{"id":"09aa1909-9a82-4449-821a-4144334340e9","arxiv_id":"2411.16316","paper_version":6,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A structured review of 2D and 3D monocular lane detection methods, with a new four-axis taxonomy and unified FPS comparisons.","lead":"This survey organizes deep-learning-based monocular lane detection into four design axes: task paradigm, lane modeling, global context supplementation, and perspective effect elimination. It covers both 2D and 3D methods, compares benchmarks, and reports re-tested inference speeds on a single GPU.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unified FPS comparison is not fully unified: three entries use borrowed reimplementations and input sizes vary, so efficiency rankings in Table VI are approximate.","rationale":"The survey provides broad coverage, a plausible four-axis organizational scheme, and useful benchmark tables; no fatal flaw in the central taxonomy was found. The most load-bearing concern is the efficiency comparison, because it is the only new empirical contribution and is advertised as a unified fair test. The paper is transparent about the borrowed entries, but that transparency does not make the comparison unified. Minor factual slips, such as LaneAF appearing in both the segmentation-based and object-detection-based blocks of Table VII and the CLGo [23] citation mismatch in Table III, are real but peripheral. They reinforce a general need for careful table auditing rather than overturning the survey's contribution. Since the reader's conditional verdict already flags the efficiency and comparability assumptions, this stress-test agrees and recommends keeping the verdict unchanged with the caveat that the efficiency section should be treated as indicative, not definitive.","tokens_in":53,"tokens_out":5094,"duration_ms":109013,"concrete_test":"Run all 17 methods in Table VI from their official open-source code on the same RTX 3090 at a single common input resolution (e.g., 320x800), measuring end-to-end latency that includes preprocessing and post-processing. Recheck SCNN, LSTR, and 3D-LaneNet in particular; if their FPS differs by more than 20% from Table VI, or if any segmentation-based vs object-detection-based ranking flips, revise the efficiency discussion and soften the 'unified setting' claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V-B/Table VI claims a unified environment for fair efficiency comparison, a claim repeated in the abstract and Contribution 3. In practice the table mixes sources: SCNN and LSTR are not run by the authors but are taken from Feng et al.'s reimplementation [20], and 3D-LaneNet is tested via Guo et al.'s reproduction [23] because the original is closed-source. Input resolutions range from 288x800 to 720x960 across methods, and post-processing ranges from none to NMS or clustering, yet the reported FPS is pure model inference without preprocessing or post-processing. Section V-C then grounds general guidance (e.g., segmentation-based methods are less efficient; UFLD is fastest) on these numbers. If the borrowed reimplementations are not faithful, or if the differing resolutions systematically favor certain architectures, the empirical recipe could mislead readers selecting baselines. The four-axis taxonomy is not invalidated by this, but the paper's claimed empirical contribution is weaker than the 'unified setting' language suggests.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript surveys deep-learning-based monocular lane detection, covering both 2D and 3D approaches. Its main organizational contribution is a four-axis taxonomy: task paradigm (instance-level discrimination), lane modeling (representing lanes as learnable parameters), global context supplementation (handling obscure lanes), and perspective effect elimination (obtaining accurate 3D lanes). The paper describes datasets and metrics, reviews representative 2D and 3D methods under this taxonomy, reports benchmark numbers on CULane, OpenLane, and other datasets, presents an efficiency comparison on a single RTX 3090 (Table VI), and adds chapters on extended tasks (multi-task perception, video lane detection, online HD map construction, lane topology reasoning) and future directions.","tokens_in":42427,"tokens_out":6384,"duration_ms":69053,"significance":"If the taxonomy is accepted, the survey provides a useful organizing framework that goes beyond the narrower structure of earlier surveys [30–32] by jointly covering 2D and 3D lane detection and by stressing design choices that matter for downstream use. The benchmark tables are a convenient reference and are transparently traced to the original papers, and the companion GitHub collection of papers and codes is a practical asset for the community. The efficiency comparison in Section V-B is a good-faith attempt to add an original empirical contribution, but, as detailed below, the claim that it is a 'unified setting' is stronger than the protocol supports. The survey