{"id":"170ccd9c-e34d-4b7c-99e5-727b9a2b0a35","arxiv_id":"2505.12246","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Using SD map priors through hybrid raster-vector fusion and an intersection keypoint auxiliary task improves lane perception and topology reasoning on OpenLane-V2 by up to 5.9 OLUS points.","lead":"SEPT is a computer-vision system that uses ordinary navigation maps (SD maps) to help self-driving cars see and understand road structure, such as lane connections and intersections, from camera images. It reports large improvements on the OpenLane-V2 benchmark over two strong baselines, which matters because cheap SD maps could reduce reliance on expensive high-definition maps in mapless driving.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The closest concurrent method, TopoSD, is cited but never benchmarked, so the central 'outperforming existing methods by a substantial margin' claim is not yet supported.","rationale":"I read SEPT as a method paper: the central claim is that fusing SD maps through a hybrid representation plus IKPD improves OpenLane-V2 perception and topology beyond existing methods. The ablations support the internal component story: each added component improves OLUS, and the FT/DGFF comparisons are useful. The weakest link is external comparison. Related Work identifies TopoSD as the closest concurrent method combining raster and vector streams, and explicitly criticizes its fusion, but the experiments do not include it. Without that number, 'outperforming existing methods by a substantial margin' is a claim about a comparison that has not been run. I would keep the conditional verdict: the internal evidence is credible, but the headline requires the missing TopoSD comparison before acceptance. This is not an accusation of bad faith; it is a standard completeness check for a SOTA claim. A single benchmark of TopoSD would settle it.","tokens_in":12210,"tokens_out":6263,"duration_ms":61631,"concrete_test":"Run TopoSD's released checkpoint or official training code on the OpenLane-V2 subset A validation split and compute OLS v1.0/v1.1 and OLUS using the same protocol as Tables I and II, then insert TopoSD into both tables. If TopoSD's OLS/OLUS is equal to or above SEPT's values, the abstract's 'substantial margin' claim over existing methods fails; if SEPT remains ahead, the comparison should be reported to verify the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II-C cites TopoSD [11] as a concurrent approach that 'combines these two distinct streams' of rasterized and vectorized SD map encoding, and then asserts its fusion is 'overly simplistic.' TopoSD is the most direct published competitor for SEPT's hybrid SD-map fusion on OpenLane-V2, yet it appears nowhere in Tables I or II. The abstract's claim of 'outperforming existing methods by a substantial margin' therefore rests on comparisons against TopoNet, LaneSegNet, SMERF, and TopoLogic, but not against the one method whose design overlaps most with SEPT's. Since TopoSD was available on arXiv in November 2024, well before this submission, the omission is not a matter of timing. The paper's internal ablations are well controlled and support the claim that the proposed modules help over a fixed baseline; the open question is external competitiveness. If TopoSD's published OLV2 subset A numbers meet or exceed SEPT's OLS/OLUS, the headline margin claim would not hold; if SEPT is better, the absence still leaves the most relevant comparison unquantified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SEPT, an extension of online BEV-based scene perception and topology reasoning models that integrates Standard-Definition (SD) map priors. The method contains three components: a hybrid SD map encoding that uses both rasterized and vectorized representations, a Feature Transformation (FT) module for aligning rasterized SD features with BEV features, a Dual Gated Feature Fusion (DGFF) module for combining the two augmented features, and an auxiliary Intersection-Aware Keypoint Detection (IKPD) task supervised by intersection locations extracted from the SD map. Experiments on the OpenLane-V2 subset A validation split are reported on two baselines: TopoNet for the OLS metric and LaneSegNet for the OLUS metric. The paper reports substantial gains, e.g., +5.4 OLS (v1.1) for TopoNet and +5.9 OLUS for LaneSegNet, and component ablations indicate that each proposed module contributes positively over the fixed baselines.","tokens_in":12364,"tokens_out":2531,"duration_ms":27715,"significance":"If the reported results hold, SEPT is a practically relevant contribution: it shows that lightweight, publicly available SD map priors can be injected into existing perception and topology pipelines with consistent improvement across detection, area, traffic element, and topology subtasks. The strength of the paper is its controlled internal evaluation: