{"id":"f4fc4499-701d-40ac-a283-426b7246bab8","arxiv_id":"2507.09216","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A spherical input resampling trick transfers planar pretrained backbone weights to 360-degree segmentation, with a channel-attention branch adding moderate gains.","lead":"This paper describes a pixel rearrangement method that lets standard image networks, trained on ordinary photos, work on 360-degree panoramas without extra distortion-correction modules. The authors test this on indoor scene segmentation and report competitive scores on the Stanford2D3D dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3.2's 4x4 spherical kernel is geometrically inverted: the labeled 'inner' circle has radius 4π/W while the 'outer' circle has radius 4π/(3W), so the outer ring is 3x closer to the pole. The construction as written cannot produce a valid planar-compatible kernel.","rationale":"I read the paper's central claim as twofold: (1) the spherical sampling of Sec. 3.2 allows a planar pretrained backbone to be used directly on ERP images, and (2) this yields state-of-the-art segmentation on Stanford2D3D. The reader rejected on the grounds that the full model is never tested at 512x1024. I agree that is a serious evidence gap, but I found an even more basic problem in the core construction. In Sec. 3.2 the inner and outer circles of the 4x4 kernel have radii 4π/W and 4π/(3W) respectively. Since 4π/W > 4π/(3W), the ring labeled 'outer' is actually closer to the north pole than the ring labeled 'inner'. The geometry described is impossible: the 12 'outer' points would sit inside the 4 'inner' points. Consequently, the kernel cannot be assembled into a valid 4x4 matrix that matches the ordering assumed by the planar ConvNeXt weights, undercutting the 'direct utilization' claim at its foundation. This is not a criticism of the author; it may be a typographical error in the radius expression, but as written the method is not reproducible. If the intended radii were e.g. 4π/W and 6π/W, the method might be sound, but that is not what the text says. The concrete test of reimplementing the sampling and measuring point distances would quickly settle whether the inversion is real or merely a typo. Either way, the current manuscript does not support acceptance, so the REJECT verdict stands.","tokens_in":899,"tokens_out":1026,"duration_ms":110069,"concrete_test":"Implement the Sec. 3.2 sampling for a 4x4 kernel on a sphere with W=512 and compute the great-circle angular distance of each of the 16 points from the kernel center. Verify that the four points designated 'inner' are closer to the center than the twelve designated 'outer'. If the outer ring is closer (as the printed radii imply), the kernel geometry is invalid; rerun the Stanford2D3D experiment with a corrected sampling (outer radius > inner radius) and check whether the mIoU improvements in Table 1 survive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that spherical sampling lets unmodified planar pretrained ConvNeXt weights run on panoramas. Section 3.2 specifies the 4x4 kernel sampling: an inner circle radius 2π/W·2 (i.e., 4π/W) with 4 points at π/4 spacing, and an outer circle radius 2π/(3W/2) (i.e., 4π/(3W)) with 12 points at π/12 spacing. On a sphere, the inner circle must have smaller angular radius than the outer; here the outer radius is one third of the inner. The sampling pattern is therefore inverted, so the 16 points cannot be arranged into a 4x4 kernel that preserves the planar kernel's spatial ordering. Because every downstream claim — direct use of pretrained weights, no extra distortion modules, superior segmentation — rests on this sampling geometry, the method as described is not implementable. This is independent of the separate issue that the 512x1024 result is from an incomplete model (Sec. 4.4).