{"id":"d8726cbb-7c3c-4ce7-a6f1-ee15a2767f6f","arxiv_id":"2501.13104","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A comprehensive survey that taxonomizes NeRF methods, challenges, applications, and resources, but offers no new methods or derivations.","lead":"This is a survey of Neural Radiance Fields (NeRFs), organizing hundreds of papers into a taxonomy of fundamentals, real-world challenges, applications, tools, and datasets. It is useful as a reference for researchers entering NeRF, but it does not present new scientific results.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The survey's value as a reference depends on the accuracy of its citations and technical summaries; the verified Plenoxels and Eq. (9) misattributions suggest a verification gap that could affect many entries.","rationale":"I read the survey in good faith and found it to be a broad, well-organized compilation that covers a wide range of NeRF topics with a generally sensible structure. The central claim is that it provides a unified taxonomy connecting technical progress to real-world challenges, and that it serves as an up-to-date reference. The reader's weakest assumption—that the survey's utility depends on the fidelity of its summaries and taxonomy—is exactly the right point to stress. The two concrete errors (Plenoxels citation in Table 3, Eq. (9) misattribution in Section 4.1.2) are verified in the text and are not fabricated. They are individually minor, but together they raise the possibility that the authors did not consistently verify details against primary sources. This is the most load-bearing concern because it directly affects whether the survey can be trusted as a reference, which is the entire point of the contribution. If a systematic audit reveals a low error rate, the concern dissolves and the conditional verdict is justified. If the error rate is high, the taxonomy itself may rest on shaky ground, and even a novel organizational scheme would not rescue the survey's broken utility. I considered whether the 'first unified taxonomy' claim is more load-bearing, but that is an external comparison that cannot be settled from the manuscript alone; the fidelity issue is internal, concrete, and actionable. The proposed audit is straightforward and would give a quantitative basis for deciding between acceptance and rejection. Since the reader already recommended CONDITIONAL and my analysis supports that, I see no reason to change the verdict, hence 'UNCHANGED'.","tokens_in":50242,"tokens_out":6115,"duration_ms":59954,"concrete_test":"Select a random sample of 30 method descriptions from Sections 2–5 (six from each of Fundamentals, Real-World Challenges, Reconstruction, Other Applications, and Tools/Datasets). Retrieve the original paper for each. Verify (1) the cited reference number matches the correct paper; (2) any equation numbers or formulas cited correspond to the described method; and (3) the core mechanism (loss, architecture, or key idea) is accurately summarized. Count substantive inaccuracies beyond trivial typos. If more than 2 of 30 are substantive, the survey's reliability as a reference is questionable and the value of the taxonomy drops correspondingly. If 0–2, the concern is minor and the conditional verdict stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that this survey provides an up-to-date, reliable reference for NeRF research through a unified taxonomy. That claim stands or falls on the fidelity of its technical descriptions and citations. The two concrete errors identified by the reader are symptomatic: Table 3 lists 'Plenoxels [94]' where [94] is HDR-Plenoxels, not the original Plenoxels [58], and Section 4.1.2 attaches 'spatially varying kernel sizes τ (in Eq. (9))' when Eq. (9) is the NeuS density-to-SDF mapping, unrelated to kernel sizes. These are not merely stylistic; they direct readers to the wrong reference or equation and indicate that at least some entries were not cross-checked against the original publications. The load-bearing risk is not that a citation is wrong, but that the error rate across the approximately 300 cited works is non-trivial. If descriptions are routinely compiled from abstracts or secondary sources, the survey's taxonomy, however logical, would organize unreliable content, and the claimed utility as a reference would be undermined. This concern is internal and testable, unlike the external 'first taxonomy' novelty claim, which depends on comparing with surveys not fully reproduced here.