{"id":"04b6a572-738a-4513-9d0f-0b791d632404","arxiv_id":"2508.07300","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"BEVANet reports 81.0% mIoU at 33 FPS on Cityscapes, 0.9% above PIDNet-M but with more parameters and FLOPs.","lead":"BEVANet is a new neural network for real-time semantic segmentation, the task of labeling every pixel in a photo with an object class. It combines several attention mechanisms and reports 79-81% accuracy on the Cityscapes driving benchmark while running at 33 frames per second on a high-end GPU.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed 0.9–1.5 pt mIoU gains rest on single runs and adopted PIDNet baselines; without seed-level or same-protocol replication the 81.0/79.3 SoTA margin is not established.","rationale":"Good-faith reading: BEVANet is a plausible combination of existing large-kernel attention modules inside a PIDNet-like bilateral frame; the code link is a plus for reproducibility. The main correctness risk is not a specific algorithmic error but evidential: the headline numbers are single-run and the external baselines are adopted from the original PIDNet paper. I agree with the reader that this is the weakest point, and the conditional verdict is appropriate. Also note a minor overclaim: the abstract says 'we introduce LKA,' although §2.2 credits VAN/LSKA; this affects novelty phrasing, not correctness. The strongest check is replication under identical protocol with seeds; that would settle whether 79.3/81.0 and the module attributions are real. Therefore no change to the CONDITIONAL verdict is needed, though a confirmed failure of replication would justify moving to REJECT.","tokens_in":8673,"tokens_out":7636,"duration_ms":72930,"concrete_test":"Using the released code, train BEVANet (with and without ImageNet pretraining) and PIDNet-M under the exact §3.1 recipe in the same codebase, with at least 5 seeds each, recording mean ± std mIoU. Also rerun the Table 4 ablation rows (at least rows 1, 2, 7, 9) with at least 5 seeds. If the mean BEVANet−PIDNet-M margin is less than 1 std, or if the cumulative 1.5% ablation gain collapses below about 0.5 pts, the central SoTA claim is not supported by current evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—BEVANet reaches 81.0 mIoU with ImageNet pretraining and 79.3 without, beating PIDNet-M by 0.9 pt (§3.2, Tables 1 and 3)—depends on the assumption that small mIoU differences in Tables 4 and 5 are real. All mIoU values are single runs: no seeds, no error bars, no significance tests. The cumulative 1.5% 'Performance Enhancement' (§3.3) is the difference between rows 2 and 9 of Table 4 (77.77→79.27), where each row is one run; intermediate increments (0.8, 0.26, 0.4, 0.5) are in the same range as typical Cityscapes seed noise. Table 1's baselines are 'adopted from PIDNet [8]' (Table 1 caption), not measured under the §3.1 recipe (batch 12, LR 0.008, 484 epochs, OHEM), so systematic protocol differences, not just random noise, could explain the 0.9 pt gap. The ablation's first row (78.22) is itself PIDNet's no-pretraining number. Thus the load-bearing evidence for both the SoTA comparison and the attribution of gains to SDLSKA/CKS/BGAF/DLKPPM is single-run, cross-paper comparisons. This does not show the architecture is wrong, only that the headline margin is not yet evidenced.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BEVANet, a bilateral network for real-time semantic segmentation. It combines two branches (detail and semantic) with four new components: an Efficient Visual Attention (EVA) block built from Sparse Decomposed Large Separable Kernel Attention (SDLSKA), a Comprehensive Kernel Selection (CKS) mechanism, a Deep Large Kernel Pyramid Pooling Module (DLKPPM), and a Boundary Guided Adaptive Fusion (BGAF) module. The main empirical claims are 81.0% mIoU on Cityscapes val with ImageNet pretraining, 79.3% without pretraining, and 33 FPS on an RTX 3090, which the paper labels state-of-the-art. Ablations in Tables 4 and 5 attribute gains of 0.26-0.8 mIoU points to individual modules and 1.5 points cumulatively.","tokens_in":9080,"tokens_out":2629,"duration_ms":27292,"significance":"If confirmed, BEVANet would offer a modest but useful accuracy improvement over PIDNet-M at real-time speeds, with the interesting property of retaining most of the gain without ImageNet pretraining. The paper contributes a concrete combination of large-kernel attention, selective kernel fusion, pyramid pooling, and boundary-guided fusion, and it releases code. These are real strengths. However, the headline