{"id":"64ac183f-58d3-4d1b-83f7-eeb32c50f775","arxiv_id":"2507.19691","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"Co-Win proposes a window-attention BEV architecture for joint LiDAR vehicle detection and instance segmentation, but the claimed state-of-the-art performance is presented without any experimental data.","lead":"Co-Win is a LiDAR perception system that detects vehicles and outlines their ground footprints in a top-down view using a windowed transformer. The paper claims major accuracy gains on KITTI, Waymo, and SemanticKITTI, but reports no quantitative results, tables, or ablation numbers to back that claim.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No experimental results exist: §4.3/§4.4 are empty placeholders, so the SOTA claim is unsupported; moreover §3.2's complexity formula describes standard windowed attention, not the stated ELU+1 linear attention.","rationale":"The reader's rejection is correct. I reinforce it with two specifics: all experimental sections are empty placeholders, and the one quantitative efficiency argument is mis-derived. The linear-attention representational-power concern is real but secondary because without any experiments the architecture's accuracy is entirely unassessed. The complexity error is independently verifiable from the equations in §3.2: the formula K·O(M^4)=O(HW·M^2) is standard window attention, not linear attention; with the stated ELU+1 kernel the total should be O(HW d^2). This matters because the paper's claimed 400× reduction is used to motivate the architecture. No formal verification, code, or parameter-free derivation exists, so the central claim of state-of-the-art performance remains unsupported. The verdict should remain as the reader's rejection; nothing in my read changes that outcome.","tokens_in":9534,"tokens_out":5781,"duration_ms":64654,"concrete_test":"Recompute the SPCN total cost under the paper's own linear-attention definition in §3.2: inside a sub-window of n=M^2 tokens, the ELU+1 kernel gives ϕ(Q)(ϕ(K)^T V) at O(M^2 d^2) per window, hence O(HW d^2) total, not O(HW·M^2). If the derived total differs from the stated formula, the sole quantitative support for the architecture's efficiency advantage is invalid; in that case, absent any experimental tables, the SOTA claim has no support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—SOTA gains on KITTI, Waymo, and SemanticKITTI—rests entirely on experiments that are missing. Section 4.3 says 'In ??, ??, and ??, the above evaluation metrics are compared with previous works' and §4.4 refers to '?? and ??' for ablation results; no tables, metrics, or training details appear anywhere. The qualitative Figure 3 caption asserts improvement but supplies no quantitative comparison. Thus the strongest claim is not merely weakly supported; it is unverifiable from the manuscript. The only quantitative argument in the paper, the SPCN efficiency analysis in §3.2, is also internally inconsistent. After introducing linear attention with φ(x)=ELU(x)+1 and claiming O(N) complexity, the analysis computes per-window cost as K·O(M^4)=O(HW·M^2). That is the cost of standard softmax attention inside M×M windows (n=M^2 tokens, n^2=M^4), not linear attention. For linear attention, per-window cost is O(M^2 d^2); summing over K=HW/M^2 windows gives O(HW d^2). The claimed 400× reduction over global attention therefore follows from windowed standard attention, not from the linear-attention kernel. This does not disprove the architecture, but it removes the stated theoretical efficiency justification and leaves the central performance claim wholly unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Co-Win, a bird's-eye-view (BEV) LiDAR perception framework for joint object detection and instance segmentation. The architecture has three components: an Axis-Fusion Network (AFN) that encodes raw point clouds into BEV representations, a Sub-window Parallel Computing Network (SPCN) backbone that applies window-based linear attention, and a mask-based query decoder with Hungarian matching. The authors claim significant performance gains over state-of-the-art methods on KITTI, Waymo Open Dataset, and SemanticKITTI, while also reducing computational complexity relative to global attention. The manuscript provides detailed architectural equations and a qualitative figure, but contains no quantitative experimental results; Section 4.3 and Section 4.4 are empty placeholders.","tokens_in":9983,"tokens_out":4166,"duration_ms":42071,"significance":"If the claimed performance were substantiated, Co-Win would be a meaningful contribution to BEV-based LiDAR perception, combining mask-based instance segmentation with windowed linear attention and reporting results on multiple major benchmarks. The system design is coherent and builds on established components such as deformable attention, transformer decoders, and Hungarian matching, and the paper is explicit about its intended computational benefits. However, because no experiments, metrics, ablations, or training details are presented, the central claim of state-of-the-art performance cannot be evaluated; the significance of the work is therefore unsubstantiated.","major_comments":[{"comment":"The