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spherical sampling method that rearranges equirectangular panoramic images so that unmodified planar pretrained ConvNeXt models can be applied 'directly' without additional distortion modules. The method constructs spherical convolution kernels by selecting points on a sphere and back-projecting them to the ERP plane, with the mapping precomputed as a lookup table. For semantic segmentation, the authors add a spherical branch whose features generate per-channel attention masks that modulate the planar backbone's predictions. Experiments are reported on Stanford2D3D at 256x512 and 512x1024 resolutions, claiming improved or competitive mIoU against several prior methods.","tokens_in":8868,"tokens_out":3894,"duration_ms":46407,"significance":"If the proposed construction were valid, the idea of reusing large planar pretrained backbones for panoramic segmentation through a fixed spherical resampling would be practically appealing: it avoids additional trainable distortion modules and could make panoramic fine-tuning cheaper. The offline precomputable lookup table is a sensible engineering choice, and the MNIST ablation (Section 4.4) provides some evidence that pretrained weights help within the spherical framework. However, the central geometric construction as written is internally inconsistent, and the headline 512x1024 results come from an incomplete model. These issues affect the core claims of the paper, so the current significance is not established.","major_comments":[{"comment":"The radii for the 4x4 spherical kernel are geometrically inverted. The text states that the 'inner circle' has radius 2π/W * 2 (i.e., 4π/W) and the 'outer circle' has radius 2π / (3W/2) (i.e., 4π/(3W)). Since 4π/(3W) is one-third of 4π/W, the so-called outer circle lies closer to the pole than the inner circle. A valid 4-point-inner / 12-point-outer kernel requires the outer radius to be larger than the inner radius. As written, the 16 sampled points cannot be arranged into a 4x4 kernel that preserves the spatial ordering of the planar convolution kernel, so the described method is not implementable and the claim of direct compatibility with pretrained planar weights is unsupported.","section":"Section 3.2"},{"comment":"The 512x1024 results in Table 1 are produced by an incomplete model. Section 4.1 states that the authors 'were unable to complete full model training at the 512×1024 resolution' and used an incomplete model; Section 4.4 specifies that at 512x1024 only Layer 1 of the spherical branch was used. Nevertheless, Section 4.2 compares these incomplete-model numbers against complete state-of-the-art systems and claims improvements (e.g., 2.18 over PanoFormer and 2.38 over CBFC in mIoU). This comparison does not support the paper's claim of superior performance for the full proposed method at 512x1024.","section":"Sections 4.1 and 4.4"},{"comment":"The comparison protocol for SGAT4PASS is inconsistent. Section 4.2 says that Table 1 uses the original results from SGAT4PASS, while the same section says the authors were unable to reproduce SGAT4PASS's reported performance and Section 4.3 states that the SGAT4PASS results shown in the figures were reproduced by the authors. Mixing original-paper numbers with self-reproduced numbers without a standardized evaluation protocol makes the quantitative comparison unreliable.","section":"Section 4.2 vs Section 4.3"},{"comment":"No error bars or repeated-run statistics are reported. Several claimed advantages are very small (e.g., 0.11 mIoU over CBFC at 256x512), and on a dataset of 1,413 images such differences may be within random variation. Without variance estimates, the 'superior performance' claim is not statistically supported.","section":"Table 1 and Section 4.2"}],"minor_comments":[{"comment":"Equation (1) is not properly formatted and its variables are not defined; please clarify the projection formula, the meaning of (x', y', z'), and the convention for arctan2.","section":"Equation (1)"},{"comment":"The phrase 'the 2.5th row and column as a center' is unclear; specify how the fractional row/column index corresponds to a spherical coordinate.","section":"Section 3.2"},{"comment":"References [3] and [17] are duplicates of the same paper (Tateno et al., ECCV 2018); please merge or remove the duplicate.","section":"References"},{"comment":"There are numerous typographical and grammatical errors, such as 'futhermore', 'each blocks is formed', and 'we proposes'; the manuscript would benefit from thorough proofreading.","section":"Throughout"},{"comment":"The header 'Base layer1 layer2 layer3 layer4 Cat' is ambiguous; please label the columns clearly and describe what each checkmark indicates.","section":"Table 2"},{"comment":"The MNIST ablation is not clearly connected to spherical sampling; explain how planar MNIST digits are mapped to the sphere and what 'spherical pre-training model' means in this experiment.","section":"Section 