{"id":"c3b599f9-e211-4c78-a012-4291919810cc","arxiv_id":"2411.16481","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"On five wide-FoV segmentation benchmarks, a Mamba-plus-deformable-convolution decoder beats common segmentation heads with the same backbones while cutting decoder FLOPs by roughly 97% versus UperHead.","lead":"This paper introduces a decoder that combines Mamba state-space scanning with deformable convolutions, and reports that it improves semantic segmentation on 180-degree fisheye and 360-degree panoramic images across five benchmarks. The decoder is designed as a drop-in replacement for standard segmentation heads, aiming to give similar or better accuracy with far fewer parameters and FLOPs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains rest on uncontrolled re-implemented baselines and mixed efficiency baselines; a controlled comparison with shared configs and published-number validation is needed to confirm the +2.5 mIoU and 72%/97% claims.","rationale":"The reader's weakest_assumption—that re-implemented baselines are trained under the same settings as the proposed model—is indeed the most load-bearing point. The headline +2.5 mIoU is the single strongest quantitative evidence for the decoder's value, and it comes from exactly the comparison (same VMamba-T backbone, different decoder) where training details can dominate. If the re-implemented VMamba† is undertrained or uses different augmentation, the central accuracy claim collapses to a training-recipe difference. The efficiency claim is secondary but also part of the central claim; it is internally inconsistent because the 72% and 97% figures come from different baseline decoders, and Table 1 directly contradicts 'Ours is smaller than UperHead' on parameters. Both issues are addressable with released configs and a controlled re-run, which is why CONDITIONAL (the reader's verdict) is the right call rather than ACCEPT or REJECT. My concern does not change that verdict; it sharpens the conditions: the authors should validate baseline fairness against published numbers and report a single consistent efficiency comparison. I agree with the reader's identification of the weakest assumption; the efficiency mixing is a related but distinct flaw that reinforces the need for caution, not a separate fatal objection.","tokens_in":17161,"tokens_out":3964,"duration_ms":39157,"concrete_test":"Request the released code and training configs for all '†re-implementation' baselines. Then run a controlled comparison on Stanford2D3D: train (a) VMamba-T + UperHead and (b) VMamba-T + Deformable Mamba Decoder under identical augmentation, input resolution, optimizer (AdamW, lr 6e-5, wd 0.01), iteration count (80K), batch size, and seed, with at least 3 seeds. Also check whether the re-implemented VMamba† reproduces the published VMamba S2D3D mIoU within 1 point; if it falls short, the +2.5 gap may be due to a weak baseline. For efficiency, recompute the abstract's 72%/97% numbers using the same baseline decoder for both metrics and report both directions (e.g., vs UperHead and vs CGRHead).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim has two pillars: (1) the Deformable Mamba Decoder improves wide-FoV segmentation, with the headline +2.5 mIoU over VMamba on Stanford2D3D, and (2) it is far more efficient, with 72% fewer parameters and 97% fewer FLOPs than widely-used decoders. Both pillars depend on comparisons that are not strictly apples-to-apples. First, the performance gain is established mainly against baselines marked '†re-implementation' (VMamba†, SegNeXt†, 360SFUDA†, Tables 4-7). Section 4.2 specifies the optimizer and schedule for the proposed method but does not state whether these re-implementations use the same augmentation, preprocessing, crop strategy, or effective training iterations. If, for example, the re-implemented VMamba† was trained with weaker augmentation or different resolution handling than DMamba, part or all of the +2.5 mIoU could be a training-detail artifact, not decoder distortion awareness. This is especially consequential because the cleanest comparison—same VMamba-T backbone, only decoder changed—is exactly the one where baseline training details matter most. Second, the efficiency claim mixes baselines across tables: the 72% parameter reduction is versus CGRHead (Table 1: 282.6M to 77.3M; Table 2: 40.6M to 11.2M), while the 97% FLOPs reduction is versus UperHead in Tables 2/3 (206.9G to 6.0G). Against the same UperHead column, Ours actually has more parameters (Table 1: 77.3M vs 40.5M; Table 2: 11.2M vs 31.5M is fewer, but Table 1 is the counterexample). Thus the abstract's efficiency headline is not a single consistent comparison and would not survive a reader checking Table 1. Neither issue is fatal on its own—both are addressable—but together they make the central claim less secure than the abstract suggests.