{"id":"5a9a845a-9821-45fd-9e20-29f4415ed110","arxiv_id":"2508.21689","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A probabilistic BEV projection with learned offsets and confidence-based temporal fusion improves online HD map accuracy and generalization.","lead":"This paper proposes a probabilistic projection module that starts from known camera geometry and learns small offsets and confidence scores to turn camera images into bird's-eye-view maps, replacing attention-based projection. It reports better HD map accuracy and fewer hallucinated road elements on nuScenes and Argoverse2, with the largest gains at long range.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'probabilistic' contribution is not isolated: no ablation removes covariance/sampling, so the gains could be entirely from deterministic offset refinement.","rationale":"The reader's weakest assumption (single Gaussian cannot capture multimodal correspondence) is a plausible limitation, but it is not the most load-bearing concern for the paper's central claim. Even if the Gaussian is misspecified, the method could still outperform SOTA by learning good offsets and confidences. The more decisive gap is internal: the paper never ablates the probabilistic components (covariance, sampling, K) from the deterministic offset-refinement components. The ablation in Table 2 moves from static projection (A) to offset+covariance (B1), but does not decompose them. Without such an ablation, the headline claim that 'probabilistic projection' is responsible for the improvements is underdetermined. This concern does not overturn the empirical results, and the paper otherwise provides a plausible, well-ablated framework with code/checkpoints promised, so the reader's CONDITIONAL verdict stands. I would keep the verdict CONDITIONAL, adding the missing decomposition as an explicit condition for full acceptance.","tokens_in":9846,"tokens_out":9117,"duration_ms":114164,"concrete_test":"Run a nuScenes short-range ablation with (1) learned foffset only, covariance fixed to a small isotropic value (e.g., 0.1·I), K=1, no confidence weighting; (2) same but with learned covariance and K=8; and (3) the full model but replacing sampled µ_i,k with the deterministic mean µ_i at inference. If (1) matches the full model's mAP (Table 2, row E: 45.4) within a single-run tolerance, then the probabilistic sampling/covariance is inert and the novelty claim should be reframed as deterministic offset refinement. If (3) also matches, stochasticity plays no role at inference, further weakening the 'probabilistic' label.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that probabilistic projection (Gaussian sampling of K locations with learned covariance and confidence weighting, Eq. 6) is what drives the reported improvements over attention-based projection. But Table 2 only compares static mapping (A) to the full offset+covariance model (B1); there is no ablation with learned offsets but fixed/identity covariance, no ablation with K=1 vs K>1, and no ablation replacing stochastic sampling with the deterministic mean at inference. If the learned covariance collapses toward zero or the K samples effectively concentrate at the mean, Eq. 6 degenerates to deterministic residual offset refinement, which is a much weaker claim than 'probabilistic projection.' The paper's own Limitations section concedes that occluded regions are handled poorly, consistent with the Gaussian being a weak model of true correspondence uncertainty, but the more immediate problem is attribution: the nuScenes gains (+12.7% short-range mAP, +18.2% long-range) are not shown to require the probabilistic machinery at all. Without isolating covariance/sampling, the strongest claim 'probabilistic projection significantly outperforms SOTA' is not yet supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes replacing the attention-based BEV projection used in online HD map estimation with a geometry-initialized, scene-adaptive probabilistic projection. Starting from a static camera-parameter pull mapping μ_pull, a lightweight CNN predicts per-pixel offsets and per-pixel covariance/confidence; for each BEV cell a Gaussian N(μ_i, Σ_i) is formed, K locations are sampled from it, and image features at those locations are pooled with likelihood/confidence weights (Eq. 6). Per-camera raw BEV features are averaged, then confidence-weighted temporal fusion with historical raw features and a MapTracker-style memory module produces the final BEV features. Experiments on new non-overlapping geographical splits of nuScenes and Argoverse2 report consistent mAP/C-mAP improvements over StreamMapNet and MapTracker, with the largest gains on nuScenes and at long range. Code and checkpoints are released.","tokens_in":10139,"tokens_out":5649,"duration_ms":66330,"significance":"If confirmed, the paper would make a useful empirical contribution: it shows that a geometric prior from camera extrinsics, combined with residual offset refinement and confidence-gated temporal accumulation, can improve generalization and reduce false positives relative to learned attention-based projection. The use of new geographical splits is a genuine generalization test, and the ablation structure is logical: A→B1 isolates offset/covariance refinement, C1→C2 isolates confidence-gated temporal fusion, and D→E isolates confidence weighting in history fusion. The release of code and checkpoints is a practical strength. However, the central attribution of the gains to probabilistic sampling is not yet supported: the ablations do not separate