{"id":"1dca07d0-8bc6-4e50-9148-f4a51cdf3185","arxiv_id":"1908.03673","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":0.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A structured survey of deep learning object detection covering two-stage and one-stage detectors, feature learning, training strategies, applications, and benchmarks up to 2019.","lead":"This paper surveys deep learning methods for object detection in images, organizing the field into detection components, learning strategies, and applications. It is a reference review, not a new scientific result, and is useful as an entry point to the 2019 state of the art.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark table entries mix unequal settings; a few rows are internally inconsistent, so the survey's comparative claims need verification of source values.","rationale":"The reader's central concern is about comparability of benchmark numbers across methods that differ in training data, test-time augmentation, and implementation details. I partially agree, but the deeper issue is not only cross-method comparability; it is that specific table entries appear to be internally inconsistent with their cited sources under the survey's own stated protocol. For example, Table 2 lists R-CNN at 66.0 mAP on VOC2007 with a footnote saying the model is trained only on VOC2007 trainval, while the original R-CNN paper reports 58.5 mAP for VOC2007 trainval training and the 66.0 figure involves additional training data and ensembling. Similarly, Table 3 lists DeepRegionlets with ResNet-101 and a 2018 date, but the original paper uses VGG-16, and the 39.3/59.8 numbers correspond to their multi-scale setting. These are concrete, checkable discrepancies that undermine the survey's central claim of providing a reliable map of the field. A conditional verdict is appropriate: the survey can be accepted as a useful qualitative review, but the benchmark tables should either be corrected to match original protocols or be clearly caveated as citing best-reported numbers under heterogeneous conditions. I do not see a need to reject the paper; the qualitative taxonomy and narrative are sound and the error is in the quantitative comparison layer. My proposed test is one independent re-derivation of the flagged table rows from the primary sources, which is a single concrete check that would settle the concern.","tokens_in":45834,"tokens_out":1817,"duration_ms":17824,"concrete_test":"For each row in Tables 2 and 3 flagged for potential inconsistency (R-CNN, SPP-net, MR-CNN, OHEM++, RON384++, RetinaNet800++, DeepRegionlets, SNIP++, SNIPER++, PANet++, and all '++' entries), re-extract the reported mAP/AP values from the cited original paper's own tables or abstract, recording the exact training data, backbone, input size, test-time augmentation, and NMS settings. Then check whether the survey's Table value matches the original paper's number under that exact protocol. If any row differs by more than 1 mAP point or lacks a matching protocol, the table needs a correction or a footnote; if all numbers match the original sources with protocols documented, the comparability concern is resolved for those rows.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The organizing claim is that Tables 2 and 3 offer a reliable comparative map of detector performance, and the reader's weakest assumption identifies the comparability issue. Spot-checking the table contents against the original papers reveals concrete inconsistencies. In Table 3, DeepRegionlets [140] is listed with ResNet-101, 2018, AP/AP50/AP75 = 39.3/59.8/--, whereas the published CVPR 2018 paper reports 39.3 AP and 59.8 AP50 with a VGG-16 backbone, using a multi-scale testing variant (the base model scores 39.3 AP, AP50 59.8). The table also lists OHEM++ and RON384++/RetinaNet800++/++ variants without specifying that these are multi-scale/test-time augmentation results, which violates the stated default comparison protocol. More seriously, Table 2 attributes R-CNN VGG-16 with 66.0 mAP on VOC2007 but labels it with an asterisk that only the model is trained on VOC2007 trainval; however, the original R-CNN paper reports 58.5 mAP for that setting, with the 66.0 figure coming from ensembling bounding-box regressions with selective search and voc2012 data. If the table entries are not corrected or footnoted with their exact evaluation protocols, the survey's utility as a reliable field map—its central purpose—is compromised. A single computational check can settle the concern: recompute or re-extract each flagged row's numbers from the cited original papers and compare against the stated training/testing protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey of deep learning methods for visual object detection, organized into detection components, learning strategies, and applications and benchmarks. It reviews two-stage and one-stage detectors, backbone architectures, proposal generation, feature learning, training and testing