{"id":"e566ecf3-f291-4076-b137-fa046b797ceb","arxiv_id":"2506.13553","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"RelTopo embeds geometric relationships into three levels of a driving-scene model (perception, reasoning, supervision) and reports state-of-the-art lane and topology reasoning scores on OpenLane-V2.","lead":"RelTopo is a system that improves how self-driving cars understand lane connections and traffic-signal relationships by building geometric cues into perception, reasoning, and training. It reports large gains on the OpenLane-V2 benchmark, including a +4.4 overall score improvement, claiming a new state of the art.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on Table 1 baseline numbers whose metric version and protocol are not established; TopoLogic's results are sourced from their paper without evidence of recomputation under V2.1.0.","rationale":"The reader identified baseline comparability as the weakest assumption, and I agree: the SOTA claim depends on Table 1, and the footnote admits that RoadPainter used old metrics while the remaining non-recomputed numbers came from the TopoLogic paper. The paper states it follows V2.1.0 but does not show that prior numbers were recomputed under that version. The internal evidence for the relational modeling idea is reasonably strong: Table 2 shows cumulative gains from geometry-biased self-attention, curve-guided cross-attention, the L2L and L2T heads, and the contrastive loss, and Table 5 shows consistent improvements across backbones. So the architectural claim is plausible, but the headline SOTA claim is not yet verified against comparable numbers. The missing Topo2D baseline reinforces the concern because Topo2D is directly relevant to L2T reasoning and would be a natural competitor. I also note minor inconsistencies: the abstract in the provided text names the method CoPo while the full text uses RelTopo, and Eq. 9 is not a standard multi-positive InfoNCE form, but neither is as load-bearing as the baseline comparison. The existing CONDITIONAL verdict already captures the needed revision, so I recommend no change.","tokens_in":16786,"tokens_out":5075,"duration_ms":51669,"concrete_test":"Run the official TopoLogic and Topo2D checkpoints (and TopoMLP if available) on the OpenLane-V2 subset A and subset B test splits using the exact V2.1.0 evaluation script referenced in Section 4.1, with the same matching thresholds and score aggregation, and recompute DET_l, TOP_ll, TOP_lt, and OLS. If TopoLogic's TOP_ll or TOP_lt changes by more than about 1 point relative to the values in Table 1, then the headline gains are inflated and Table 1 must be regenerated; Topo2D's V2.1.0 numbers should also be added to the comparison.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's externally communicated claim is that RelTopo sets a new state-of-the-art on OpenLane-V2 with gains of +3.1 DET_l, +5.3 TOP_ll, +4.9 TOP_lt, and +4.4 OLS. That claim rests entirely on Table 1, and Table 1's comparability is not established. The header says all results use V2.1.0, but the footnote discloses that RoadPainter used old metrics and that 'other results were sourced from the TopoLogic paper' without stating whether TopoLogic's numbers were recomputed with the V2.1.0 evaluation code. TopoLogic and TopoMLP are the closest prior systems (subset A: TopoLogic 44.1, TopoMLP 44.5, ours 48.9; subset B: TopoMLP 43.6, ours 49.7), so even a small metric-version shift in their TOP_ll or TOP_lt could materially reduce the claimed margin. Additionally, Topo2D, a method specifically targeting L2T with 2D lane priors, is absent from Table 1, so the 'new state-of-the-art' statement is incomplete even under a single metric version. The internal ablations in Table 2 and the backbone generalization study in Table 5 support the value of the relational modules, but they do not repair the external comparison because they use only the authors' own baseline. Thus the central SOTA claim is conditional on an external reproducibility check that the paper does not report.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RelTopo, a unified framework for lane detection and road topology reasoning on the OpenLane-V2 benchmark. The method integrates relational modeling at three levels: perception (geometry-biased self-attention and curve-guided cross-attention in a Bezier-based lane decoder), reasoning (a geometry-enhanced L2L head and a cross-view L2T head), and supervision (a contrastive InfoNCE objective with hard-negative mining). The authors report state-of-the-art results on both subA and subB of OpenLane-V2, with gains of +3.1 DET_l, +5.3 TOP_ll, +4.9 TOP_lt, and +4.4 OLS on subA, and larger gains on subB. The manuscript also includes ablations, comparisons with alternative Bezier representations, and a backbone-generalization study.","tokens_in":17205,"tokens_out":3231,"duration_ms":30786,"significance":"If the reported results are reproducible and the comparisons are fair, the multi-level relational modeling idea is a solid contribution to driving-scene topology reasoning. The paper's internal ablations (Table 2) show monotonic improvements when adding each module, which supports the claim that relational perception, reasoning, and supervision complement each other. The appendix provides useful additional experiments, including a