central claim of 'significant performance gains over existing state-of-the-art methods' is unsupported. Section 4.3 contains only the sentence 'In ??, ??, and ??, the above evaluation metrics are compared with previous works,' with no tables, metrics, error bars, or comparison results, and Section 4.4 similarly refers to '?? and ??' for ablation results. The paper therefore provides no experimental evidence for its primary contribution, and the claim cannot be verified from the manuscript.","section":"Abstract; §4.3"},{"comment":"The computational complexity analysis is internally inconsistent. After introducing linear attention with the kernel feature map φ(x)=ELU(x)+1 and claiming O(N d²) complexity, the 'Computational Complexity Analysis' paragraph gives Ω(SPCN)=K·O(M⁴)=O(HW·M²). The latter is the cost of standard softmax attention inside M×M windows, not of the linear-attention formulation; for linear attention, the per-window cost is O(M² d²), and summing over K=HW/M² windows yields O(HW d²). Consequently, the claimed 400× reduction over global attention is not a consequence of the stated linear-attention mechanism, and the theoretical efficiency justification needs to be corrected.","section":"§3.2"},{"comment":"The qualitative comparison in Figure 3 does not support the claim of 'significant improvement': the figure caption asserts this improvement without quantitative backing, and the accompanying text notes objects that 'didn't be recognized by previous method' based on a single visual example. No detection or segmentation metrics are reported, so this result is anecdotal and cannot substantiate the abstract's performance claim.","section":"§4.5"},{"comment":"Several components of the proposed architecture are specified at a level that prevents reproducibility. For example, the statistical fusion function S, the feature similarity measure sim(·,·), the token generator T for GGIT, and the exact structure of the 'Sub-window Block' are not defined unambiguously; additionally, the number of layers L, block depths D_l, channel counts C, and loss weights λ are never given, nor is any training schedule or data preprocessing detail provided. As a methods paper, the absence of these specifications is a load-bearing deficiency even beyond the missing experiments.","section":"§3.1; §3.3"}],"minor_comments":[{"comment":"There are multiple unresolved placeholders, including 'Figure. ??' in §3.1, 'Fig. ??' in §3.3, and '??' in §4.3/§4.4; these should be replaced with actual references before submission.","section":"Throughout"},{"comment":"The caption of Figure 3 contains grammatical errors and an unsupported evaluative statement; for example, 'ours shows a significant improvement (e.g. arrows in (b) (c) (d)) didn't be recognized by previous method' is not a complete sentence and should be rewritten.","section":"§4.5"},{"comment":"In the related work, the description of PointRCNN++ as 'enhanced accuracy through distance bin-based encoding' is not reflected in the cited reference [17]; consider clarifying the contribution or revisiting the citation.","section":"§2"},{"comment":"The notation for sub-window size M conflicts with the use of M in the multi-scale deformable attention formula in §3.3, where M denotes the number of attention heads; consider renaming one of them to avoid ambiguity.","section":"§3.2; §3.3"},{"comment":"The phrase 'multi-modality inherent in environmental understanding' is vague; the method itself is LiDAR-only, so the relevance of 'multi-modality' is not clear.","section":"§1"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an incomplete draft, with unresolved '??' placeholders, missing figure references, and empty experimental sections. Even setting aside the absent experiments, the inconsistency in the complexity analysis in §3.2 undermines a stated theoretical contribution. The recommendation is based solely on the technical content, which lacks verifiable support for the paper's central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper has no experiments—§4.3 and §4.4 are literally placeholders with '??' where tables and numbers should be. And the one quantitative argument it does make, the efficiency analysis in §3.2, is internally inconsistent.\n\nWhat's actually new: the three-part architecture—AFN encoder with multi-perspective projection and geometric fusion, the SPCN backbone with windowed attention, and the mask-based decoder with Hungarian matching—is a coherent recombination of known pieces from Swin, Mask2Former, and MaskBEV. The specific integration is not something I've seen as a whole, so there is a kernel of a contribution there.\n\nWhere it falls apart: the abstract claims 'significant performance gains' on KITTI, Waymo, and SemanticKITTI. No table, no metric, no training configuration appears anywhere. Figure 3 shows qualitative output, but without numbers it's just a picture. The ablation study is also empty. For a systems paper, this is fatal.