strategies, and specialized tasks such as face and pedestrian detection, along with public benchmarks and future directions. The paper also provides two large comparison tables (Pascal VOC and MS COCO) intended to summarize the state of the art.","tokens_in":46268,"tokens_out":6750,"duration_ms":66663,"significance":"If the survey's content is accurate, it offers a useful structured overview of a fast-moving field, with broad coverage and a substantial reference list. The taxonomy (two-stage vs. one-stage, components, learning strategies, applications) is reasonable and the milestone timeline is helpful for orientation. However, the central value of such a survey depends on the reliability of its compiled benchmark results, and the concrete misattributions identified below materially reduce that reliability. The contribution is therefore of moderate significance and needs a careful revision of the benchmark tables and related citations before it can serve as a dependable map of the field.","major_comments":[{"comment":"The R-CNN row reports 66.0 mAP on VOC2007 with an asterisk claiming that the model is trained only on VOC2007 trainval. This is inconsistent with the original R-CNN paper, which reports 58.5 mAP for a VGG-16 model trained on VOC2007 trainval; the 66.0 figure corresponds to a different configuration (with additional VOC2012 data and/or ensembled bounding-box regression). As written, the entry misleads readers about the training protocol and inflates the comparison for that row. Please re-verify the number and either correct the value or clarify the exact configuration in the footnote.","section":"Table 2 (Section 7)"},{"comment":"The DeepRegionlets row lists ResNet-101 as the backbone, but the original ECCV 2018 paper reports its COCO test-dev result (39.3 AP, 59.8 AP50) with a VGG-16 backbone and a multi-scale testing variant. The combination of a wrong backbone and the missing '++' marker means the row is not comparable to other rows in the same table that use standard inference. Please correct the backbone and indicate the test-time protocol, or remove the row if the original protocol cannot be cleanly accommodated.","section":"Table 3 (Section 7)"},{"comment":"The table mixes base models, entries marked '++' (multi-scale testing, horizontal flip, etc.), and entries such as DeepRegionlets that use multi-scale testing without the marker. The caption defines '++' but does not state that rows with '++' and rows without it are not directly comparable. Given that the survey's organizing claim is to provide a reliable comparative map of detector performance, the table should include per-row protocol indications (or a separate column) so that readers can make valid comparisons; otherwise the benchmark tables may mislead rather than inform.","section":"Table 3 (Section 7)"}],"minor_comments":[{"comment":"The SPP-net row cites reference [2], but the correct reference for SPP-net is [47].","section":"Table 2"},{"comment":"The footnote contains duplicated words ('the the model is trained') and a subject-verb agreement error ('the model are trained'); please correct both.","section":"Table 2 footnote"},{"comment":"The name 'Single-Shot Mulibox Detector' should be 'Single-Shot Multibox Detector'.","section":"Section 3.2.2"},{"comment":"The phrase 'Precise ROI Pooing' should be 'Precise ROI Pooling'.","section":"Section 3.5.2"},{"comment":"The name 'Hosong et al.' should be 'Hosang et al.' (the authors of 'Learning non-maximum suppression').","section":"Section 4.2.1"},{"comment":"The word 'sveral' should be 'several'.","section":"Section 8"},{"comment":"The opening sentence lists 'Pascal VOC2007, VOC2007 and MSCOCO'; the second occurrence should be 'VOC2012'.","section":"Section 7"}],"recommendation":"major_revision","confidential_remarks":"I recommend asking the authors to re-extract and verify all benchmark numbers from the original papers, and to clearly document the protocol for each row in Tables 2 and 3. The identified errors are correctable and do not undermine the survey's taxonomy, but the comparative tables are a central part of the contribution and need to be reliable. I would also ask the authors to review the citation assignment in the benchmark tables for similar misattributions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. This survey gives a genuinely organized map of deep learning object detection as of mid-2019, with a sensible three-part split (components, learning strategies, applications/benchmarks) and a broad, accurate description of the main detector families. Nothing in it is new—it is explicitly a survey—but that is fine; the contribution is the organization and the coverage. The second thing: the benchmark tables are the weakest part, and the paper's central comparative claim rests on them. The stress-test note is right. Table 2 reports R-CNN at 66.0 mAP on VOC2007 trained on VOC2007 trainval only; the original R-CNN paper reports 58.5 for that exact setting, and 66.0 comes from a different, larger training setup. Table 3 lists DeepRegionlets with ResNet-101, but the published paper uses VGG-16 and reports 39.3/59.8 with multi-scale testing. Several '++' entries are flagged only by a footnote; the table never states what the default protocol is, so readers can't tell which rows are directly comparable. There are also citation slips—SPP-net is cited as [2] instead of [47] in Table 2—and assorted typos.