{"id":"5e2ed4fc-61f5-4c88-bcb7-1c78d18b0f57","arxiv_id":"2502.10444","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A comprehensive survey organizes deep learning for omnidirectional 360-degree vision into representation learning, optimization strategies, tasks, and applications, with benchmark tables and an open-source repository.","lead":"This paper maps the recent deep-learning literature for 360-degree omnidirectional vision into a taxonomy: imaging and projections, representation learning, optimization strategies, tasks, and applications. It is a survey, not a new algorithm, so its value lies in whether the paper selection and benchmark tables are accurate and complete.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (1) in Sec. 2.2 gives internally inconsistent spherical-to-Cartesian and inverse mappings, undermining the paper's foundational projection exposition.","rationale":"The reader identified the undocumented selection process as the weakest assumption. That is a legitimate concern for any survey claiming comprehensiveness, but it is inherently difficult to falsify from the manuscript alone and could be remedied by adding a methodology section. The internal inconsistency in Eq. (1) is a sharper, fully verifiable defect: it sits in the section the paper itself advertises as a thorough introduction to imaging principles, and it affects the mathematical foundation on which projection formats and many reviewed methods are built. Because the paper is a survey rather than a derivation-heavy method paper, a localized equation error does not destroy the overall value of the taxonomy or the reference lists, so the appropriate verdict remains CONDITIONAL rather than REJECT. The reader's selection-bias concern and the equation-error concern both point to the same conclusion: the survey needs revision before it can serve as a reliable reference map. I therefore agree with the CONDITIONAL verdict but for a more concrete and load-bearing reason.","tokens_in":51417,"tokens_out":4217,"duration_ms":39080,"concrete_test":"Take θ=π/3, ϕ=π/4, ρ=1. Compute p from the paper's own definition p=[sinθ cosϕ, sinθ sinϕ, cosθ] ≈ (0.612, 0.354, 0.707). Apply Eq. (1)'s inverse: θ' = arctan(0.612/0.707) ≈ 0.713 rad (≠ π/3) and ϕ' = arccos(0.354) ≈ 1.209 rad (≠ π/4). Then apply Eq. (1)'s forward formula to the original θ,ϕ; it gives (0.612, 0.5, 0.354), not p. This demonstrates the round-trip fails and the two halves of Eq. (1) contradict each other. If the authors confirm a notation error, the same check should be repeated after replacing Eq. (1) with the standard convention and verifying that all downstream formulas (ERP, TP, and any projection-dependent discussion) remain consistent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim includes a 'thorough introduction to the principles of omnidirectional imaging and commonly explored projections.' That claim is undercut by a concrete mathematical defect in the very first projection equation. In Sec. 2.2, the paper defines the spherical coordinate of a point as p = [sinθ cosϕ, sinθ sinϕ, cosθ]^T, with θ the polar angle and ϕ the azimuth angle. Under that standard convention, the correct forward mapping is x = ρ sinθ cosϕ, y = ρ sinθ sinϕ, z = ρ cosθ, and the correct inverse is θ = arccos(z/ρ), ϕ = arctan2(y,x). Eq. (1) instead states θ = arctan(x/z), ϕ = arccos(y/ρ), and x = ρ sinθ sinϕ, y = ρ cosϕ, z = ρ cosθ sinϕ. These formulas are mutually incompatible and fail round-trip: for θ=π/2, ϕ=0, the text's definition gives p=(1,0,0), while Eq. (1)'s forward formula gives (0,ρ,0). This is not a harmless typo: the ERP mapping (u,v)→(θ,ϕ), the tangent projection in Eqs. (2)-(3), and many distortion-aware methods discussed later all rely on a correct spherical coordinate convention. A reader implementing Eq. (1) directly will obtain wrong projections. This internal inconsistency is more damaging to the survey's foundational value than the absence of a search protocol, because it is a definite correctness error, not a matter of completeness judgment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript surveys deep learning for omnidirectional vision, covering acquisition and projection formats (ERP, CP, TP, polyhedron, and others), datasets, representation learning (Euclidean and non-Euclidean), optimization strategies (unsupervised/semi-supervised, transfer, multi-task, and deep reinforcement learning), and a taxonomy of tasks ranging from visual enhancement and scene understanding to 3D geometry and motion estimation. It also discusses applications such as AR/VR, robot navigation, and autonomous driving, and concludes with challenges and future directions. The authors claim that this is the first comprehensive survey of deep learning for omnidirectional vision, summarize over 200 representative works, provide benchmark tables, and maintain an open-source repository with code links.","tokens_in":51716,"tokens_out":3799,"duration_ms":35049,"significance":"If the technical content is corrected, this survey could be a valuable unified entry point to the field. Its strengths include the hierarchical taxonomy in Fig. 2, the cross-task organization of methods, benchmark tables (Tables 2-8) that allow quick comparisons, and the openly maintained repository. The paper goes beyond listing papers; it draws useful cross-task insights, such as the trade-offs between distortion-aware convolutions and attention