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript is a comprehensive survey of Neural Radiance Fields (NeRFs), organized around four main axes: fundamentals of the NeRF pipeline (sampling, encoding, radiance-field estimation, volume rendering), real-world challenges (degraded views, sparse views, inaccurate poses, lighting, in-the-wild conditions, complex scenes, uncertainty, generalization), applications (reconstruction, robotics, recognition, generation/editing), and tools/datasets. The authors claim this is the first unified taxonomy that connects NeRF technical developments to real-world deployment challenges and evaluation protocols. The core mathematical presentation (Eqs. 1-4) is correct, and the survey covers a very broad corpus of recent work, which would make it a useful reference if the individual method descriptions and citations are reliable.","tokens_in":50473,"tokens_out":3591,"duration_ms":41459,"significance":"If the survey's descriptions are faithful, it fills a genuine gap: existing NeRF surveys are either broader (neural rendering, neural fields) or narrower in scope, and none systematically links technical progress to deployment-related challenges such as degraded inputs, pose noise, uncertainty, and scene complexity. The survey also provides useful curated resources (Table 2, tools) and a structured discussion of open problems. The central value is therefore as a reference work, which means the accuracy of every method-reference and equation-reference pair is load-bearing. The authors have made the right high-level choices in scope and organization, but the verified citation and cross-reference errors described below indicate that a systematic verification pass is needed before the survey can serve that reference role reliably.","major_comments":[{"comment":"The entry 'Plenoxels [94]' is incorrect: reference [94] is HDR-Plenoxels, not the original Plenoxels paper. The original Plenoxels is [58], which is correctly cited in Sections 2.3.4 and 3.6.3. This is not a purely cosmetic issue, because a reader using Table 3 to locate the leading method will be directed to the wrong paper. I ask the authors to correct this entry and, more importantly, to audit every entry in Tables 2 and 3 against the actual cited papers, since the table is one of the most-used parts of a survey.","section":"Table 3, row 'Static / Novel View Synthesis'"},{"comment":"The sentence 'Wang et al. [262] utilize spatially varying kernel sizes τ (in Eq. (9))' is factually wrong: Eq. (9) is the NeuS density-to-SDF mapping, where τ is the sharpness parameter of the logistic sigmoid in NeuS, and it has no connection to the spatially varying kernel sizes in Adaptive Shells [262]. This cross-reference sends the reader to an unrelated equation and undermines confidence in the technical fidelity of the surrounding discussion. The sentence should either be rewritten without the Eq. (9) reference or the correct equation from [262] should be provided.","section":"Section 4.1.2, Eq. (9)"},{"comment":"The survey's central claim of being a reliable 'first unified taxonomy' depends on the accuracy of roughly 300 method descriptions. The two verified mismatches above (Table 3 and Section 4.1.2) are concrete evidence that at least some entries were not cross-checked against the original publications. I therefore request a full systematic pass over method-reference pairs and equation references before publication, with particular attention to tables and inline cross-references. This is a load-bearing requirement for the survey's utility, not a stylistic preference.","section":"Section 1.1 and passim"}],"minor_comments":[{"comment":"The view direction d is written as an element of R^2, but a view direction is naturally a unit vector in R^3 (S^2). The accompanying text's '5D coordinate (three for position and two for view direction)' suggests the authors are thinking of a two-dimensional angular parameterization, but writing d ∈ R^2 is dimensionally inconsistent with the later discussion of the MLP input. Please clarify the notation, for example by writing d ∈ S^2.","section":"Eq. (1)"},{"comment":"The caption uses the symbols '!' and '%' extensively in the resolution columns, but their meaning is not defined in the caption or in the table notes. Please add an explicit legend explaining what '!' and '%' represent.","section":"Table 2 caption"},{"comment":"References [186] and [187] are duplicate entries for the same CLIP paper (Radford et al., 2021). This is visible in Sections 5.2 and 5.3, where the same work is cited twice with different numbers. Please consolidate them or cite them consistently.","section":"References [186] and [187]"},{"comment":"The sentence listing camera-trajectory parameterizations cites [147] twice (Lie algebra and cubic B-spline) and then [112] and [111]. It would be clearer to cite each method once and to specify which work uses which parameterization, since the current phrasing suggests repeated citation of the same work for two different options.","section":"Section 3.1, blurriness paragraph"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a well-structured survey with a plausible and useful scope, and the core NeRF equations are correct. However, the two verified citation/equation errors are concerning because the survey's value is precisely as a reliable pointer into the literature. I do not believe these errors require rejection, but they do require the authors to audit the entire manuscript systematically; a superficial correction of just these two instances would not address the underlying verification gap. I would also encourage the editor to ask the authors to confirm that the 'first unified taxonomy' claim in Section 1.1 is supported by a direct comparison with the most recent competing surveys, rather than only by the summary in Table 1."