controlled comparison with TopoLogic's geometric distance topology (Table 3), alternative Bezier variants (Table 4), and consistent gains across ResNet-50, ResNet-101, and Swin-Base (Table 5). The authors state that code will be released, which would further strengthen reproducibility. However, the headline 'state-of-the-art' claim depends on the comparability of Table 1 baselines, which is not fully established.","major_comments":[{"comment":"The state-of-the-art claim rests on Table 1, but the comparability of the baseline numbers is not established. The table header states that all results use the V2.1.0 metric, yet the footnote discloses that RoadPainter‡ used the old metrics and that 'other results were sourced from the TopoLogic paper' without stating whether TopoLogic, TopoNet, MapTR, and VectorMapNet numbers were recomputed with the V2.1.0 evaluation code. Since TopoLogic (44.1 OLS) and TopoMLP (44.5 OLS) are the closest baselines on subA, a small metric-version shift in TOP_ll or TOP_lt could materially reduce the claimed +5.3 and +4.9 gains. Please clarify the exact evaluation protocol for each baseline and, if applicable, rerun the official V2.1.0 evaluation for all methods.","section":"Table 1 and Section 4.3"},{"comment":"Topo2D (Li et al., 2024), a method explicitly designed for L2T reasoning with 2D lane priors, is discussed in the related work but is absent from Table 1. Given that one of the paper's claimed strengths is +4.9/+6.0 TOP_lt gains, omitting the most directly competing L2T method makes the 'new state-of-the-art' statement incomplete even under a single metric version. Please include Topo2D in the comparison or justify its exclusion.","section":"Table 1 (missing baseline)"},{"comment":"All experimental comparisons, including the central ablations in Table 2 and the GDT comparison in Appendix Table 3, are reported without error bars or multiple seeds. Differences of 0.6-1.7 points in TOP_ll and TOP_lt could plausibly arise from training variance, especially given the DET_t fluctuations the authors themselves analyze in Appendix A.3.5. Please report the mean and standard deviation over at least three seeds for the main ablations, or explicitly justify the stability of single-run results.","section":"Table 2 and Sections 4.4, A.3"},{"comment":"The ablation baseline (#1) is the authors' own reimplementation of a deformable-DETR decoder with lightweight MLP heads, said to follow Wu et al. (2023). The baseline DET_l is 27.7, whereas Table 1 lists TopoMLP's DET_l as 28.5. This discrepancy raises the question of whether the baseline faithfully reproduces TopoMLP or differs in implementation details. Please state whether the ablation baseline matches the published TopoMLP numbers, and discuss how any implementation gap affects the attribution of gains to the proposed modules.","section":"Section 4.4 and Appendix A.2.1"}],"minor_comments":[{"comment":"The model is called 'CoPo' in the abstract but 'RelTopo' in the title, introduction, and the rest of the paper. Please unify the name.","section":"Abstract and Title"},{"comment":"The text repeatedly says 'OSL' (e.g., 'RelTopo achieves an OSL of 48.9') but the metric is abbreviated OLS in Eq. (10) and Table 1. Please correct the typo.","section":"Section 4.3 and Table 1"},{"comment":"The caption text is garbled: 'associations between lanes , and among lanes and associations between lanes and traffic elements (L2T)' repeats phrases and the inline symbols are missing. Please rewrite the caption to clearly describe the pipeline, including the definitions of Q_lane, Q_te, P_lane, and P_te.","section":"Figure 2 caption"},{"comment":"The multi-positive extension of InfoNCE is written as log[1 + sum_{v+} sum_{v-} exp(v- - v+)]. This treats each positive-negative pair independently rather than the standard multi-positive softmax. Please clarify the derivation and cite the exact source; if this is intentional, explain how it relates to the usual multi-positive contrastive loss.","section":"Equation (9)"},{"comment":"The name 'Topologic' is used inconsistently alongside 'TopoLogic' (e.g., in Sections 3.2.1 and 3.3.1). Please use the official spelling 'TopoLogic' consistently.","section":"Throughout"},{"comment":"The table title contains a typo: 'Comparsion' should be 'Comparison'.","section":"Appendix Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper's central architectural idea is credible, and the internal ablations suggest the relational modules do help. My main concern is the external comparability of Table 1, which is the basis for the abstract's SOTA claim. The authors should be asked to either recompute all baselines with the V2.1.0 metric or clearly state which numbers were not recomputed, and to include Topo2D. I also recommend requiring error bars for the main ablations. These are fixable within the scope of a revision, so I do not recommend rejection. I would also note that the self-citation of Luo et al. (2023) is appropriate given the method builds on LATR-style Bezier lane detection, but no other citation concerns arose."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the method is a sensible and well-ablated integration of relational cues into lane detection and topology reasoning, and the internal experiments support the value of each added module. The headline claim of a new state of the art, however, depends on Table 1 baseline numbers whose evaluation conditions are not fully pinned down. I'd send it to a serious referee but ask for baseline recomputation and a few missing controls.