4.4"}],"recommendation":"reject","confidential_remarks":"The core idea is timely, but the manuscript has a load-bearing geometric error in Section 3.2 and the primary high-resolution results are from an incomplete model. Correcting these would require re-deriving the kernel construction and rerunning the full 512x1024 experiments, which is beyond a routine revision. The SGAT4PASS comparison inconsistency and the lack of statistical validation of small mIoU differences further weaken the empirical claims. I recommend rejection, though the authors could consider resubmission after a substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: the paper's premise — that you can resample an equirectangular image so that stock planar ConvNeXt kernels sample spherical neighborhoods, and thereby reuse pretrained weights on 360° data — is a legitimate and practically useful idea. The specific packaging (precomputed lookup table, kernel-size-specific layouts, channel-wise attention masks) is a real combination not in the prior work they cite. And they are honest enough to state that their 512×1024 result was produced by an incomplete model.\n\nUnfortunately, the geometry as written does not hold together. In Section 3.2, for the 4×4 kernel, they define an inner circle radius 2π/W·2 and an outer circle radius 2π/(3W/2). That makes the inner radius 4π/W and the outer radius 4π/(3W), so the \"outer\" ring sits three times closer to the pole than the \"inner\" ring. On a sphere, that inverts the arrangement and cannot correspond to a 4×4 kernel that preserves the planar ordering. Since every downstream claim — direct weight reuse, no extra modules, superior segmentation — rests on this sampling construction, the method as described is not implementable. This is not a minor typo; it needs to be corrected and the corrected geometry verified.\n\nThe experimental evidence is also weaker than the abstract implies. The 512×1024 comparison uses an incomplete model while the full-method claim is supported only by an ablation at 256×512. The SGAT4PASS baselines are taken from the original paper while others are their own replications, mixing protocols. There are no error bars, and the choice of layer 3 for the attention branch appears to be selected based on the same test set.\n\nWhat they do well: the MNIST ablation, although on a non-panoramic task, shows a large jump from pretrained weights (66% to 96%), which suggests the resampling is not destroying the pretrained weights. And the 256×512 results are competitive with strong baselines, even if the gains over CBFC are small.\n\nThis paper is for researchers working on 360° perception who want to reuse pretrained backbones without architectural changes. The idea is worth discussing, but the current manuscript needs a major revision: fix the geometry, rerun 512×1024 with the full model, align the baselines, and release code.\n\nRecommendation: this deserves a serious referee, but heavy revision is required. I would not cite it in its current form.","headline":"A plausible idea for reusing planar pretrained models on 360° imagery, but the spherical sampling geometry is internally inconsistent and the headline experiment uses an incomplete model.","tokens_in":9370,"tokens_out":3356,"would_cite":false,"duration_ms":38239,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a fixed spherical resampling of an equirectangular panorama, computed only from the image width, lets an unmodified planar ConvNeXt pretrained on ordinary photos reach state-of-the-art semantic segmentation on…","keywords":["spherical convolution","panoramic image segmentation","pretrained models","equirectangular projection","distortion elimination","ConvNeXt","Stanford2D3D","attention mechanism"],"falsifier":"A reader could run the full model at 512x1024 on Stanford2D3D with all four spherical downsampling layers and compare its mIoU against (a) the reported 54.58, (b) a plain ConvNeXt trained on the original ERP image with the same decoder, and (c) SGAT4PASS's 55.30; if the spherical branch does not beat the plain ERP baseline, the claimed distortion correction is not producing the reported gains.","tokens_in":8443,"feed_emoji":"🌐","tokens_out":7097,"duration_ms":73907,"temperature":0.7,"pith_summary":"Panoramic images stretch and tear at the poles, so networks pretrained on ordinary perspective photos lose accuracy when applied to 360-degree input. This paper claims that a fixed