mechanisms, and identifies underexplored directions such as 3D robustness for panoramic segmentation and panoramic panoptic segmentation. However, the foundational projection equations in Sec. 2.2 contain a concrete mathematical inconsistency, and the claim of comprehensiveness is not backed by a documented selection methodology. Both issues must be resolved before the survey can be relied upon as an authoritative reference.","major_comments":[{"comment":"Eq. (1) is internally inconsistent with the spherical-coordinate convention stated immediately above it. The text defines p = [sinθ cosϕ, sinθ sinϕ, cosθ]^T, so θ is the polar angle and ϕ is the azimuth angle; the correct inverse is then θ = arccos(z/ρ), ϕ = arctan2(y,x), and the correct forward mapping is x = ρ sinθ cosϕ, y = ρ sinθ sinϕ, z = ρ cosθ. Instead, Eq. (1) gives θ = arctan(x/z), ϕ = arccos(y/ρ), and x = ρ sinθ sinϕ, y = ρ cosϕ, z = ρ cosθ sinϕ. These formulas do not round-trip: for θ=π/2 and ϕ=0 the stated definition gives p=(1,0,0), while the forward formula gives (0,ρ,0). Because the ERP mapping (u,v)→(θ,ϕ), the tangent projection in Eqs. (2)-(3), and many distortion-aware methods reviewed later depend on this convention, Eq. (1) must be corrected or the convention must be stated explicitly and used consistently.","section":"§2.2, Eq. (1)"},{"comment":"The central claim of being the \"first comprehensive\" survey and of covering \"over 200 representative published top-tier works\" is not supported by a systematic methodology. The manuscript does not report search databases, a time window, keywords, inclusion/exclusion criteria, screening steps, or a log of excluded works. Since the taxonomy and the state-of-the-art tables are presented as a map of the field, the authors should either document the selection process or explicitly qualify the selection as author-curated rather than comprehensive; otherwise a reader cannot assess possible selection bias.","section":"§1, Contributions (I)-(II); §2.3"}],"minor_comments":[{"comment":"The abstract says the paper covers \"four main contents\" and then lists five items, (i)-(v). This mismatch should be fixed, for example by replacing \"four\" with \"five\" or by merging two items.","section":"Abstract / §1"},{"comment":"In the quantitative VQA comparison, the row for Assessor360 is cited as [72], but reference [72] is the AHGCN paper; the text identifies Assessor360 as [251]. This citation duplication should be corrected so that the table entries can be traced.","section":"Table 3"},{"comment":"References [342] and [343] are the same work (\"Spherical view synthesis for self-supervised 360 depth estimation\"), yet the text in §5.3.1 cites both as if they were distinct. The duplicate entry should be removed and the citations aligned.","section":"References [342] and [343]"},{"comment":"The data size for Deep360 is listed as \"1,2000 (RGB)\"; this appears to be a typo for 12,000. Please verify and correct the entry.","section":"Table 1"},{"comment":"The saliency method SalGAIL is attributed to Ma et al. [86], but reference [86] is the generative adversarial imitation learning paper; the actual SalGAIL work appears to be [265]/[266]. The citations should be re-checked and assigned to the correct publications.","section":"§5.2.3"}],"recommendation":"major_revision","confidential_remarks":"The Eq. (1) error is likely to be noticed by any reader who attempts to implement the formulas, and it directly undermines the survey's foundational value. The comprehensiveness claim also needs either a documented methodology or a more modest framing. I recommend major revision, not rejection, because the overall organization and benchmark compilation are useful and the defects appear fixable within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you want a map of deep learning for 360° images. The survey covers more ground than prior surveys—representation learning, optimization strategies, and tasks from generation to SLAM—with a reasonable taxonomy in Fig. 2 and SOTA tables compiled from the literature. And there's a real correctness problem: Eq. (1) in Sec. 2.2, the paper's foundational spherical-to-Cartesian mapping, is internally inconsistent. The paper first defines p = [sinθ cosϕ, sinθ sinϕ, cosθ]^T, then gives inverse θ=arctan(x/z), ϕ=arccos(y/ρ), and forward x=ρ sinθ sinϕ, y=ρ cosϕ, z=ρ cosθ sinϕ. These don't round-trip; for θ=π/2, ϕ=0 the definition gives (1,0,0) while the forward formula gives (0,ρ,0). That is not a harmless typo—the ERP mapping and tangent projection in Eqs. (2)-(3) build directly on this, and any reader implementing Eq. (1) will get wrong projections.\n\nThe actual novelty is the synthesis, not any new method or data. The paper organizes roughly 200 papers into a hierarchical taxonomy and provides benchmark tables across depth, layout, segmentation, VQA, and more. For a newcomer, that is genuinely useful. The discussion of optimization strategies (unsupervised, transfer, multi-task, RL) as a cross-cutting theme is a good framing, and the open-source repository with code links is a nice addition. I give real credit for the breadth.\n\nSoft spots: the claimed 'first comprehensive survey' is asserted, not demonstrated. No search protocol, inclusion/exclusion criteria, or screening log is given. For a survey whose value is its coverage, that is load-bearing—the taxonomy and SOTA tables are only as good as the selection. There are also concrete editing errors: the abstract says 'four main contents' and then lists five; refs [342] and [343] are the same paper, and there are likely other duplicates. These are fixable, but they undermine trust.