margin (0.9 points over PIDNet-M with pretraining, 1.05 points without) rests entirely on single-run experiments and on baseline numbers adopted from prior papers. That is insufficient evidence for the 'state-of-the-art' label as it stands, though the issue is fixable with additional experiments or more cautious claims.","major_comments":[{"comment":"The ablation study reports one run per configuration and no error bars or significance tests. The individual increments (0.8, 0.26, 0.4, and the 0.53 in Table 5) and the cumulative 1.5-point 'Performance Enhancement' are within the range of seed-to-seed variation commonly observed for Cityscapes mIoU. The statement that 'SDLSKA significantly outperforms its components' and the attribution of the 1.5-point gain to the proposed modules are therefore not statistically supported. Please provide repeated-seed results (mean and standard deviation) for the key rows, or at least for the baseline, final model, and the four module increments, and state whether differences are reproducible.","section":"Section 3.3, Table 4"},{"comment":"The SoTA comparison is cross-paper: Table 1 states that most baselines are 'adopted from PIDNet [8]', and Table 3 reuses PIDNet's no-pretraining numbers. BEVANet is trained with a different recipe (batch size 12, LR 0.008, 484 epochs, OHEM) from that used by PIDNet in its original paper. Thus the 0.9-point advantage over PIDNet-M in Table 1 and the 1.05-point advantage in Table 3 could be due to protocol or environment differences rather than the architecture. Please re-run PIDNet-M and, if feasible, one or two other baselines under the exact same training and evaluation protocol, and report the comparison including variance.","section":"Tables 1-3 and Section 3.1"},{"comment":"The paper calls BEVANet 'state-of-the-art' and 'a superior balance of speed and accuracy,' but the accuracy gain over PIDNet-M is 0.9 mIoU at the cost of 40.8 additional GFLOPs, 24.2 million extra parameters, and 7 FPS. Without an accuracy-efficiency metric or a controlled comparison, the 'superior balance' claim is not established by the table alone. This is secondary to the reproducibility issue, but the claim should be calibrated to the actual cost difference.","section":"Section 3.2, Table 1"}],"minor_comments":[{"comment":"The model name is typeset inconsistently as 'BEV ANet' and 'BEVANet'; please standardize. Also 'Resolusion' in Table 1 should be 'Resolution'.","section":"Global"},{"comment":"The sentence 'This enables continuous branch interaction to improve semantic understanding and object boundary .' appears incomplete; the section ends abruptly without explaining how the two branches communicate beyond referring to Fig. 3.","section":"Section 2.1"},{"comment":"The introduction says BEVANet-S 'further achieves 83% mIoU on CamVid,' while Table 2 reports 83.1; please be precise. Also, the claim that BEVANet-S beats PIDNet-S-Wider by 1.1% mIoU while having 40 GFLOPs fewer should be checked against the same single-run caveat as the Cityscapes numbers.","section":"Section 3.2, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The core concern is evidentiary, not architectural. The proposed modules are coherent and the code is promised, but the empirical evidence for the headline margin is too thin: single runs and adopted baselines. I would be willing to accept after the authors provide multi-seed results for the main tables and either re-run key baselines in-house or substantially soften the SoTA claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"BEVANet is a competently assembled real-time segmentation network that recombines known large-kernel attention ideas into a bilateral architecture. It is not a breakthrough, but it is a concrete, code-released system that could be useful for practitioners. The real contribution is the specific module designs—SDLSKA, CKS, DLKPPM, and BGAF—and the ablation in Table 4 shows each adds a bit, though the increments are small.\n\nThe main weakness is evidential. All mIoU numbers are single-run, no error bars or seed repetitions. The cumulative 1.5-point gain over the BA baseline (Table 4, row 2 to row 9) could easily be within run-to-run noise. And the comparison against PIDNet-M in Table 1 uses baselines adopted from the PIDNet paper, not re-run under the same protocol, so the 0.9-point margin is not controlled. The efficiency trade-off is also not favorable: BEVANet-M uses 58.6M params and 238 GFLOPs for 0.9 points over PIDNet-M's 34.4M and 197 GFLOPs. Finally, the abstract's 'we introduce the Large Kernel Attention mechanism' is inaccurate—LKA has been around since VAN.