deterministic offset refinement from Gaussian sampling, and the uncertainty model itself is not evaluated. The reported comparisons also rest on single-run numbers with no variance estimates, and on Argoverse2 some metrics are actually lower than MapTracker. These issues make the central claim stronger than the current evidence.","major_comments":[{"comment":"The reported gain from A to B1 (41.0→43.5 mAP) conflates three mechanisms: the learned residual offset in Eqs. (3)–(4), the predicted covariance in Eq. (5), and stochastic K-sample pooling in Eq. (6). No row in Table 2 removes the Gaussian covariance while keeping the offset, no row fixes K=1 with deterministic sampling at the mean, and no row compares K=1 vs K>1. If the learned Σ collapses or the K samples concentrate at μ_i, Eq. (6) degenerates to deterministic bilinear sampling at a refined location, and the abstract's claim that 'probabilistic projection' drives the improvement is unsupported. Please add ablations for (i) deterministic sampling at μ_i without Σ, (ii) learned Σ but K=1/deterministic mean, and (iii) fixed isotropic Σ with K>1, and report learned covariance statistics to demonstrate that the stochastic component is actually active.","section":"§3.1 Eq. (6); §4.3 Table 2"},{"comment":"All results appear to be from a single run, and no error bars or significance tests are reported. On Argoverse2 the short-range boundary AP is below MapTracker (67.5 vs 68.9) and short-range C-mAP is below MapTracker (62.7 vs 63.2); the long-range C-mAP gain is only 0.7 points. Without variance estimates, 'significantly outperforms' across 'nearly all metrics' is too strong. Please provide mean/std over multiple seeds or a paired significance test, and clarify exactly which Table 1 entries are re-runs versus numbers taken from prior papers.","section":"§4, Table 1"},{"comment":"The term 'probabilistic' is not validated as uncertainty modeling. The covariance Σ and confidence α are trained only through the downstream map loss; no calibration, entropy, or failure-mode analysis of the Gaussian is presented, and the Limitations paragraph concedes that occluded regions are handled poorly. If the contribution is intended to be probabilistic projection rather than merely offset-refined sampling, the revision should evaluate the uncertainty estimates (e.g., covariance behavior in occluded vs visible cells, or calibration of confidence) or soften the probabilistic framing.","section":"§3.1, Fig. 2, Limitations"}],"minor_comments":[{"comment":"The equation uses Braw on both sides without distinguishing the current and updated variables; make the warping and update explicit with time subscripts or a separate update operator.","section":"§3.2, Eq. (8)"},{"comment":"The text says the Gaussian likelihood is normalized by the sum of K likelihoods, but the displayed equation omits this normalization; align the formula with the text.","section":"§3.1, Eq. (6)"},{"comment":"The sentence 'Removing confidence scores has the least impact on mAP (D)' is confusing because D vs E drops 2.0 mAP, which is comparable to the other ablated components; clarify the intended comparison.","section":"§4.3, Table 2 discussion"},{"comment":"C-mAP is used as a headline metric but is never defined; provide the formula or a precise citation.","section":"§4, Experimental Setup"},{"comment":"The note about image-space vs BEV-space notation for μ_pull and μ is difficult to parse; consider simplifying the notation or moving the clarification to the main text.","section":"Fig. 2 caption"}],"recommendation":"major_revision","confidential_remarks":"I agree with the reader's conditional assessment. The paper is technically solid and the split-based evaluation is a strength, but the attribution of the gains to probabilistic sampling is the load-bearing point and is currently under-supported. This is fixable with additional ablations and uncertainty diagnostics, so I recommend major revision rather than rejection. Please also ask for variance estimates given the weak Argoverse2 margins."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this paper on probabilistic BEV projection for HD mapping. The idea is simple: instead of attention-based projection (BEVFormer-style) used by MapTracker, they start from known camera geometry, learn per-pixel offsets, and add a Gaussian sampling step with confidence weights. They also use confidences to merge historical BEV features temporally. That's a real, useful direction, and the paper shows consistent gains on nuScenes and long-range Argoverse2, with fewer false positives. The ablation in Table 2 is genuinely informative: replacing attention with static geometric pull already beats MapTracker slightly (41.0 vs 40.3), and learning offsets adds +2.5 mAP. Code and checkpoints are promised.\n\nNow the soft spots. The biggest one is attribution. The \"probabilistic\" part—covariance, K samples, stochastic sampling—is never isolated. Table 2 only compares static mapping (A) to the full offset+covariance model (B1). There's no fixed-covariance or K=1 ablation, and no test of deterministic mean vs sampling at inference. So we don't know if the gains come from the Gaussian machinery or just from learned deterministic residual offsets. The stress-test concern lands: the paper's strongest claim, that probabilistic projection is what drives the improvements, is not yet supported by the evidence as reported. That's not fatal—the offsets alone are still a contribution—but the framing overreaches.