\n\nThe conceptual sections are much better. The descriptions of R-CNN, Fast/Faster R-CNN, FPN, YOLO, SSD, RetinaNet, anchor-free methods, and the discussion of learning strategies like loss design and NMS are consistent with the literature and are a fair representation of the field. The applications section, especially face and pedestrian detection, is a good entry point for someone outside those subareas. The survey reads like a competent review, not a careless collage.\n\nBut the tables are not decoration; they are the main practical reason to use a survey like this. If a practitioner compares numbers without checking the original papers, they can easily be misled. The good news is that this is fixable: every flagged row needs a re-extraction of the source value and a footnote about the exact training/testing protocol. That is a mechanical correction, not a research problem.\n\nWho is this for? A new graduate student who wants the landscape of detectors before 2020, or a researcher from another area who needs a quick orientation. I would not cite its numbers in work right now until the tables are corrected. But I would send it to review: the scope is useful, the narrative is sound, and the errors are concentrated in a few rows. A serious referee would catch them, and a revision would make this a dependable reference.","headline":"A useful but imperfect map of deep learning object detection as of 2019; the narrative is solid, but the benchmark tables need fixing before you trust them.","tokens_in":46549,"tokens_out":3211,"would_cite":false,"duration_ms":30713,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This survey claims that deep-learning object detection is best organized as a two-stage/one-stage dichotomy, with components, learning strategies, and benchmarks making up the rest of the map.","keywords":["object detection","deep learning","deep convolutional neural networks","two-stage detectors","one-stage detectors","feature pyramid networks","anchor-free detection","instance segmentation"],"falsifier":"Retrain two detectors from the same benchmark table under an identical protocol, using the same training data, backbone, augmentation, and test settings, and compare mAP; if the ranking inverts or the gap collapses, the survey's cross-method comparisons are not reliable.","tokens_in":45668,"feed_emoji":"🎯","tokens_out":6470,"duration_ms":66887,"temperature":0.7,"pith_summary":"This survey argues that recent deep-learning object detection is best understood through a single split: two-stage detectors that first propose regions and then classify them, versus one-stage detectors that predict classes and boxes directly from feature maps. It claims that performance is determined by a small set of reusable components, namely backbone architecture, proposal generation, and feature representation, together with learning strategies such as data augmentation, imbalance sampling, localization refinement, and cascade learning. If the map is right, it gives researchers and practitioners a reliable way to locate any detector, compare the accuracy-versus-speed tradeoff, and choose where to improve next. The survey also assembles benchmark tables and reviews face and pedestrian detection as the main applied settings.","feed_headline":"Two families organize seven years of deep object detectors","feed_subtitle":"A systematic review sorts detectors by components, learning strategies, applications, and benchmarks, and maps where the field is heading.","key_machinery":"The machinery that carries the survey is its taxonomy, built on the two-stage/one-stage dichotomy as the root. The taxonomy organizes every reviewed method into detection components, learning strategies, and applications and benchmarks, so that a detector is described by choices within each category rather than by a single headline number. Within that structure, the review treats feature pyramids and multi-scale feature learning as the key mechanism for handling scale variation, and class-imbalance handling through hard negative mining and focal loss as the key mechanism for training one-stage detectors. The IoU-based evaluation metrics and the benchmark tables are the instruments that make the taxonomy's comparisons concrete.","core_discovery":"On the paper's own terms, the central claim is that every important deep object detector falls into one of two paradigms: two-stage detectors, which use a proposal generator to produce a sparse set of candidate regions and then classify and refine each one, and one-stage detectors, which skip proposal generation and make dense