\n\nWhat the paper does well: it is clearly written, the modules are well-illustrated, and the code is public. The CamVid result for BEVANet-S looks promising (83.1 vs 82.0 for PIDNet-S-Wider with fewer FLOPs), but again, those baselines are likely adopted.\n\nWho should read this: people in real-time segmentation who want to see a clean integration of LKA-style modules into a bilateral net. The paper doesn't establish a new state of the art, but it offers a useful architecture and a fair amount of design space exploration.\n\nFor peer review: I'd send it to a serious referee. The architecture is concrete and the code is available, and the questions raised are empirical. However, I'd expect the experimental protocol to be strengthened: multiple seeds, confidence intervals, and ideally re-running the strongest baselines under the same setup.","headline":"BEVANet is a competently assembled architecture-engineering paper that recombines known LKA components, but the state-of-the-art claim is undercut by single-run results and adopted baselines.","tokens_in":9594,"tokens_out":5605,"would_cite":true,"duration_ms":46025,"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":"BEVANet bundles large-kernel attention modules into a bilateral network, claiming 81.0% mIoU on Cityscapes at 33 FPS.","keywords":["semantic segmentation","real-time","large kernel attention","bilateral network","feature fusion","Cityscapes","boundary guidance","kernel selection"],"falsifier":"Run the Table 4 ablation five or more times with different random seeds and report mean and standard deviation. If the cumulative 1.5 percent mIoU gain of BEVANet over the PIDNet baseline is within one standard deviation of run-to-run variance, the central claim that the new modules cause the improvement is falsified. Alternatively, replace SDLSKA with plain LSKA at matched FLOPs and parameters; if mIoU does not drop, the sparse decomposition is not the cause.","tokens_in":8583,"feed_emoji":"🚗","tokens_out":4732,"duration_ms":40176,"temperature":0.7,"pith_summary":"The paper proposes BEVANet, a real-time semantic segmentation network that uses large kernel attention to capture both broad context and fine detail. Its central claim is that a bilateral architecture with four new modules—SDLSKA, CKS, DLKPPM, and BGAF—achieves 81.0 percent mIoU on Cityscapes after pretraining and 79.3 percent without, running at 33 frames per second. If correct, this shows that large-kernel attention can be made efficient enough for real-time use while improving accuracy over the PIDNet baseline. The authors intend to establish that each module contributes meaningfully, attributing a cumulative 1.5 percent mIoU gain to them.","feed_headline":"BEVANet hits 81% mIoU on Cityscapes at 33 FPS","feed_subtitle":"Large-kernel attention and bilateral fusion beat PIDNet-M by 0.9 points without breaking real time.","key_machinery":"The central mechanism is the Efficient Visual Attention (EVA) block built from Sparse Decomposed Large Separable Kernel Attention (SDLSKA) and Comprehensive Kernel Selection (CKS). SDLSKA expands the receptive field to 35 through a 5×5 convolution plus 1×11 and 11×1 strip kernels with dilation 3; CKS fuses small, large-horizontal, and large-vertical kernel features using joint channel and spatial attention, letting the network adapt its effective field per feature.","core_discovery":"BEVANet introduces the LKA mechanism into a two-branch network: a high-level branch for semantics and a low-level branch for details, with frequent communication via a Boundary Guided Adaptive Fusion module. The Sparse Decomposed Large Separable Kernel Attention expands the receptive field to 35 with low computation by decomposing a large kernel into a 5×5 convolution and two dilated strip kernels, while the Comprehensive Kernel Selection mechanism dynamically fuses features from kernels of different sizes. The Deep Large Kernel Pyramid Pooling Module enriches context while reducing pooling information loss. On Cityscapes, BEVANet reaches 81.0 percent mIoU after ImageNet pretraining and 79.3","pith_inferences":["If the per-module gains are real, a natural extension is to apply the SDLSKA-CKS pair to backbone networks in detection or instance segmentation, where large receptive fields also help.","The claim that CKS outperforms decoupled channel-spatial selection (LSKNet/SKNet) predicts a testable hypothesis: coupling channel and spatial attention during kernel fusion should consistently beat separate selection