official baseline implementations are retained, ablations isolate the rasterized/vectorized branches, the FT module, and the fusion strategy, and the parameter overhead is modest. The main limitation is external: the closest concurrent hybrid SD-map method, TopoSD, is cited but never benchmarked, so the headline claim of outperforming existing methods by a substantial margin is not fully supported. In addition, all quantitative claims rest on single-run validation numbers without error bars or significance tests.","major_comments":[{"comment":"The manuscript cites TopoSD [11] as a concurrent approach that also combines rasterized and vectorized SD map streams and then dismisses its fusion strategy as 'overly simplistic', but TopoSD appears nowhere in the quantitative comparisons in Tables I or II. Because TopoSD is the most direct published competitor for SEPT's hybrid SD-map fusion on OpenLane-V2, this omission is load-bearing for the abstract claim of 'outperforming existing methods by a substantial margin'. The authors should add TopoSD to the experimental tables, or if a direct comparison is infeasible, provide a quantitative comparison using TopoSD's published results and clearly discuss the discrepancy.","section":"Section II-C, Tables I and II"},{"comment":"All reported metrics are single-run validation numbers without standard deviations, confidence intervals, or significance tests. With a training batch size of 8 and typical run-to-run variance in this benchmark, margins such as +5.9 OLUS in Table II or +5.4 OLS in Table I may not be distinguishable from noise. The authors should report results over multiple seeds with variance, or at minimum state whether the observed gains are stable across repeated runs.","section":"Section IV-B, Tables I and II"},{"comment":"The fusion weights mu and nu in Eq. (4) are selected on the validation set (Table V), giving best performance at mu = nu = 0.5. Since the same validation split is used for both ablations and final comparisons, the possibility of selection-induced overfitting is not addressed. The authors should either fix the weights a priori, report the sensitivity of the final OLS/OLUS numbers to nearby weight choices, or explicitly justify that the balanced choice is a natural default rather than a tuned hyperparameter.","section":"Section IV-A and Table V"},{"comment":"For the v1.1 metric, several rows in Table I report '-' and the text says that only TopoNet and SMERF v1.1 results are available from official checkpoints. This makes the v1.1 comparison incomplete: the most competitive SD-map-augmented baselines (TopoNet w/ OLV2, OSMG, OSMR) are absent from the ranking metric used in the current benchmark. The authors should either obtain these numbers or clearly restrict their 'substantial margin' claim to the configurations actually compared.","section":"Section IV-B, Table I"}],"minor_comments":[{"comment":"The weights mu and nu in Eq. (4) are introduced without a normalization constraint; please state whether they are constrained to sum to one or are free scalars, and how they are initialized.","section":"Eq. (4), Section III-C3"},{"comment":"The Gaussian radius used to construct intersection keypoint heatmaps is not specified; please report the radius value and any sensitivity of the IKPD task to this choice.","section":"Section III-D1"},{"comment":"The sentence 'without increasing the number of learnable parameters' refers to the sigmoid gating operation, but the surrounding DGFF module includes FFN and projection networks that do add parameters; please rephrase to avoid misunderstanding.","section":"Section III-C3"},{"comment":"The qualitative results show a case where the SD map is outdated and the model correctly prioritizes online perception; please consider adding quantitative analysis of scenarios with outdated SD maps, since this is an important practical robustness property.","section":"Section IV-D"}],"recommendation":"major_revision","confidential_remarks":"The internal ablations are generally well controlled and support the claim that the proposed modules improve over a fixed baseline. The main risk is external: the omission of TopoSD, the closest concurrent method, from all quantitative comparisons weakens the paper's central 'substantial margin' claim. I would also encourage the editor to require multi-seed reporting or at least a clear statement about variance, as the single-run numbers are currently the only evidence for all margins. The novelty overlap with TopoSD should be carefully positioned in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: SEPT is a competently built modular extension that feeds SD maps into an existing BEV perception-and-topology pipeline through hybrid raster/vector fusion plus a keypoint auxiliary loss. The internal evidence is decent; the external case is incomplete because the closest published competitor is cited but never benchmarked.