\n\nWhat's actually new: the three-level framing (geometry-biased self-attention and curve-guided cross-attention at perception, geometry-enhanced L2L and cross-view L2T heads at reasoning, and a contrastive loss at supervision) is a coherent package, not just a single trick. The cross-view L2T head, which projects 3D lane points into the front view and samples features there, looks like a genuine fix for the BEV-FV misalignment that prior methods either ignore or patch with a separate 2D decoder. The ablation table shows monotonic gains as modules are added, and the supplementary comparisons against TopoLogic's GDT and against BezierFormer/BeMapNet strengthen the case that the design choices matter.\n\nSoft spots: the biggest one is Table 1. The footnote says RoadPainter used old metrics and other results came from TopoLogic's paper, without saying whether those numbers were recomputed under the V2.1.0 metric. Since TopoLogic and TopoMLP are the closest competitors, even a small metric-version shift in their TOP scores could shrink the claimed +5.3/+4.9 gains. The missing Topo2D baseline is also a problem given that Topo2D specifically targets L2T with 2D priors. Beyond the comparison: no error bars or multiple runs, code isn't available yet, and the multi-positive contrastive loss in Eq. (9) looks non-standard — I'd want the authors to clarify the derivation and show it actually behaves like InfoNCE. These are all fixable in revision.\n\nBottom line: the relational-modelling story is credible and the ablations are honest, but I wouldn't bank on the absolute numbers until the baseline issue is resolved. If you work on driving topology, it's worth a close read; if you're outside the subfield, the architecture is a useful reference but not a transformative result. I'd accept it for peer review with a request for recomputed baselines, a Topo2D comparison, and variance reporting.","headline":"A well-ablated multi-level relational model for lane topology, but the 'new SOTA' headline rests on baseline numbers whose metric version isn't established, so treat the absolute gains as provisional.","tokens_in":17696,"tokens_out":3097,"would_cite":false,"duration_ms":26700,"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":"RelTopo shows that embedding geometric relations into lane perception, topology heads, and contrastive supervision jointly raises detection and topology accuracy, reporting new state-of-the-art scores on OpenLane-V2.","keywords":["3D lane detection","topology reasoning","lane-to-lane connectivity","lane-to-traffic-element association","Bézier curve representation","geometry-biased attention","contrastive InfoNCE supervision","OpenLane-V2"],"falsifier":"Re-run the official OpenLane-V2 V2.1.0 evaluation on the released TopoLogic model and on RelTopo under identical training data and schedule; if the TOP_ll gap on subset A is not near +5.3 (and TOP_lt not near +4.9), the headline comparison is wrong even though the paper's internal ablations could still hold.","tokens_in":16616,"feed_emoji":"🛣️","tokens_out":7902,"duration_ms":62929,"temperature":0.7,"pith_summary":"The paper argues that road-topology reasoning fails when lane detection, lane-to-lane connectivity, and lane-to-traffic-element association are optimized separately, because none of those stages exploits the geometric structure of roads. Its proposed system, RelTopo, embeds relational cues at three levels: lane features are built with geometry-biased attention and curve-guided Bézier sampling; topology heads fuse geometry and cross-view features; and a contrastive loss pulls connected lane/traffic embeddings together. On the OpenLane-V2 benchmark, the paper reports consistent gains over prior methods, notably +5.3 in lane-to-lane topology and +4.9 in lane-to-traffic topology on subset A, with a new state-of-the-art overall score. If the comparison is fair, the result strengthens the case that perception and topology reasoning should be trained jointly rather than as a two-stage pipeline.","feed_headline":"Geometry-guided reasoning lifts lane topology scores by 5.3","feed_subtitle":"A three-level relational model ties lane detection, lane-to-lane and lane-to-traffic reasoning into one end-to-end pipeline on OpenLane-V2.","key_machinery":"The load-bearing object is the relation-aware lane decoder built on a cubic Bézier representation. Geometry-biased self-attention adds an MLP-embedded bias, computed from each lane pair's minimum end-to-start distance and orientation difference, to the attention logits; curve-guided cross-attention samples K=11 points along each Bézier curve and uses a single shared lane query to predict offsets and weights, so context is aggregated along the lane's shape rather than at sparse control points. This decoder feeds a geometry-enhanced L2L head, which broadcasts predecessor and successor query embeddings with sinusoidal positional encodings and adds a learned distance embedding of endpoint separation, and a cross-view L2T head, which projects 3D lane points into the front view, samples features by grid sampling, and forms pairwise lane-traffic embeddings. A contrastive