\n\nWho is this for? Graduate students and researchers entering 360° vision, or anyone needing a quick orientation across tasks. It deserves a serious referee, but the referee should insist on fixing Eq. (1), adding a methodology section, and cleaning the reference list. As is, I wouldn't cite it yet; after revision I might.","headline":"A genuinely broad 360° vision survey with a useful taxonomy and benchmark tables, but the spherical projection equation in Sec. 2.2 is internally inconsistent and the 'comprehensive' claim lacks a documented selection process.","tokens_in":52216,"tokens_out":3147,"would_cite":false,"duration_ms":28385,"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":"This survey claims to be the first comprehensive map of deep learning for omnidirectional vision, organizing imaging principles, projection formats, datasets, representation learning, optimization strategies, and applications into one…","keywords":["omnidirectional vision","360-degree images","deep learning survey","spherical representation learning","equirectangular projection","omni-directional image tasks","panoramic scene understanding","representation learning"],"falsifier":"A reader can test the map's completeness with a fixed protocol: enumerate every paper published in a defined set of major vision venues over the last five years whose title or abstract matches 360-degree, panoramic, or omnidirectional deep learning, then check whether the survey's taxonomy and benchmark tables systematically omit any of them. The claim of being the first comprehensive survey can be checked directly by searching for any earlier review that already covers representation learning, optimization, and applications for omnidirectional vision together.","tokens_in":51198,"feed_emoji":"🌐","tokens_out":4124,"duration_ms":40484,"temperature":0.7,"pith_summary":"The paper sets out to establish that deep learning for omnidirectional (360-degree) vision can be organized into a single coherent structure: how 360-degree images are captured and projected, how representations are learned from them, how models are optimized when annotations are scarce, and how the resulting methods sort into tasks from generation and super-resolution to depth, optical flow, and room layout estimation. The authors claim this is the first survey to cover all of these together, rather than a single sub-domain as earlier surveys did. A sympathetic reader's takeaway is a previously unavailable reference map: over 200 recent works placed in a taxonomy, benchmark tables for direct comparison, discussion of open problems, and a maintained code repository to make the map usable.","feed_headline":"A first unified map of deep learning for 360° vision","feed_subtitle":"The survey organizes over 200 methods into one taxonomy, with benchmark tables and code links.","key_machinery":"The carrying object is the hierarchical taxonomy (Fig. 2 of the paper), which classifies methods along three axes: representation learning (Euclidean methods on ERP-style planes versus non-Euclidean methods on spherical meshes and graphs), optimization strategy (unsupervised and semi-supervised, transfer, multi-task, deep reinforcement learning), and task family (generation, super-resolution, quality assessment, detection, segmentation, saliency, depth, optical flow, room layout, SLAM). It rests on the imaging-projection framework of Sec. 2, where the sphere is the native domain and the equirectangular projection is the default planar format whose pole distortion is the recurring problem, and on the per-task comparison tables that rank representative methods. The taxonomy does the argumentative work: it turns many individual papers into a map and supports the paper's claims about which directions are settled and which are open.","core_discovery":"On the paper's own terms, the central claim is that the development of deep learning for omnidirectional vision has matured enough to be reviewed end to end, and that it can be organized by three cross-cutting concerns: representation learning on spherical data (Sec. 3), optimization strategies beyond plain supervision (Sec. 4), and task families (Sec. 5, from visual enhancement through scene understanding to 3D geometry and motion estimation). The paper asserts that this is the first comprehensive review of that development, that its roughly 200 cited works are representative of top-tier output in the last five years, and that its hierarchical taxonomy plus quantitative tables give readers intra-task comparisons that were not previously collected in one place. If the paper is right, the field now has a unified entry point and a shared map of where it stands.","pith_inferences":["My reading of the paper's own comparison tables is that the winning recipe in monocular depth estimation combines less-distorted projections (tangent rather than cubemap) with attention or transformer backbones, a template that plausibly generalizes to other spherical dense-prediction tasks.","The novelty claim of being the first comprehensive survey is directly testable: a systematic literature search for any earlier review that already unites representation learning, optimization, and