\n\nWhat's actually new: not much conceptually. The FT alignment is a FiLM-style modulation, the cross-attention branch is straight from SMERF, and DGFF is a small gated fusion network. The IKPD task is a simple auxiliary heatmap loss generated from the SD map itself. None of this opens a new technology class. What the paper does well is execution: it keeps the official baselines unchanged, runs component ablations in Table II, validates FT in Table III, compares DGFF against simpler fusions in Table IV, and tests on two different baselines. The gains are consistent and monotonic as modules are added. That is real evidence that the architecture works as claimed.\n\nSoft spots, in proportion. The biggest one is the missing TopoSD comparison. Section II-C calls TopoSD's hybrid fusion \"overly simplistic,\" and TopoSD was on arXiv in November 2024, well before this submission. Yet TopoSD appears in no experimental table. The abstract's \"outperforming existing methods by a substantial margin\" is therefore not yet supported against the one method whose design overlaps most with SEPT's. If TopoSD's published OLV2 numbers are at or above SEPT's, the headline claim fails; if SEPT is better, the absence still leaves the most relevant comparison unquantified. Second, all numbers are single-run validation results, no error bars or multiple seeds. On sub-metrics where the margin is 1-3 points, that matters. Third, no code release, which makes independent verification harder. Minor: the fusion weights mu and nu were selected on the validation split, and since IKPD targets are a deterministic function of the SD map input, the auxiliary task partly measures the network reading the map back. That's a regularizer, not a flaw, but it should be stated.\n\nWho this is for: researchers working on SD-map-enhanced online perception, especially those wanting a clean example of how to fuse raster and vector priors. It's not a breakthrough but it is useful engineering. It deserves a serious referee. I'd send it to review with a mandatory request to add TopoSD to the tables, report seed variance, and release code.","headline":"SEPT is a sound, modular SD-map fusion method with strong internal ablations, but the missing benchmark against the closest competitor, TopoSD, undercuts the headline margin claim.","tokens_in":12951,"tokens_out":1764,"would_cite":true,"duration_ms":19881,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Cheap SD map priors give self-driving scene perception a large, measurable boost.","keywords":["standard-definition maps","bird's-eye-view perception","lane topology reasoning","OpenLane-V2","feature fusion","keypoint detection","autonomous driving","mapless driving"],"falsifier":"Retrain SEPT with the same parameter budget and fusion hyperparameters but with the SD map replaced by random or scrambled geometry; if the OLUS/OLS gains persist, the specific SD map prior content is not the cause. Alternatively, train the baseline with the same number of parameters as SEPT to test the capacity explanation.","tokens_in":11981,"feed_emoji":"🗺️","tokens_out":3588,"duration_ms":33086,"temperature":0.7,"pith_summary":"The paper tries to show that a self-driving system can perceive road structure and reason about lane topology more accurately when it fuses cheap, readily available standard-definition (SD) map data with its camera-based bird's-eye-view features. It reports consistent gains on the OpenLane-V2 benchmark, including a 5.9-point OLUS improvement over the LaneSegNet baseline and a 5.4-point OLS improvement over TopoNet. The value of the claim is that it points a path toward reducing reliance on costly, hard-to-maintain HD maps while still handling long-range and occluded scenes.","feed_headline":"Cheap map data gives self-driving perception a 5.9-point boost","feed_subtitle":"Fusing standard-definition maps with camera features improves scene and topology understanding on OpenLane-V2.","key_machinery":"The load-bearing mechanism is the hybrid SD map encoding plus the alignment and fusion modules: a Feature Transformation (FT) module that applies FiLM-style per-channel scaling and bias to align rasterized SD features with BEV space; a cross-attention network that aligns vectorized SD tokens; and a Dual Gated Feature Fusion (DGFF) network that combines the two aligned streams using sigmoid gates and weighted projections. The auxiliary Intersection-Aware Keypoint Detection (IKPD) head supervises a Gaussian heatmap of SD-derived intersection points, injecting road-topology priors into the shared BEV feature.","core_discovery":"The central discovery is that a hybrid representation of an SD map—using both a rasterized image-like encoding and a vectorized polyline encoding—can be aligned to BEV features and fused so that the two modalities complement each other: rasterization boosts area detection, while vectorization