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Deformable Mamba Decoder, a decoder that combines a quadri-directional Mamba scan branch with a deformable convolution (DCNv2) branch and PixelShuffle upsampling, and demonstrates it as a plug-in decoder for CNN-, Transformer-, and Mamba-based backbones in wide field-of-view (180° fisheye and 360° panoramic) semantic segmentation. The method is evaluated on five datasets (Stanford2D3D, Matterport3D, SynPASS, WoodScape, SynWoodScape) and is reported to improve mIoU over prior decoders while reducing parameters and FLOPs, with a headline +2.5 mIoU gain on Stanford2D3D over a VMamba-based baseline. The paper includes ablations of scan direction, deformable convolution, and upsampling method, and releases code.","tokens_in":17591,"tokens_out":3860,"duration_ms":35835,"significance":"If the empirical claims are sustained, the paper makes a useful contribution: it identifies decoder-level distortion awareness as a modular component that can be attached to diverse backbones, and it provides one of the first decoder-focused uses of Mamba for dense prediction. The experimental scope is a strength: five datasets, three backbone families, and ablations that isolate DCN, scan direction, and upsampling. The availability of code is also a concrete asset. However, the main quantitative claims currently rest on comparisons whose fairness and precise accounting need to be verified before the stated advantages can be accepted.","major_comments":[{"comment":"The performance gains over re-implemented baselines (VMamba†, SegNeXt†, 360SFUDA†) are load-bearing for the +2.5 mIoU claim, but the manuscript does not establish that these re-implementations were trained under identical conditions. Section 4.2 specifies the optimizer, schedule, and iterations for the proposed method but does not state whether the re-implemented baselines used the same augmentation, cropping/resolution policy, preprocessing, or loss weighting. Since the cleanest comparison (same VMamba-T backbone, only the decoder differs) is exactly the case where training details matter most, the paper should report the full training recipe for the re-implemented baselines and ideally release their configuration files, or compare against published numbers from the original papers.","section":"§4.2 and Tables 4-7"}],"minor_comments":[{"comment":"The opening sentence is grammatically incomplete: 'Recent advancements in the Mamba architecture, with its linear computational complexity, being a promising alternative...' should be revised.","section":"Abstract"},{"comment":"There is a typo: 'Deformable Mamaba Decoder' should be 'Deformable Mamba Decoder'.","section":"§4.3"},{"comment":"'desin' in 'lacking deformable desin' should be 'design'.","section":"§4.4"},{"comment":"The PixelShuffle formula's subscript notation is unclear: 'PS(T)h,w,2c = T2h,2w,c/2' mixes input and output dimensions in a way that is hard to parse. Please rewrite the index notation to show the rearrangement explicitly.","section":"§3.2, Eq. (3)"},{"comment":"It should be stated explicitly whether the reported parameter and FLOPs counts refer to the decoder alone or to the full encoder-decoder model; Tables 1-3 list decoder-level numbers while §4.5 discusses overall reductions.","section":"§4.2 and Tables 1-3"},{"comment":"The limitations paragraph does not actually state a limitation of the current method; it only points to future LLM integration. Please add a genuine discussion of limitations, for example the sensitivity of the reported gains to re-implementation details and the mixed efficiency baselines.","section":"Limitations and Future Work"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical study with good breadth of datasets and ablations, but the headline efficiency numbers need to be restated with consistent baselines, and the re-implemented baseline protocol needs to be documented. I would encourage the editor to request that the authors release baseline configurations alongside the code, since the +2.5 mIoU claim depends on those comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid engineering paper with one real reporting flaw. The decoder itself is a sensible combination, but the efficiency headline in the abstract doesn't match any single comparison in the tables.