applications for omnidirectional vision would settle it; the paper offers no search protocol to back the claim.","Because no inclusion or exclusion criteria are given for the roughly 200 works, the taxonomy's neutrality cannot be verified from the paper alone; annotating each entry with venue, year, and selection basis would make the map auditable.","The paper frames spherical transformers mostly as a future direction, yet its own tables show transformer-based methods leading in classification and depth estimation, suggesting this direction will grow faster than the survey's cautious framing implies."],"forward_implications":["A newcomer to 360-degree vision can use the taxonomy to locate any method by its projection choice, learning strategy, and task, and the benchmark tables to compare reported results without reading dozens of papers.","The paper's structure makes the field's recurring trade-offs explicit, especially the tension between ERP's convenience and its pole distortion, and between planar projections' compatibility with pretrained models and their discontinuity, which future method design must navigate.","The identified open problems, such as data-efficient learning, panoramic optical aberration correction, multi-modal spherical understanding, and robustness to adversarial attacks, point to where the next wave of work is most likely to land.","The maintained open-source repository with code links gives the community a single access point to implementations, lowering the barrier to reproducing and building on existing work."],"supporting_citations":[{"why":"Earlier survey of Manhattan room layout reconstruction from single 360-degree images; it is the domain-restricted precedent the paper extends and contrasts against for the layout task.","marker":"[346]"},{"why":"Earlier survey of 3D scene geometry estimation from 360-degree imagery; it defines the sub-domain scope the paper broadens into a field-wide review.","marker":"[48]"},{"why":"Earlier review of panoramic imaging systems and optical technologies; it supplies the imaging-principle background the survey builds on.","marker":"[74]"},{"why":"Earlier survey of 360-degree video and image transmission, compression, and quality assessment; it anchors the quality-assessment discussion the paper contrasts with.","marker":"[261]"},{"why":"Earlier survey of top-view omnidirectional imaging applications; another domain-specific precedent that supports the paper's claim to be the first field-wide review.","marker":"[290]"},{"why":"SphereNet, the foundational distortion-aware spherical convolution work; it grounds the representation-learning section and appears as a baseline in benchmark tables.","marker":"[45]"},{"why":"PanoSwin, the transformer-based representation method the paper identifies as holding current state-of-the-art classification results on spherical benchmarks.","marker":"[144]"}],"fun_headline_variants":["First unified map of deep learning for 360° vision","How deep learning tackles omnidirectional vision","A survey of AI for 360° imagery: from pixels to 3D","Deep learning for panoramic vision: a structured review","Mapping 200+ methods in 360° vision deep learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole map stands on the assumption that the roughly 200 cited works were selected to be representative of top-tier research in the field, so that the taxonomy and the state-of-the-art tables are a fair picture rather than an author-selected subset.","fun_headline_variants_meta":{"raw":{"variants":["First unified map of deep learning for 360° vision","How deep learning tackles omnidirectional vision","A survey of AI for 360° imagery: from pixels to 3D","Deep learning for panoramic vision: a structured review","Mapping 200+ methods in 360° vision deep learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1357,"prompt_tokens":958,"completion_tokens":399,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":318}},"tokens_in":574,"tokens_out":399,"duration_ms":4530,"temperature":1.0,"reasoning_tokens":318,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:03:11.985971+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader can test the map's completeness with a fixed protocol: enumerate every paper published in a defined set of major vision venues over the last five years whose title or abstract matches 360-degree, panoramic, or omnidirectional deep learning, then check whether the survey's taxonomy and benchmark tables systematically omit any of them. The claim of being the first comprehensive survey can be checked directly by searching for any earlier review that already covers representation learning, optimization, and applications for omnidirectional vision together.","supporting_citations":[{"cited_title":"State-of-the-art in 360◦ video/image processing: Percep- tion, assessment and compression.IEEE J-STSP, 2020","cited_arxiv_id":null,"evidence_quote":"Earlier survey of 360-degree video and image transmission, compression, and quality assessment; it anchors the quality-assessment discussion the paper contrasts with."},{"cited_title":"Ap- plications of deep learning for top-view omnidirectional imaging: A survey.CVPR Workshop, 2023","cited_arxiv_id":null,"evidence_quote":"Earlier survey of top-view omnidirectional imaging applications; another domain-specific precedent that supports the paper's claim to be the first field-wide review."}],"review_version":1}