusion is that deep learning largely solved the semantic limitations of handcrafted features but increased complexity and resource demands, leaving deployment efficiency as an open issue.","pith_inferences":["The authors do not say this, but their taxonomy suggests a deployment-selection rule: resource-constrained edge devices might prefer classical lightweight models or one-stage detectors, while accuracy-critical cloud applications favor two-stage detectors.","A natural extension the paper leaves implicit is a third family of transformer-based detectors; their absence implies the review's comparison is bounded to the CNN era.","One testable consequence of the review's framing: if the listed limitations are the real bottlenecks, then tiny-object benchmarks such as COCO small-object AP should separate two-stage and one-stage models more sharply than overall mAP does."],"forward_implications":["If the comparison is correct, the field's progress can be read as the progressive removal of handcrafted components: each two-stage model removed one bottleneck, and one-stage models removed the proposal stage entirely.","Readers can use the reported benchmarks as a shortcut for choosing a detector: the R-CNN family favors accuracy and segmentation, while YOLO and SSD favor real-time speed.","Mask R-CNN's placement in the taxonomy implies that instance segmentation is a natural extension of two-stage detection rather than a separate task.","The review's future directions point to the concrete limitations of current detectors: small objects, crowded scenes, 3D understanding, multi-modal inputs, and few-shot regimes."],"supporting_citations":[{"why":"Supplies the HOG-features-plus-SVM detector that anchors the classical half of the review.","marker":"[6]"},{"why":"Introduces the Deformable Part Model, the classical method the review contrasts with deep models.","marker":"[8]"},{"why":"Defines the two-stage R-CNN pipeline that the review presents as the start of CNN-based detection.","marker":"[17]"},{"why":"Introduces Fast R-CNN with ROI pooling and end-to-end training, a key step in the two-stage lineage.","marker":"[18]"},{"why":"Introduces the Region Proposal Network, making proposals learnable in Faster R-CNN.","marker":"[19]"},{"why":"Extends Faster R-CNN to instance segmentation with masks and RoIAlign.","marker":"[20]"},{"why":"Defines YOLO, the one-stage detector that the review credits with real-time speed.","marker":"[21]"},{"why":"Defines SSD's multi-scale anchor design, the other one-stage model in the comparison.","marker":"[22]"},{"why":"Introduces SPP-Net's spatial pyramid pooling, the intermediate architecture between R-CNN and Fast R-CNN.","marker":"[23]"}],"fun_headline_variants":["From SIFT to YOLO: A survey of object detection","Object detection review: classic features vs CNNs","Deep CNN detectors excel but cost more compute","Two-stage vs one-stage: the object detection survey","Handcrafted to learned features: object detection review"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The review's usefulness as a reference depends on the accuracy of the compiled performance numbers in Table III and on the technical descriptions of the models being correct; if those figures or attributions are wrong, the overview can mislead rather than guide.","fun_headline_variants_meta":{"raw":{"variants":["From SIFT to YOLO: A survey of object detection","Object detection review: classic features vs CNNs","Deep CNN detectors excel but cost more compute","Two-stage vs one-stage: the object detection survey","Handcrafted to learned features: object detection review"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000496,"raw_usage":{"total_tokens":2411,"prompt_tokens":900,"completion_tokens":1511,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":1445}},"tokens_in":516,"tokens_out":1511,"duration_ms":10542,"temperature":1.0,"reasoning_tokens":1445,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:49:30.782500+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Look up reference [18] and check whether it actually describes Faster R-CNN rather than Fast R-CNN, since the text credits Faster R-CNN to that same citation; a mismatch would expose an attribution error. Separately, re-run the six detectors on Pascal VOC, COCO, ImageNet, and Open Images under a single evaluation protocol and compare the resulting mAP, FPS, and model size to Table III; any large mismatch would show the table is not a dependable summary.","supporting_citations":[{"cited_title":"A discriminatively trained, multiscale, deformable part model,","cited_arxiv_id":null,"evidence_quote":"Introduces the Deformable Part Model, the classical method the review contrasts with deep models."}],"review_version":1}