\n\nWhat's new: pairing a single quadri-directional Mamba scan (the SS2D block from VMamba) with a DCNv2 branch and PixelShuffle upsampling, as a plug-in decoder that can sit on top of CNN, Transformer, or Mamba backbones. The components are all prior art, but the specific module and the claimed distortion-awareness for wide-FoV segmentation is new. The paper's strongest evidence is the ablation table: removing the deformable convolution drops mIoU by 1.85 points on Stanford2D3D, and switching from quadri-directional to uni- or bi-directional scanning also costs about a point. That's a clean internal story.\n\nThe breadth of experiments is genuinely good: five datasets, two fisheye and three panoramic, real and synthetic, and the decoder is tested on three backbone families. The code is supposedly on GitHub, which helps.\n\nNow the soft spots, in order of seriousness:\n\nFirst, the efficiency claim as written is wrong. The abstract says 'reducing 72% parameters and 97% FLOPs' without saying the baselines differ. Reading the tables: 72% parameter reduction is versus CGRHead, 97% FLOPs reduction is versus UperHead. Against UperHead on ResNet-50, the decoder actually has almost twice as many parameters (77.3M vs 40.5M). That's a real misreporting issue and should be fixed.\n\nSecond, the performance gains versus the '†re-implementation' baselines (VMamba†, SegNeXt†, 360SFUDA†) are only as trustworthy as the training configs. Section 4.2 gives the schedule for the proposed method but doesn't state whether the re-implemented baselines used the same augmentation, preprocessing, and iterations. On MP3D the gain over SegNeXt is only +0.3 mIoU, which is within run-to-run noise; no error bars or multiple seeds are reported. These issues are fixable, not structural.\n\nThird, the novelty is modest. The decoder is a re-combination of existing blocks with a new purpose, and the paper says so accurately. The motivation—that Mamba decoders are under-explored and that distortion-aware decoders should be plug-in—is reasonable.\n\nOverall: the central argument—that a deformable branch helps a Mamba-based decoder handle wide-FoV distortion—is supported by the ablation. The paper is not a breakthrough, but it is a useful drop-in module for a subfield where efficiency matters. A serious referee should look at it, but the authors need to correct the efficiency claim and provide baseline setup transparency.","headline":"A useful plug-in decoder with a clean ablation story, held back by a misleading efficiency headline and uncontrolled re-implemented baselines.","tokens_in":18126,"tokens_out":2352,"would_cite":true,"duration_ms":20213,"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":"The paper argues that distortion awareness for wide-field-of-view segmentation can be supplied entirely by the decoder, without modifying the backbone, and validates this with a decoder that works across CNN, Transformer, and Mamba…","keywords":["semantic segmentation","wide field of view","panoramic images","fisheye images","Mamba decoder","deformable convolution","state space models","efficient decoder"],"falsifier":"Run a head-swap experiment on Stanford2D3D where UperHead, MusterHead, CGRHead, and the proposed decoder share the exact same augmentation, data split, effective iterations, and seeds; if the proposed decoder does not beat the best baseline by a margin consistent with the paper's +2.5 points, the claim that the decoder alone imparts distortion awareness is not supported.","tokens_in":16910,"feed_emoji":"🧩","tokens_out":6697,"duration_ms":56015,"temperature":0.7,"pith_summary":"The paper argues that distortion awareness for wide-field-of-view segmentation can be supplied entirely by the decoder, without modifying the backbone. It proposes the Deformable Mamba Decoder, whose