\n\nThe complexity analysis only makes it worse. They introduce linear attention with φ(x)=ELU(x)+1 and claim O(N) complexity, then compute Ω(SPCN) = K·O(M^4) = O(HW·M^2). That is the cost of standard softmax attention inside M×M windows, not linear attention. The 400× reduction over global attention follows from windowing, not from the linear-attention kernel. So the paper's only quantitative claim is wrong about its own mechanism. This doesn't disprove the architecture, but it means the theoretical justification as written doesn't hold.\n\nMinor notes: the abstract says 'variational approach' but no variational objective appears in the loss. GGIT is described at a high level without implementation detail. The citation pattern looks fine—no self-citations, and MaskBEV is properly referenced.\n\nBottom line: this is an architecture sketch, not a paper. No referee should spend time on it until the authors actually run the experiments and fix the complexity analysis. If they do, the ideas might be worth a second look. As submitted, desk reject.","headline":"No experiments, and the one quantitative argument contradicts itself—this is a placeholder submission, not a paper.","tokens_in":10444,"tokens_out":2952,"would_cite":false,"duration_ms":30891,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Co-Win claims that predicting LiDAR objects as complete footprint masks, using a windowed linear-attention backbone, outperforms prior box-based and mask-based bird's-eye-view methods.","keywords":["LiDAR point clouds","bird's-eye view perception","instance segmentation","3D object detection","linear attention","mask-based decoder","autonomous driving","BEV detection"],"falsifier":"Train Co-Win with global softmax attention in place of SPCN linear attention under identical settings and compare mask IoU on heavily occluded vehicles in SemanticKITTI; if the linear-attention version drops materially, the paper's claim that linear attention preserves representational power is falsified, and a plot of per-instance mask IoU against occlusion fraction would also test the boundary-completion claim directly.","tokens_in":9353,"feed_emoji":"🚗","tokens_out":6429,"duration_ms":67187,"temperature":0.7,"pith_summary":"This paper is trying to establish that bird's-eye-view perception for autonomous driving should predict each vehicle as a complete footprint mask rather than as a bounding box, and that this mask-based formulation can be made efficient with a windowed linear-attention backbone. The proposed Co-Win pipeline has three components: an Axis-Fusion Network (AFN) that encodes raw point clouds from three orthogonal projections, a Sub-window Parallel Computing Network (SPCN) whose linear attention reduces complexity from $O(N^2)$ to $O(N)$, and a mask-based decoder that outputs class labels and instance masks from learned object queries. The authors claim that this combination yields significant performance gains over prior state-of-the-art methods on KITTI, Waymo Open Dataset, and SemanticKITTI, with qualitatively better footprint completion than prior BEV mask work. If the central claim is right, LiDAR perception can deliver planner-ready object shapes at lower compute than global-attention alternatives.","feed_headline":"Co-Win detects vehicles as footprint masks, not boxes","feed_subtitle":"A windowed linear-attention backbone keeps mask prediction cheap while recovering precise vehicle shapes in bird's-eye view.","key_machinery":"The load-bearing mechanism is the Sub-window Parallel Computing Network (SPCN), a windowed transformer backbone whose linear attention is written as $\\phi(Q)(\\phi(K)^T V)$ with kernel feature map $\\phi(x) = \\text{ELU}(x) + 1$. This reordering lets the network compute the key-value product first, reducing complexity from $O(N^2 d)$ to $O(N d^2)$, and partitioning the BEV map into non-overlapping sub-windows reduces memory relative to global attention; the paper estimates a 400x complexity reduction for a 200x200 map with window size 10. The mask-based decoder completes the argument by taking features from four SPCN stages and producing instance masks through mask embeddings dotted with pixel features, while the AFN supplies multi-plane geometric features and global tokens that are injected into both the backbone and decoder. What this machinery does is convert a normally box-regression problem into a mask-completion problem without paying global-attention cost.","core_discovery":"The paper's central claim is that a single feedforward pass over a BEV point-cloud representation can jointly detect and segment instances by predicting binary footprint masks, and that Co-Win's specific architecture—AFN encoding, SPCN windowed linear attention, and query-based mask decoder—makes that mask prediction both more accurate and cheaper than previous box-based or mask-based BEV methods. The authors report that using the ELU+1 feature map in linear attention preserves representational power while lowering attention complexity, and that mask prediction with Hungarian matching over classification, binary cross-entropy, and Dice losses recovers object boundaries, including completing the unobserved portions of occluded vehicles. They evaluate on KITTI, Waymo Open Dataset, and SemanticKITTI and state that Co-Win achieves significant performance gains over existing state-of-the-art methods.","pith_inferences":["Because the decoder is query-based and mask-based