does not introduce fitted quantities or predictions, so there is no circular-reasoning concern; the main risk is that readers may over-interpret the efficiency ranking and the cross-paper benchmark numbers.","major_comments":[{"comment":"The claim of a 'unified environment' for fair efficiency comparison is not fully supported by the protocol. The table mixes input resolutions from 288×800 to 720×960, post-processing that ranges from none to NMS to clustering, and three entries are borrowed from other groups' reimplementations rather than run by the authors: SCNN and LSTR use Feng et al.'s reimplementation [20], and 3D-LaneNet uses Guo et al.'s reproduction [23]. Since FPS is measured as pure model inference without preprocessing or post-processing, the ranking cannot be read as an end-to-end deployment comparison. Section V-C then uses these numbers to ground practical guidance (e.g., that segmentation-based methods are less efficient, that UFLD is the fastest). I recommend either rerunning the speed tests at matched input resolutions with the required post-processing included, or explicitly reframing Table VI as per-configuration reference numbers and softening the corresponding statements in the abstract, Contribution 3, and Section V-C.","section":"§V-B, Table VI; Contribution 3 in §I"},{"comment":"The benchmark tables present numbers collected from different original papers side by side, and readers may infer that the F1 values are directly comparable across rows. The paper describes the standard metrics in Section II-B but does not state that evaluation details can differ across papers (for example, matching thresholds, possible test-time augmentation, or ensemble configurations). I ask the authors to add an explicit caveat that reported numbers are copied as published and that small cross-row differences should be interpreted with caution. This does not undermine the survey's taxonomy, but it is part of the paper's empirical contribution.","section":"§V-A, Tables IV and V"}],"minor_comments":[{"comment":"The table marks both '2D' and '3D' for LLAMAS, but Section II-A1 discusses LLAMAS among the 2D lane detection datasets; please clarify whether 3D annotations exist or remove the 3D checkmark.","section":"Table I, LLAMAS row"},{"comment":"CLGo is cited as [23], but the text attributes this method to Liu et al. [131] (AAAI'22); the citation should be corrected for consistency with Table VIII.","section":"Table III, CLGo row"},{"comment":"LaneAF appears in both the segmentation-based and the object detection-based blocks; please remove the duplicate or indicate its intended category.","section":"Table VII"},{"comment":"The sentence 'The Apollo 3DLane dataset [23] is generated using the game engine' cites Gen-LaneNet [23], but the dataset was introduced in the 3D-LaneNet paper [22]; please update the citation.","section":"§II-A2"},{"comment":"There are several typos: 'The sturcture of this paper' in Section I, a stray 't' after 'segmentation process.' in Section III-B1, and 'Derictly Modeling 3D Lanes' in Section IV-C.","section":"§I and §III-B1 and §IV-C"},{"comment":"The spelling of Gen-LaneNet is inconsistent ('GenLaneNet' in Table V and 'Gen-LaneNet' elsewhere), and the BézierLaneNet row contains a stray space in the method name.","section":"Table VI and text"}],"recommendation":"major_revision","confidential_remarks":"The survey's central taxonomy is sound and the paper makes a useful contribution to the literature. The main issue is the overclaimed 'unified' efficiency comparison; it is fixable by reframing or rerunning the experiment, so rejection is not warranted. I would also gently check that the claim of being the 'first survey' covering both 2D and 3D lane detection is accurate with respect to very recent work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this survey is worth sending to referees. It is the first to my knowledge to put 2D and 3D monocular lane detection under one taxonomy, and the four axes—task paradigm, lane modeling, global context supplementation, perspective effect elimination—are a real organizational advance, not just repackaged section headings. The benchmark tables cite original papers, the coverage runs through 2024, and the extended-works sections (video, HD map, topology reasoning) make it a useful one-stop map for someone entering the field or choosing baselines. The maintained GitHub repo is a plus.\n\nThe main flaw is the efficiency comparison. The abstract and Contribution 3 say the authors 'investigate their inference speed under a unified setting for fair comparison,' but Table VI is not as unified as advertised. Three entries are borrowed from other people's reimplementations (SCNN and LSTR from Feng et al., 3D-LaneNet from Guo et al.), and input resolutions range from 288x800 to 720x960 with post-processing ranging from none to NMS or clustering. The FPS numbers are therefore only loosely comparable, and Section V-C draws general lessons ('segmentation-based methods are less efficient', 'UFLD is fastest') from that table. The qualitative conclusions probably survive—the differences are large enough—but the precise rankings and the 'unified setting' language overstate the evidence. This is a moderate, fixable problem: either re-run everything at matched resolutions without post-processing, or table the numbers with clear provenance and caveats.