fusion stages pair a single quadri-directional Mamba scan with a deformable-convolution branch, and plugs this head into CNN, Transformer, and Mamba backbones. Across five fisheye and panoramic benchmarks the decoder reports consistent mIoU gains, including +2.5 points on Stanford2D3D, while cutting decoder parameters by 72 percent and FLOPs by 97 percent relative to widely used heads such as UperHead. A sympathetic reader would care because existing wide-FoV methods usually couple distortion handling to a specific architecture, forcing retraining; a decoder-only fix would let pretrained backbones be reused across sensor types.","feed_headline":"A new decoder head lifts wide-FoV segmentation accuracy by 2.5 points","feed_subtitle":"Swapping in one decoder head beats specialized wide-FoV models while cutting decoder compute by 97 percent.","key_machinery":"The Deformable Mamba Fusion (DMF) block: a fusion-and-upsampling module that takes an encoder feature $E_i$ and a previously fused decoder feature $D_j$, applies one quadri-directional selective scan (SS2D) to $D_j$ and one DCNv2 deformable convolution to $E_i$, concatenates the two branches, fuses them with 3×3 convolutions, and upsamples with PixelShuffle. It is the mechanism that adds data-dependent offsets—the distortion-awareness—without touching the encoder, which is why the same decoder can be attached to CNN, Transformer, and Mamba backbones.","core_discovery":"On the paper's own terms, the discovery is that a Mamba-based decoder can be made distortion-aware by combining the quadri-directional 2D selective scan from VMamba with a parallel DCNv2 deformable-convolution branch in each Deformable Mamba Fusion block. This gives the decoder adaptive sampling offsets that compensate for fisheye and equirectangular deformation while retaining Mamba's linear complexity. With the same VMamba-T backbone, the proposed decoder raises mIoU on Stanford2D3D from 56.8 for the VMamba-plus-UperHead baseline to 59.3, improves results on Matterport3D, SynPASS, WoodScape, and SynWoodScape, and reduces decoder parameters by 72 percent and FLOPs by 97 percent compared with UperHead.","pith_inferences":["Editorial inference: because the decoder is the only component that changes, the same DMF block could plausibly transfer distortion-awareness to other dense prediction tasks on wide-FoV inputs, such as depth estimation, panoptic segmentation, or open-vocabulary segmentation.","Editorial inference: the small margin on Matterport3D (+0.3 to +0.9 mIoU) suggests the benefit of decoder-level deformation is larger on datasets with heavy distortion or fine elongated classes; a controlled head-swap study would reveal which property drives the gain.","Editorial inference: a testable extension is to visualize the learned DCN offsets on equirectangular images; if the offsets systematically align with the spherical-to-planar projection, that would confirm the mechanism is distortion compensation rather than generic capacity.","Editorial inference: since the paper trains all models with one schedule and no auxiliary losses, part of the gain could come from regularization; comparing against a matched-capacity conventional decoder with identical training would isolate the deformable mechanism."],"forward_implications":["A single decoder design can replace task-specific heads across CNN, Transformer, and Mamba backbones, so wide-FoV models no longer need distortion handling baked into the encoder.","The decoder's compute is small enough that the head stops being the FLOP bottleneck: on VMamba-T, UperHead uses 206.9 GFLOPs while the proposed head uses 6.0 GFLOPs, making the total model much cheaper.","Consistent gains on synthetic and real, indoor and outdoor, 180-degree and 360-degree datasets indicate the approach generalizes across distortion types rather than overfitting one sensor.","Quadri-directional scanning contributes more than uni- or bi-directional scanning in the decoder, supporting the claim that scan diversity matters for spatially structured wide-FoV features.","Replacing deformable convolution with ordinary convolution drops mIoU by 1.85 points, so the deformable branch is the component responsible for the distortion-aware gain."],"supporting_citations":[{"why":"Supplies the VMamba backbone and the quadri-directional SS2D scanning block