modules in other architectures.","A direct test of the non-pretraining claim would be to scale BEVANet up without pretraining on tougher benchmarks like ADE20K; if the advantage persists, the design may genuinely compensate for missing pretraining.","The absence of repeated-seed experiments leaves open the possibility that some of the 0.1-point improvements are noise; a multi-seed rerun of Table 4 would clarify which modules are load-bearing."],"forward_implications":["If the reported results hold, BEVANet establishes that large-kernel attention can be computed in real time and still outperform standard convolutional real-time baselines on Cityscapes.","The 79.3 percent mIoU without ImageNet pretraining suggests the design reduces dependence on large pretraining datasets, easing deployment in data-limited settings.","The bilateral architecture with boundary-guided fusion may generalize to other pixel-labeling tasks, such as depth estimation or edge detection, where boundary accuracy matters.","The modular design (SDLSKA, CKS, DLKPPM, BGAF) provides a recipe that other lightweight networks can adopt."],"supporting_citations":[{"why":"PIDNet supplies the baseline architecture, the BAG fusion module, and most comparison numbers in Tables 1-3.","marker":"[8]"},{"why":"LSKA provides the large separable kernel attention design that SDLSKA builds on and is a direct ablation comparison.","marker":"[14]"},{"why":"SLaK contributes the sparse decomposition of large kernels into strip convolutions, used by SDLSKA.","marker":"[12]"},{"why":"DAPPM is the pyramid pooling baseline that DLKPPM extends and is compared against in Table 5.","marker":"[7]"},{"why":"LSKNet is the large selective kernel approach that CKS improves upon in the ablation study.","marker":"[15]"},{"why":"SKNet is the selective kernel network that motivates the kernel selection idea and is the prior work CKS claims to surpass.","marker":"[16]"},{"why":"Cityscapes is the primary benchmark dataset whose validation set supplies all reported mIoU results.","marker":"[4]"}],"fun_headline_variants":["BEVANet: 81% mIoU at 33 FPS with large-kernel attention","Real-time segmentation hits 81% mIoU: BEVANet's bilateral fusion","BEVANet's large-kernel attention: 81% mIoU at 33 FPS","BEVANet: 79.3% mIoU without pretraining, 33 FPS"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that the single-run mIoU differences of 0.1 to 0.5 points in Table 4 are statistically real; if they fall within run-to-run noise, the claimed contributions of the four modules are not supported.","fun_headline_variants_meta":{"raw":{"variants":["BEVANet: 81% mIoU at 33 FPS with large-kernel attention","Real-time segmentation hits 81% mIoU: BEVANet's bilateral fusion","BEVANet's large-kernel attention: 81% mIoU at 33 FPS","BEVANet: 79.3% mIoU without pretraining, 33 FPS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001143,"raw_usage":{"total_tokens":4593,"prompt_tokens":768,"completion_tokens":3825,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":3721}},"tokens_in":512,"tokens_out":3825,"duration_ms":25171,"temperature":1.0,"reasoning_tokens":3721,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:12:30.727405+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the Table 4 ablation five or more times with different random seeds and report mean and standard deviation. If the cumulative 1.5 percent mIoU gain of BEVANet over the PIDNet baseline is within one standard deviation of run-to-run variance, the central claim that the new modules cause the improvement is falsified. Alternatively, replace SDLSKA with plain LSKA at matched FLOPs and parameters; if mIoU does not drop, the sparse decomposition is not the cause.","supporting_citations":[{"cited_title":"Fully convolutional networks for semantic segmenta- tion,","cited_arxiv_id":null,"evidence_quote":"DAPPM is the pyramid pooling baseline that DLKPPM extends and is compared against in Table 5."},{"cited_title":"U-net: Convolutional networks for biomedical im- age segmentation,","cited_arxiv_id":null,"evidence_quote":"PIDNet supplies the baseline architecture, the BAG fusion module, and most comparison numbers in Tables 1-3."},{"cited_title":"Pidnet: A real-time semantic segmentation network inspired by pid controllers,","cited_arxiv_id":null,"evidence_quote":"LSKNet is the large selective kernel approach that CKS improves upon in the ablation study."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cityscapes is the primary benchmark dataset whose validation set supplies all reported mIoU results."}],"review_version":1}