\n\nThe camera-projection math is standard and correct. There is no new experimental claim beyond the FPS re-test, which is what makes the overclaim there matter more than it would in a research paper. The 'first survey covering both 2D and 3D lane detection' claim is plausible given the cited prior surveys focus on one side. The criteria for selecting representative methods could be stated more explicitly, but that is minor.\n\nBottom line: the taxonomy alone justifies engagement. With the efficiency section cleaned up, this would be a solid reference. I would send it to referees now and ask for a revision on the speed comparison.","headline":"Useful 2D+3D lane detection survey with a genuinely new taxonomy; the 'unified' FPS table is the one place that needs an honest fix.","tokens_in":42945,"tokens_out":2053,"would_cite":true,"duration_ms":19172,"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":"A survey of deep-learning lane detection organizes every method around four core design choices.","keywords":["lane detection","monocular 3D lane detection","deep learning survey","autonomous driving perception","bird's-eye view","inverse perspective mapping","lane modeling","instance discrimination"],"falsifier":"Re-run the open-source methods from the efficiency table at a single input resolution with identical post-processing on the same GPU; if the FPS ordering changes substantially, the unified efficiency claim is weakened. Also, a method that achieves state-of-the-art results on CULane or OpenLane while fitting none of the four proposed axes would count against the taxonomy.","tokens_in":1627,"feed_emoji":"🛣️","tokens_out":1566,"duration_ms":49232,"temperature":0.7,"pith_summary":"This survey tries to establish that all deep-learning monocular lane detection methods, both 2D and 3D, can be understood through four core design axes: task paradigm (how lane instances are told apart), lane modeling (how a lane is encoded as learnable parameters), global context supplementation (how obscure lanes are inferred), and perspective effect elimination (how 3D lanes are obtained). It argues that this four-axis view is more useful than earlier surveys that focused on network structures and loss functions. It also claims to deliver the first survey covering both 2D and 3D lane detection together, along with a unified efficiency comparison of representative methods. A sympathetic reader would care because a correct taxonomy tells practitioners which design choices actually drive accuracy, speed, and usability for downstream planning.","feed_headline":"Four design choices explain every deep-learning lane detector","feed_subtitle":"A new survey unifies 2D and 3D lane detection methods and re-tests their speed on one GPU.","key_machinery":"The central organizing device is the four-axis decomposition of lane detection algorithms. The axes are: instance-level discrimination (two-stage segmentation-based versus one-stage object detection-based), lane modeling (mask, grids, keypoints, line anchors, or curves), global context supplementation (special modules for obscure lanes), and perspective effect elimination (IPM, learnable BEV transformation, depth-based projection, or direct 3D modeling). This decomposition carries the survey's argument by providing the classification structure for all reviewed methods, the categories used in the benchmark tables, and the framework for the efficiency comparison and the discussion of empirical recipes.","core_discovery":"On the paper's own terms, the central claim is that the design space of monocular lane detection is spanned by four decisions. The task paradigm determines whether instance-level discrimination happens before, after, or alongside lane localization. Lane modeling determines whether a lane is represented as a mask, a grid of row-wise classifications, a set of keypoints, a line anchor, or a parametric curve. Global context supplementation covers mechanisms added to infer lanes that are occluded, shadowed, or otherwise visually weak. Perspective effect elimination covers how the camera's perspective distortion is handled, either through inverse perspective mapping under a flat-ground assumption, learnable front-view-to-BEV transformation, depth-based projection, or direct prediction of 3D lanes. The paper asserts that these four axes give a general pipeline for each method family, make 2D and 3D methods comparable, and are supported by the benchmark and speed results it reports.","pith_inferences":["Editorial inference: the four-axis taxonomy likely transfers to adjacent road-perception tasks such as online HD