boosts lane segment and topology metrics. Adding an auxiliary intersection-aware keypoint detection head trained on SD-map-derived intersection heatmaps further improves all subtasks. The paper argues that these components, rather than raw capacity, explain the gains.","pith_inferences":["A testable extension: retrain the system with the IKPD head but without any SD map fusion, to isolate how much of the gain comes purely from the topology prior rather than from feature alignment.","The reported gains may vary by city because OpenStreetMap accuracy differs; a per-city breakdown would show where the method helps most and where stale maps hurt.","The same gating and alignment scheme could be applied to other priors, such as previous-frame HD maps or crowdsourced map fragments, which face similar misalignment problems.","The paper does not report inference latency; if the added modules are cheap, this becomes a practical drop-in upgrade for existing BEV pipelines."],"forward_implications":["If SD priors add this much at low cost, mapless driving systems can cover long-range and occluded scenarios without HD maps.","The hybrid fusion pattern (raster plus vector with explicit alignment) generalizes beyond lanes to areas and traffic elements.","Intersection keypoint supervision is a cheap way to inject topology priors into any BEV-based perception model.","The gains appear on both OLS and OLUS metrics and across two different baselines, suggesting the method transfers."],"supporting_citations":[{"why":"Supplies the OpenLane-V2 dataset, task definition, and official OLS/OLUS metrics used in all experiments.","marker":"[5]"},{"why":"TopoNet is the baseline model that SEPT augments for the OLS evaluation.","marker":"[2]"},{"why":"LaneSegNet is the baseline model that SEPT augments for the OLUS evaluation.","marker":"[6]"},{"why":"SMERF provides the vectorized SD-map encoding and cross-attention alignment approach that SEPT builds on.","marker":"[10]"},{"why":"These prior SD-map-augmented methods (OLV2, OSMG, OSMR) are the comparison points for the OLS table.","marker":"[9]"},{"why":"TopoLogic is a recent topology reasoning model on which SEPT is also applied to demonstrate generality.","marker":"[23]"}],"fun_headline_variants":["SD map fusion boosts self-driving perception by 5.9 points","Hybrid SD map fusion sharpens self-driving perception","Intersection-aware keypoints enhance SD map fusion for driving","SD map priors boost scene and topology reasoning in driving","Raster plus vector SD map fusion aids autonomous driving"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The measured gains come from the proposed modules rather than from the extra parameters, the choice of fusion weights, or the quality of OpenStreetMap data in the evaluation subset.","fun_headline_variants_meta":{"raw":{"variants":["SD map fusion boosts self-driving perception by 5.9 points","Hybrid SD map fusion sharpens self-driving perception","Intersection-aware keypoints enhance SD map fusion for driving","SD map priors boost scene and topology reasoning in driving","Raster plus vector SD map fusion aids autonomous driving"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000897,"raw_usage":{"total_tokens":3820,"prompt_tokens":857,"completion_tokens":2963,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":2882}},"tokens_in":473,"tokens_out":2963,"duration_ms":19091,"temperature":1.0,"reasoning_tokens":2882,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:37:47.334516+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain SEPT with the same parameter budget and fusion hyperparameters but with the SD map replaced by random or scrambled geometry; if the OLUS/OLS gains persist, the specific SD map prior content is not the cause. Alternatively, train the baseline with the same number of parameters as SEPT to test the capacity explanation.","supporting_citations":[{"cited_title":"Openlane-v2: A topology reasoning benchmark for unified 3d hd mapping,","cited_arxiv_id":null,"evidence_quote":"Supplies the OpenLane-V2 dataset, task definition, and official OLS/OLUS metrics used in all experiments."},{"cited_title":"Augmenting lane perception and topology understanding with standard definition navigation maps,","cited_arxiv_id":null,"evidence_quote":"SMERF provides the vectorized SD-map encoding and cross-attention alignment approach that SEPT builds on."},{"cited_title":"Enhancing online road network perception and reasoning with standard definition maps,","cited_arxiv_id":null,"evidence_quote":"These prior SD-map-augmented methods (OLV2, OSMG, OSMR) are the comparison points for the OLS table."},{"cited_title":"Topologic: An interpretable pipeline for lane topology reasoning on driving scenes,","cited_arxiv_id":null,"evidence_quote":"TopoLogic is a recent topology reasoning model on which SEPT is also applied to demonstrate generality."}],"review_version":1}