that the decoder's Mamba branch reuses.","marker":"[34]"},{"why":"Defines the selective state-space model whose linear complexity is the efficiency premise for the decoder.","marker":"[17]"},{"why":"Provides DCNv2 deformable convolution with modulation offsets, the mechanism that gives the decoder adaptive distortion awareness.","marker":"[74]"},{"why":"The widely used UperHead decoder that serves as the main baseline and compute bottleneck in the efficiency comparison.","marker":"[54]"},{"why":"A pinhole-designed SegNeXt baseline that the paper re-implements to show narrow-FoV models degrade on wide-FoV inputs.","marker":"[20]"},{"why":"The panoramic 360SFUDA baseline re-implemented for comparison on 360- and 180-degree datasets.","marker":"[70]"},{"why":"A distortion-aware transformer for panoramas whose whole-architecture coupling motivates the decoder-level decoupling.","marker":"[64]"},{"why":"The Stanford2D3D dataset where the headline +2.5 mIoU improvement is measured.","marker":"[1]"},{"why":"The WoodScape fisheye dataset used to demonstrate gains on 180-degree images.","marker":"[62]"}],"fun_headline_variants":["Distortion-aware Mamba decoder cuts FLOPs 97%, lifts mIoU 2.5","Deformable Mamba decoder: 2.5% mIoU gain, 97% fewer FLOPs","Distortion-aware decoder: 97% less compute, better wide-FoV","Mamba decoder adapts to distortion, slashes compute 97%","New wide-FoV decoder: +2.5 mIoU, 72% fewer params"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume the re-implemented baselines were trained under the same data splits, preprocessing, augmentation, and optimization settings as the proposed model, so that swapping the decoder head is the only meaningful change.","fun_headline_variants_meta":{"raw":{"variants":["Distortion-aware Mamba decoder cuts FLOPs 97%, lifts mIoU 2.5","Deformable Mamba decoder: 2.5% mIoU gain, 97% fewer FLOPs","Distortion-aware decoder: 97% less compute, better wide-FoV","Mamba decoder adapts to distortion, slashes compute 97%","New wide-FoV decoder: +2.5 mIoU, 72% fewer params"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000609,"raw_usage":{"total_tokens":2824,"prompt_tokens":920,"completion_tokens":1904,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":1783}},"tokens_in":536,"tokens_out":1904,"duration_ms":12407,"temperature":1.0,"reasoning_tokens":1783,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:03:38.265063+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a head-swap experiment on Stanford2D3D where UperHead, MusterHead, CGRHead, and the proposed decoder share the exact same augmentation, data split, effective iterations, and seeds; if the proposed decoder does not beat the best baseline by a margin consistent with the paper's +2.5 points, the claim that the decoder alone imparts distortion awareness is not supported.","supporting_citations":[{"cited_title":"De- formable convnets v2: More deformable, better results","cited_arxiv_id":null,"evidence_quote":"Provides DCNv2 deformable convolution with modulation offsets, the mechanism that gives the decoder adaptive distortion awareness."},{"cited_title":"Unified perceptual parsing for scene understand- ing","cited_arxiv_id":null,"evidence_quote":"The widely used UperHead decoder that serves as the main baseline and compute bottleneck in the efficiency comparison."},{"cited_title":"Segnext: Rethink- ing convolutional attention design for semantic segmenta- tion","cited_arxiv_id":null,"evidence_quote":"A pinhole-designed SegNeXt baseline that the paper re-implements to show narrow-FoV models degrade on wide-FoV inputs."},{"cited_title":"Semantics distortion and style matter: Towards source-free uda for panoramic segmentation","cited_arxiv_id":null,"evidence_quote":"The panoramic 360SFUDA baseline re-implemented for comparison on 360- and 180-degree datasets."},{"cited_title":"Bending reality: Distortion-aware transformers for adapting to panoramic se- mantic segmentation","cited_arxiv_id":null,"evidence_quote":"A distortion-aware transformer for panoramas whose whole-architecture coupling motivates the decoder-level decoupling."},{"cited_title":"Woodscape: A multi-task, multi-camera fisheye dataset for autonomous driving","cited_arxiv_id":null,"evidence_quote":"The WoodScape fisheye dataset used to demonstrate gains on 180-degree images."}],"review_version":1}