map construction, where map elements also need instance discrimination, parametric modeling, global context, and geometric projection.","Editorial inference: read together with the benchmark tables, the unified FPS tests imply that speed numbers from original papers are not reliable for method selection, and a standardized resolution-plus-post-processing protocol would be needed for deployment decisions.","Editorial inference: because line-anchor and grid-based models struggle with U-shaped and near-horizontal lanes, a shift toward mask-based or curve-based modeling is plausible as multi-camera surround-view lane detection grows.","Editorial inference: the convergence of top 3D methods on BEV-free direct modeling suggests that camera-pose estimation and 3D lane modeling are becoming coupled in one network, which could eventually make IPM-based pipelines obsolete."],"forward_implications":["If the four-axis view is right, a new lane detector can be described and positioned by where it falls on each axis, without needing to know its network architecture or loss function.","The flat-ground assumption of inverse perspective mapping is identified as the main weakness of 2D-to-3D projection, so methods that model 3D lanes directly avoid an error source that accurate 2D detection cannot fix.","Segmentation-based paradigms pay a clear efficiency cost for instance discrimination compared with one-stage object detection-based paradigms.","Supplementing global context is a near-universal design choice, and the survey associates its absence with weaker performance on occluded or otherwise obscured lanes.","State-of-the-art results on the OpenLane benchmark currently come from BEV-free methods that model 3D lanes directly rather than through a BEV feature intermediate."],"supporting_citations":[{"why":"Supplies the foundational segmentation-based approach with spatial CNN message passing and defines the CULane benchmark used throughout the survey.","marker":"[36]"},{"why":"Introduces the first end-to-end monocular 3D lane detection network with a differentiable IPM, the starting point for 3D methods.","marker":"[22]"},{"why":"Provides an IPM-based 3D lane detection framework with scalable supervision and establishes the Apollo 3DLane benchmark and its evaluation metrics.","marker":"[23]"},{"why":"Introduces learnable front-view-to-BEV view transformation via deformable attention and the OpenLane benchmark, the main 3D comparison dataset.","marker":"[24]"},{"why":"Provides the real-world ONCE-3DLanes benchmark and the SALAD depth-estimation-based BEV-free method, including the chamfer-distance evaluation metric.","marker":"[44]"},{"why":"Introduces dynamic kernel generation for grid-based lane modeling with top-down instance discrimination, a representative CondLaneNet pipeline.","marker":"[17]"},{"why":"Establishes the line-anchor-based object detection approach with ROIGather and Line IoU loss, a strong 2D baseline cited in comparisons.","marker":"[19]"},{"why":"Provides the Anchor3DLane BEV-free approach that projects 3D line anchors onto front-view features, a key reference for direct 3D lane modeling.","marker":"[25]"},{"why":"Supplies the LATR transformer method with dynamic 3D ground position embeddings, reported as a top performer on OpenLane.","marker":"[27]"}],"fun_headline_variants":["Four design choices drive all lane-detection systems","Lane detection: one survey, four design axes","All deep lane detectors stem from four choices","Survey: four decisions unify 2D and 3D lane detection","Four design choices explain every lane detector"],"cache_read_input_tokens":45184,"weakest_assumption_plain":"The survey's comparisons assume that accuracy numbers taken from the original papers, and FPS numbers measured on one GPU at each method's own input resolution and post-processing pipeline, are comparable across methods.","fun_headline_variants_meta":{"raw":{"variants":["Four design choices drive all lane-detection systems","Lane detection: one survey, four design axes","All deep lane detectors stem from four choices","Survey: four decisions unify 2D and 3D lane detection","Four design choices explain every lane detector"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000499,"raw_usage":{"total_tokens":2454,"prompt_tokens":970,"completion_tokens":1484,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":1409}},"tokens_in":586,"tokens_out":1484,"duration_ms":11917,"temperature":1.0,"reasoning_tokens":1409,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:14:00.549333+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the open-source methods from the efficiency table at a single input resolution with identical post-processing on the same GPU; if the FPS ordering changes substantially, the unified efficiency claim is weakened. Also, a method that achieves state-of-the-art results on CULane or OpenLane while fitting none of the four proposed axes would count against the taxonomy.","supporting_citations":[],"review_version":1}