\n\nSecond, the experimental comparisons are thin. Only two baselines, one taken as reported, and no error bars. On Argoverse2 the method slips on boundary AP and C-mAP in the short range, which the paper acknowledges but doesn't stress. The improvements on nuScenes are large, but a single run leaves some uncertainty about stability.\n\nThird, the single-Gaussian assumption for correspondence uncertainty is a reasonable first cut but likely weak in occluded or repetitive-texture areas; the authors themselves note occluded regions are handled poorly.\n\nWho is this for? People working on online HD mapping and BEV projection will get value. It deserves a serious referee: the direction is promising, the experiments are mostly clean, and the open questions are addressable with a few additional ablations and maybe repeated runs. I'd want the authors to do the K=1 / fixed-covariance ablation and clarify whether probabilistic sampling is actually load-bearing. If they can show that, this is a solid paper. If they can't, the claim needs to be scaled back to \"learned residual offsets plus confidence-gated temporal fusion.\" Either way, worth engaging.\n\nRecommendation: send to peer review, require the missing ablation before acceptance.","headline":"A useful BEV projection idea with a real attribution gap: the paper never isolates the probabilistic sampling from the learned offsets, so the central claim overreaches even though the direction is sound.","tokens_in":10553,"tokens_out":2617,"would_cite":true,"duration_ms":28631,"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":"Replacing attention-based projection with a geometry-initialized, confidence-gated probabilistic sampler yields more accurate online HD maps with fewer hallucinated road elements, especially at long range.","keywords":["online HD mapping","bird's-eye view projection","probabilistic sampling","camera-to-BEV transform","temporal fusion","autonomous driving","generalization","confidence gating"],"falsifier":"Run the model on a sequence where a lane line is split by occlusion into two separated image regions: a single Gaussian per cell can sample only near one mode, so if the predicted map shows a gap or collapsed confidence where an attention baseline recovers both segments, the unimodal assumption is the binding limit. Separately, evaluating on the original overlapping nuScenes split would reveal how much of the reported margin depends on the region-disjoint evaluation the paper adopts.","tokens_in":9799,"feed_emoji":"🗺️","tokens_out":9370,"duration_ms":105935,"temperature":0.7,"pith_summary":"Online HD mapping for self-driving cars needs to move image evidence onto a bird's-eye-view grid, and this paper's claim is that the usual way of doing that — learned, attention-based projection — is the main source of error: it memorizes where road elements tend to appear and invents them at the wrong place. The authors propose to start from the explicit geometric mapping given by camera parameters and refine it scene-by-scene with a probabilistic sampler: each grid cell learns a small Gaussian over where in the image to pull features, and a confidence score per pixel filters out irrelevant content like pedestrians and walls. The same confidences gate temporal fusion, so only reliable historical features accumulate. On geographically disjoint splits of nuScenes and Argoverse2 the method beats the current best baselines on nearly every metric, with the biggest gains in the long perception range, which is why a reader should care: accuracy and generalization to unseen areas are precisely what online mapping has been missing.","feed_headline":"Up to 18% better HD maps: probabilistic projection beats attention","feed_subtitle":"Geometry-initialized, confidence-gated sampling lifts accuracy and cuts hallucinated road elements on two driving benchmarks.","key_machinery":"The central object is the probabilistic projection module. It starts from a static pull mapping defined by camera intrinsics and extrinsics, then predicts per-pixel offset distributions and covariances with small CNN heads — foffset receives the camera features plus a distance mask, and fconf predicts covariance and confidence. For each BEV grid cell i, the refined parameters define a Gaussian N(µ_i, Σ_i) over image sampling locations; K locations are sampled, and the cell's feature is built as a weighted sum of bilinearly sampled features, with each weight equal to the sampled pixel's confidence times the Gaussian likelihood of the sample. A second use of the same confidence scores implemen","core_discovery":"The paper establishes that the camera-to-BEV projection — not the vector-map head — is where online HD mapping loses accuracy. Replacing the deformable-attention projection inside the MapTracker pipeline with a plain geometric pull mapping from camera intrinsics and extrinsics already matches or slightly beats the original, which the authors read as evidence that attention's learned geometry memorizes the training distribution and hallucinates road elements where they usually appear. Their method then makes the projection probabilistic: a static BEV-to-image correspondence is refined by learned offsets, each BEV cell draws K samples from the resulting Gaussian N(µ_i, Σ_i), and the cell's fea","pith_inferences":["A natural extension the authors leave open: replacing the single Gaussian per cell with a mixture or a sampled heatmap could recover occluded features that are split across disjoint image regions.","The confidence map is an interpretable by-product; a planner could consume it as per-cell evidence quality and treat low-confidence