predictions at every location. The survey further claims that the factors controlling detection quality decompose into detection components, including backbone networks, proposal generation, multi-scale and deformable feature learning, and region encoding, plus learning strategies such as data augmentation, imbalance sampling, localization refinement, cascade learning, and test-time processing such as non-maximum suppression. It presents benchmark tables on Pascal VOC and MS COCO that track the accuracy and speed of representative detectors, and it identifies anchor-free, keypoint-based detectors and AutoML-designed architectures as the most active directions going forward.","pith_inferences":["The two-family split will probably blur as components migrate, since two-stage detectors adopt anchor-free heads and one-stage detectors adopt cascaded refinement, so a later survey may need a finer-grained axis.","Because the benchmark tables mix training data, backbones, and test-time augmentation, the reported rankings should be read as approximate; identical-protocol re-runs could reorder methods.","A testable extension of the survey's decomposition is to ablate components one at a time within a fixed backbone, and the largest mAP swings would identify which part of the taxonomy carries the field's progress.","The review's list of open problems suggests that low-shot detection and detection-specific backbones may matter as much as anchor-free design in the next stage."],"forward_implications":["If the taxonomy is right, a new detector can be located and understood by its place in the two-stage/one-stage split and by its component choices, giving the field a stable reference map as of mid-2019.","The accuracy-versus-speed tradeoff between the two families becomes a design expectation rather than an accident: two-stage detectors set accuracy records while one-stage detectors set real-time records.","Class imbalance and scale variation emerge as first-order training problems, so progress in sampling strategies, loss design, and feature pyramids should keep improving accuracy across both families.","The benchmark tables imply that anchor-free keypoint detectors and AutoML-based architectures are the directions most likely to push state-of-the-art results next."],"supporting_citations":[{"why":"Supplies the residual backbone that most detectors in the benchmark tables use as the feature extractor.","marker":"[1]"},{"why":"Establishes the two-stage detection paradigm and the practice of starting from pretrained CNN features.","marker":"[2]"},{"why":"Extends detection to instance segmentation with mask predictions and ROI Align, defining the mask-level setting.","marker":"[3]"},{"why":"Introduces the Region Proposal Network, making proposal generation learned and end-to-end for two-stage detectors.","marker":"[34]"},{"why":"Introduces shared feature computation and end-to-end multi-task training for two-stage detectors.","marker":"[38]"},{"why":"Introduces the feature pyramid network that most later detectors use to handle objects across scales.","marker":"[39]"},{"why":"Defines the one-stage multi-scale anchor-based detector and its hard-negative-mining training recipe.","marker":"[42]"},{"why":"Introduces focal loss, the tool one-stage detectors use to counter foreground-background class imbalance.","marker":"[43]"},{"why":"Establishes the keypoint-based anchor-free paradigm by detecting objects as pairs of corners.","marker":"[63]"}],"fun_headline_variants":["Survey sorts deep detectors into one-stage and two-stage","Two-stage vs one-stage: the deep detection divide","Deep detectors fall into two camps: one-stage and two-stage","Survey maps deep detectors: components, strategies, benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey assumes that the benchmark scores reported by different papers can be compared directly in Tables 2 and 3, even though those papers differ in training data, test-time augmentation, backbone, and implementation details.","fun_headline_variants_meta":{"raw":{"variants":["Survey sorts deep detectors into one-stage and two-stage","Two-stage vs one-stage: the deep detection divide","Deep detectors fall into two camps: one-stage and two-stage","Survey maps deep detectors: components, strategies, benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001012,"raw_usage":{"total_tokens":4247,"prompt_tokens":890,"completion_tokens":3357,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":3292}},"tokens_in":506,"tokens_out":3357,"duration_ms":20972,"temperature":1.0,"reasoning_tokens":3292,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:05:00.051891+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain two detectors from the same benchmark table under an identical protocol, using the same training data, backbone, augmentation, and test settings, and compare mAP; if the ranking inverts or the gap collapses, the survey's cross-method comparisons are not reliable.","supporting_citations":[],"review_version":1}