rearrangement of the panorama's pixels, computed from the equirectangular image width alone, removes that distortion before the network sees the image. Because the rearrangement happens in the input, the unmodified weights of a planar pretrained backbone such as ConvNeXt can be applied directly, avoiding the need for million-scale panoramic pretraining. The paper demonstrates this on indoor semantic segmentation, reporting state-of-the-art mIoU on Stanford2D3D at $256\\times512$ and competitive results at $512\\times1024$ with an incomplete model.","feed_headline":"Spherical pixel repacking lets 2D pretrained models segment panoramas","feed_subtitle":"A fixed resampling of equirectangular images removes pole distortion, so ConvNeXt's original weights transfer directly to 360-degree…","key_machinery":"The carrying mechanism is a spherical discrete sampling of the equirectangular input: for the $4\\times4$ convolution, two circles centered at the north pole, with radii $\\frac{2\\pi}{W}\\cdot 2$ and $\\frac{2\\pi}{3W/2}$ and four and twelve sampled points respectively, are rotated to the kernel's center and back-projected to ERP coordinates; for $7\\times7$ convolution the image is expanded sevenfold and three circles with spacings $\\pi/8$, $\\pi/16$, $\\pi/24$ are used, with the stride changed to 7. These point positions can be precomputed and stored as a lookup table, so adapting the pretrained model costs only bilinear interpolation at runtime. The second component is a dual-branch attention head that computes per-channel masks $m_i$ from spherical features and fuses them into the planar predictions as $\\mathrm{sem} = \\mathrm{ReLU}(\\mathrm{sem}_i + \\mathrm{sem}_i \\ast m_i)$.","core_discovery":"The central claim is that planar convolution kernels can be made spherical by resampling the ERP image itself. For each kernel size the paper selects spherical points around the kernel's center, with circle radii derived from the image width $W$ and point spacings of $\\pi/4$ and $\\pi/12$ for $4\\times4$ kernels and $\\pi/8$, $\\pi/16$, $\\pi/24$ for $7\\times7$ kernels, projects those points back to ERP coordinates, and rearranges the sampled pixels into the original kernel-shaped matrix. The pretrained ConvNeXt weights are then used unchanged, because the network's view of each neighborhood is spherical rather than planar. On Stanford2D3D the full dual-branch model achieves 53.91 mIoU with RGB-D and 52.03 mIoU with RGB at $256\\times512$, while an incomplete model at $512\\times1024$ reports 54.58 mIoU, surpassing several published baselines while remaining below SGAT4PASS's 55.30.","pith_inferences":["A natural next test is applying the same input resampling to ResNet or ViT backbones; the paper's recipe is tied to ConvNeXt's specific kernel sizes, so the generality of 'planar pretrained models' remains an open extension.","If the resampling is as cheap as a precomputed lookup table plus bilinear interpolation, the method could also be dropped into other dense 360-degree tasks such as depth estimation or object detection without retraining the backbone.","The reported $512\\times1024$ result comes from a deliberately incomplete model using only the first spherical feature layer, so the full-model comparison at the standard resolution is a direct, feasible check of whether the claimed direct utilization holds at scale."],"forward_implications":["If the spherical resampling works as claimed, any pretrained planar backbone can be adapted to 360-degree data by resampling the input image, with no architectural changes and no extra branch for distortion removal.","The reported 66%-to-96% jump on MNIST when pretrained weights are added suggests the pretrained weights remain usable after spherical resampling, and that the same recipe may transfer to other planar backbones.","At $256\\times512$, the full model outperforms PanelNet by 5.73 mIoU in RGB input and PanoFormer by 5.01 mIoU in RGB-D input, indicating the dual-branch attention adds value beyond the spherical backbone alone.","At $512\\times1024$, even the incomplete model reports mIoU above PanoFormer, CBFC, SFSS, Trans4PASS+, and 360BEV, though below SGAT4PASS, so the method is positioned as a strong RGB-only panorama segmentation baseline."],"supporting_citations":[{"why":"Supplies the ConvNeXt architecture and its pretrained weights, the planar backbone the spherical sampling is designed to reuse