rather than class-specific, the same architecture could be pointed at other BEV map elements—lane markings, curbs, pedestrians—by adding queries and mask annotations, turning Co-Win into a unified panoptic BEV predictor.","The paper claims a 400x complexity reduction for a 200x200 map with window size 10; a direct wall-clock comparison of SPCN against global softmax attention at the same feature resolution would tell whether that theoretical reduction survives in practice on embedded hardware.","The ELU+1 linear-attention kernel is a natural point of comparison: swapping in other positive kernels and measuring mask IoU would show whether the exact kernel or just linear attention per se is responsible for the reported quality."],"forward_implications":["LiDAR object detection can be framed as mask prediction instead of box regression without sacrificing efficiency.","A windowed linear-attention backbone with the ELU+1 kernel is sufficient for joint detection and instance segmentation in BEV point clouds.","A single pass can both localize vehicles and complete their occluded footprint masks, giving downstream planners shape information that boxes cannot provide.","The reported complexity reduction supports real-time BEV perception on resource-constrained hardware.","The method transfers across KITTI, Waymo Open Dataset, and SemanticKITTI, suggesting the mask-based formulation generalizes across LiDAR benchmarks."],"supporting_citations":[{"why":"Supplies the mask-as-segmentation reframing that the decoder adapts to BEV.","marker":"[4]"},{"why":"Prior BEV mask-based joint detection and footprint completion work that Co-Win compares against.","marker":"[7]"},{"why":"Establishes joint object detection and instance segmentation via a mask branch in 2D, the lineage the decoder extends.","marker":"[8]"},{"why":"Pillar-based BEV encoding baseline; also used as the ablation baseline encoder.","marker":"[14]"},{"why":"BEV-based detection baseline that Co-Win claims to outperform.","marker":"[29]"},{"why":"Provides the KITTI benchmark for AP70 comparisons.","marker":"[6]"},{"why":"Provides the Waymo Open Dataset benchmark with instance masks for mAP and mAPH comparisons.","marker":"[23]"},{"why":"Provides the SemanticKITTI benchmark for AP50, AP70, mAP, and mIoU evaluations.","marker":"[1]"}],"fun_headline_variants":["Co-Win predicts vehicle shapes as masks, not boxes","Windowed attention makes LiDAR detection and segmentation one pass","Footprint masks from LiDAR: Co-Win does detection and segmentation","One BEV pass for both detection and mask segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the SPCN's windowed linear attention with the feature map $\\phi(x) = \\text{ELU}(x) + 1$ keeps enough spatial detail to predict precise footprint masks while cutting complexity from quadratic to linear; if it loses fine boundary or long-range information, the whole architecture underperforms regardless of the decoder.","fun_headline_variants_meta":{"raw":{"variants":["Co-Win predicts vehicle shapes as masks, not boxes","Windowed attention makes LiDAR detection and segmentation one pass","Footprint masks from LiDAR: Co-Win does detection and segmentation","One BEV pass for both detection and mask segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000486,"raw_usage":{"total_tokens":2358,"prompt_tokens":871,"completion_tokens":1487,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":1418}},"tokens_in":487,"tokens_out":1487,"duration_ms":11921,"temperature":1.0,"reasoning_tokens":1418,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:08:37.176409+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Co-Win with global softmax attention in place of SPCN linear attention under identical settings and compare mask IoU on heavily occluded vehicles in SemanticKITTI; if the linear-attention version drops materially, the paper's claim that linear attention preserves representational power is falsified, and a plot of per-instance mask IoU against occlusion fraction would also test the boundary-completion claim directly.","supporting_citations":[{"cited_title":"Cheng, A","cited_arxiv_id":null,"evidence_quote":"Supplies the mask-as-segmentation reframing that the decoder adapts to BEV."},{"cited_title":"Guimont-Martin, J.-M","cited_arxiv_id":null,"evidence_quote":"Prior BEV mask-based joint detection and footprint completion work that Co-Win compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes joint object detection and instance segmentation via a mask branch in 2D, the lineage the decoder extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Pillar-based BEV encoding baseline; also used as the ablation baseline encoder."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BEV-based detection baseline that Co-Win claims to outperform."},{"cited_title":"Geiger, P","cited_arxiv_id":null,"evidence_quote":"Provides the KITTI benchmark for AP70 comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Waymo Open Dataset benchmark with instance masks for mAP and mAPH comparisons."},{"cited_title":"Behley, M","cited_arxiv_id":null,"evidence_quote":"Provides the SemanticKITTI benchmark for AP50, AP70, mAP, and mIoU evaluations."}],"review_version":1}