areas as untrusted before acting on them.","The static-pull baseline's surprisingly strong showing suggests other BEV lifting tasks (object detection, segmentation) may also be over-relying on learned projection when camera geometry is available.","A hybrid that keeps geometric sampling for most cells but activates context-aware attention only where confidence is low could address occlusion while preserving the geometry anchor."],"forward_implications":["If the claim holds, attention-based BEV projection is not required for accurate online HD mapping; a geometry-seeded, scene-adapted sampler is more accurate and generalizes to unseen regions.","Long-range mapping (100 m × 50 m) benefits most, so the approach could extend effective perception range without adding sensors.","Confidence-gated temporal accumulation improves consistency-aware performance on nuScenes, implying the gating signal is a reusable mechanism for streaming perception.","Fewer false positives near junctions and crossings means downstream planning receives a cleaner map, not just a higher mAP.","The gains are dataset-dependent (large on nuScenes, modest on Argoverse2, slightly negative for the boundary class at short range on Argoverse2), so the method's advantage is strongest where attention baselines struggle."],"supporting_citations":[{"why":"The strong baseline the paper replaces and builds on: supplies the vector map head, memory mechanism, and training pipeline, with its projection swapped for the probabilistic one.","marker":"[4]"},{"why":"Supplies the static bilinear-sampling pull mapping from camera parameters that seeds the probabilistic projection and already matches the attention baseline.","marker":"[9]"},{"why":"The deformable attention projection used by MapTracker that the paper removes; its tendency to hallucinate road elements is the failure mode the method targets.","marker":"[15]"},{"why":"Contributes the geographically disjoint nuScenes and Argoverse2 splits used to test generalization, plus a baseline its results are compared with.","marker":"[29]"},{"why":"Primary evaluation dataset, where the largest improvements (up to 18.2% mAP at long range) are reported.","marker":"[1]"},{"why":"Second evaluation dataset, showing the gains transfer to a different sensor setup, with smaller margins.","marker":"[27]"}],"fun_headline_variants":["Skeptic mapping: geometry beats learned attention for HD maps","Probabilistic projection reduces HD map hallucinations","Up to 18% better HD maps with probabilistic BEV projection","Confidence-scored mapping improves HD map generalization","Skeptic's map: probabilistic BEV projection outperforms attention"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Each BEV cell's true image source is assumed to be single-peaked — one Gaussian per cell — so K samples from that Gaussian can pull the right features; where the true source is split across separate image regions (occlusion, repeated textures) the model cannot represent it, which is why occluded targets remain the paper's admitted limitation.","fun_headline_variants_meta":{"raw":{"variants":["Skeptic mapping: geometry beats learned attention for HD maps","Probabilistic projection reduces HD map hallucinations","Up to 18% better HD maps with probabilistic BEV projection","Confidence-scored mapping improves HD map generalization","Skeptic's map: probabilistic BEV projection outperforms attention"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000589,"raw_usage":{"total_tokens":2602,"prompt_tokens":748,"completion_tokens":1854,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":1773}},"tokens_in":492,"tokens_out":1854,"duration_ms":15501,"temperature":1.0,"reasoning_tokens":1773,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:01:04.824204+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the model on a sequence where a lane line is split by occlusion into two separated image regions: a single Gaussian per cell can sample only near one mode, so if the predicted map shows a gap or collapsed confidence where an attention baseline recovers both segments, the unimodal assumption is the binding limit. Separately, evaluating on the original overlapping nuScenes split would reveal how much of the reported margin depends on the region-disjoint evaluation the paper adopts.","supporting_citations":[{"cited_title":"Maptracker: Tracking with strided memory fusion for consistent vector hd mapping","cited_arxiv_id":null,"evidence_quote":"The strong baseline the paper replaces and builds on: supplies the vector map head, memory mechanism, and training pipeline, with its projection swapped for the probabilistic one."},{"cited_title":"Simple-bev: What really matters for multi-sensor bev perception? In ICRA, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the static bilinear-sampling pull mapping from camera parameters that seeds the probabilistic projection and already matches the attention baseline."},{"cited_title":"Bevformer: Learning bird’s-eye-view representation from multi- camera images via spatiotemporal transformers","cited_arxiv_id":null,"evidence_quote":"The deformable attention projection used by MapTracker that the paper removes; its tendency to hallucinate road elements is the failure mode the method targets."},{"cited_title":"Streammapnet: Streaming mapping network for vectorized online hd map construction","cited_arxiv_id":null,"evidence_quote":"Contributes the geographically disjoint nuScenes and Argoverse2 splits used to test generalization, plus a baseline its results are compared with."}],"review_version":1}