unchanged.","marker":"[28]"},{"why":"Provides the spherical-kernel construction idea and the earlier equirectangular/spherical fusion that the paper adapts for the 4x4 and 7x7 kernels.","marker":"[22]"},{"why":"Supplies the Stanford2D3D dataset on which all segmentation comparisons are run.","marker":"[13]"},{"why":"PanelNet is the RGB baseline compared at 256x512, where the full model reports a 5.73 mIoU gain.","marker":"[19]"},{"why":"PanoFormer is an RGB-D and RGB baseline surpassed in the reported comparisons.","marker":"[15]"},{"why":"CBFC is the recent fusion-based baseline the method compares against in both resolutions.","marker":"[11]"},{"why":"SGAT4PASS is the state-of-the-art baseline whose published numbers set the comparison target at 512x1024.","marker":"[12]"},{"why":"Trans4PASS+ is the distortion-aware transformer baseline compared at 512x1024.","marker":"[16]"}],"fun_headline_variants":["Resampling equirectangular images adapts 2D nets to 360° segmentation","Spherical sampling lets pretrained 2D models handle panoramas","ConvNeXt weights transfer to spherical segmentation via pixel repacking","Panorama segmentation bypasses distortion with weight-aware sampling","New sampling maps planar kernels to sphere for 360° vision"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the fixed spherical resampling computed from the panorama width alone maps standard planar convolution kernels onto spherical neighborhoods faithfully enough that unmodified pretrained ConvNeXt weights keep working for segmentation, a claim the paper never tests with the complete model at 512x1024, the resolution where it reports comparisons against complete state-of-the-art systems.","fun_headline_variants_meta":{"raw":{"variants":["Resampling equirectangular images adapts 2D nets to 360° segmentation","Spherical sampling lets pretrained 2D models handle panoramas","ConvNeXt weights transfer to spherical segmentation via pixel repacking","Panorama segmentation bypasses distortion with weight-aware sampling","New sampling maps planar kernels to sphere for 360° vision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000711,"raw_usage":{"total_tokens":3182,"prompt_tokens":909,"completion_tokens":2273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":2182}},"tokens_in":525,"tokens_out":2273,"duration_ms":18367,"temperature":1.0,"reasoning_tokens":2182,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:01:12.432770+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could run the full model at 512x1024 on Stanford2D3D with all four spherical downsampling layers and compare its mIoU against (a) the reported 54.58, (b) a plain ConvNeXt trained on the original ERP image with the same decoder, and (c) SGAT4PASS's 55.30; if the spherical branch does not beat the plain ERP baseline, the claimed distortion correction is not producing the reported gains.","supporting_citations":[{"cited_title":"A convnet for the 2020s,","cited_arxiv_id":null,"evidence_quote":"Supplies the ConvNeXt architecture and its pretrained weights, the planar backbone the spherical sampling is designed to reuse unchanged."},{"cited_title":"Estimating depth of monocular panoramic image with teacher-student model fusing equirectangular and spherical representations,","cited_arxiv_id":null,"evidence_quote":"Provides the spherical-kernel construction idea and the earlier equirectangular/spherical fusion that the paper adapts for the 4x4 and 7x7 kernels."},{"cited_title":"Panelnet: Understanding 360 indoor environment via panel representation,","cited_arxiv_id":null,"evidence_quote":"PanelNet is the RGB baseline compared at 256x512, where the full model reports a 5.73 mIoU gain."},{"cited_title":"Panoformer: Panorama transformer for indoor 360 depth estimation,","cited_arxiv_id":null,"evidence_quote":"PanoFormer is an RGB-D and RGB baseline surpassed in the reported comparisons."},{"cited_title":"Complementary bi-directional feature compression for indoor 360deg semantic segmentation with self-distillation,","cited_arxiv_id":null,"evidence_quote":"CBFC is the recent fusion-based baseline the method compares against in both resolutions."},{"cited_title":"Behind Every Domain There is a Shift: Adapting Distortion-aware Vision Transformers for Panoramic Semantic Segmentation","cited_arxiv_id":"2207.11860","evidence_quote":"Trans4PASS+ is the distortion-aware transformer baseline compared at 512x1024."}],"review_version":1}