InfoNCE loss with hard-negative mining regularizes the pairwise embeddings at supervision time.","core_discovery":"On the paper's own terms, the discovery is that relational modeling is not a post-processing step but a unifying inductive bias: injecting inter-lane geometry (distance and angle) directly into attention biases, sampling features along Bézier-curve lane representations, embedding endpoint distances into pairwise L2L logits, aligning BEV lane features to front-view traffic elements through projected geometry, and regularizing relational embeddings with a symmetric InfoNCE objective all cooperate. The paper demonstrates this with ablations: geometry-biased self-attention alone raises TOP_ll by +2.9, curve-guided cross-attention raises DET_l by +4.6, and the full stack reaches 48.9 OLS on subset A and 49.7 on subset B, outperforming tested baselines in every metric.","pith_inferences":["Editorial inference: the same three-level relational recipe could transfer to other pairwise map-reasoning tasks, such as lane-segment graphs or intersection topology, where distance/angle attention biases and contrastive relational supervision apply without the OpenLane-V2 task split.","Editorial inference: the size of the headline margin depends on metric-version alignment; if TopoLogic's published numbers come from an earlier evaluation, re-running both under V2.1.0 could narrow the +5.3 and +4.9 gains even though the internal ablations would still hold.","Editorial inference: since DET_t fluctuates across ablations while lane metrics rise, the design appears to move capacity between tasks; a gradient-balancing variant might stabilize traffic-element detection further."],"forward_implications":["Gains appear in both perception and topology at once (DET_l +3.1, TOP_ll +5.3, TOP_lt +4.9 on subset A), supporting the paper's mutual-enhancement claim rather than a trade-off.","Topology predictions no longer depend on inference-time geometric post-processing; ensembling the paper's model with TopoLogic's geometric distance scoring actually lowers its TOP_ll by 0.5.","Lane-to-traffic association improves by aligning BEV lane features to front-view traffic features via projected 3D geometry, avoiding a second 2D lane decoder.","The relational modules are encoder-agnostic, improving OLS consistently with ResNet-50, ResNet-101, and Swin-Base backbones."],"supporting_citations":[{"why":"Supplies the OpenLane-V2 benchmark, the L2L and L2T task definitions, and the DET_l/TOP_ll/TOP_lt/OLS metrics on which all claims are measured.","marker":"Wang et al. (2024)"},{"why":"The main state-of-the-art baseline; its geometric distance topology (GDT) and inference-time post-processing are the direct contrast for the paper's learnable geometry embeddings.","marker":"Fu et al. (2024)"},{"why":"Provides the TopoMLP baseline and the default decoder and head configuration (300 lane queries, 100 traffic queries, MLP heads) that RelTopo extends.","marker":"Wu et al. (2023)"},{"why":"Supplies the BEVFormer encoder that generates the 100x200 BEV feature map used by the lane branch.","marker":"Li et al. (2022c)"},{"why":"Provides the deformable attention mechanism that curve-guided cross-attention extends with Bézier-sampled reference points.","marker":"Zhu et al. (2020)"},{"why":"Supplies the InfoNCE contrastive objective that the paper's relational supervision loss is based on.","marker":"Oord et al. (2018)"},{"why":"Supplies the explicit position-relation prior idea that geometry-biased self-attention adapts to lane pairs.","marker":"Hou et al. (2025)"},{"why":"The BézierFormer comparison that motivates single-query curve-guided cross-attention instead of per-point grid sampling.","marker":"Dong et al. (2024)"}],"fun_headline_variants":["Lane topology +5.3 with geometry-guided CoPo","CoPo: geometry-guided relational model for driving scenes","Geometry in attention lifts lane and traffic topology","CoPo ties lane perception and topology with geometry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline state-of-the-art claim assumes every baseline was evaluated under the same metric version (V2.1.0) and training protocol as RelTopo; the paper notes RoadPainter used old metrics and lists other baseline scores as sourced from the TopoLogic paper rather than recomputed.","fun_headline_variants_meta":{"raw":{"variants":["Lane topology +5.3 with geometry-guided CoPo","CoPo: geometry-guided relational model for driving scenes","Geometry in attention lifts lane and traffic topology","CoPo ties lane perception and topology with geometry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1523,"prompt_tokens":1034,"completion_tokens":489,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":650,"completion_tokens_details":{"reasoning_tokens":427}},"tokens_in":650,"tokens_out":489,"duration_ms":5265,"temperature":1.0,"reasoning_tokens":427,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:58:47.509313+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the official OpenLane-V2 V2.1.0 evaluation on the released TopoLogic model and on RelTopo under identical training data and schedule; if the TOP_ll gap on subset A is not near +5.3 (and TOP_lt not near +4.9), the headline comparison is wrong even though the paper's internal ablations could still hold.","supporting_citations":[],"review_version":2}