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, well-organized NeRF survey with a genuinely useful taxonomy, but it has a couple of verified citation/equation mismatches that make me want a careful audit before I'd rely on it as a reference.\n\nThe paper does what a good survey should: it covers fundamentals, real-world challenges (degraded views, sparse views, pose inaccuracy, lighting, in-the-wild, dynamic/indoor/unbounded scenes, uncertainty, generalization), reconstruction applications, robotics/recognition/generation, and tools/datasets. The organization is clear, the coverage is current, and the decision to exclude Gaussian Splatting with a pointer to a dedicated survey is sensible. The core NeRF equations are correct, and most method descriptions I checked align with the original papers. Table 1's comparison with other surveys gives a reasonable basis for the 'first unified taxonomy' claim, even if that claim is hard to fully verify from the text alone.\n\nThe soft spots are real but localized. The stress test flagged two concrete errors, and they check out. Table 3 lists 'Plenoxels [94]' as a leading static-view method, but [94] is HDR-Plenoxels; the original Plenoxels is [58]. Section 4.1.2 says Wang et al. [262] use 'spatially varying kernel sizes τ (in Eq. (9))', but Eq. (9) is NeuS's density-to-SDF mapping where τ controls sigmoid sharpness, not a spatial kernel size. These are exactly the kind of mismatches a reader will trip over in a reference work. They also suggest some entries were compiled without cross-checking against the original papers. For a survey of ~300 references, that is a verification gap, not proof of widespread unreliability, but it is enough to warrant a systematic audit before publication.\n\nThe paper is not trying to break new technical ground, and it doesn't. Its value is as a map. For a PhD student or a researcher entering NeRF, this is a helpful entry point. For an established researcher, it's a checklist of methods and gaps. The self-citations are unremarkable for a survey. The open-challenges section is a bit generic, but that's typical.\n\nI'd recommend sending this to peer review. A serious referee can ask for a reference audit and a fix for the equation cross-reference, and after that, it would be a usable reference. My own verdict is somewhere between accept and major revision, depending on how deep the audit goes.","headline":"A broad, well-organized NeRF survey that needs a reference audit before it can be fully trusted as a map.","tokens_in":50978,"tokens_out":3740,"would_cite":true,"duration_ms":37982,"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 provide the first unified taxonomy that connects NeRF technical developments to real-world deployment challenges, evaluation protocols, applications, and datasets, and to identify open challenges and future directions.","keywords":["neural radiance fields","novel view synthesis","3D reconstruction","implicit neural representations","real-world robustness","survey","datasets and evaluation","neural rendering"],"falsifier":"Take a random sample of fifty papers cited in Sections 2 and 3, compare each paragraph description and category assignment against the cited paper's own abstract and contributions, and tally mismatches; a substantial mismatch rate would show that the taxonomy does not faithfully represent the field.","tokens_in":50053,"feed_emoji":"🗺️","tokens_out":5786,"duration_ms":58133,"temperature":0.7,"pith_summary":"The paper is a survey of neural radiance fields written for a field that has expanded rapidly since NeRF appeared in 2020. It tries to establish that the literature is best understood not as a list of incremental variants but as a challenge-driven taxonomy that pairs technical improvements with the real-world conditions that break the original formulation, such as degraded views, sparse views, inaccurate camera poses, complex lighting, unconstrained photo collections, dynamic or unbounded scenes, uncertainty, and generalization to unseen scenes. On that basis it catalogues applications in reconstruction, robotics, recognition, and 3D generation and editing, and it assembles the datasets and toolkits used across the field. A sympathetic reader would take the contribution to be an up-to-date reference map that lets a researcher match a problem to a known solution and see where the open gaps are.","feed_headline":"One taxonomy ties every NeRF fix to a real-world failure","feed_subtitle":"A new survey maps five years of NeRF research onto eight real-world challenge categories.","key_machinery":"The central instrument is the taxonomy itself, laid out in Table 1 and Figure 1. It is organized as two axes: the technical pipeline (sampling, encoding, radiance-field estimation, volume rendering, and NeRF-agnostic enhancements, plus alternative scene representations) and the set of real-world conditions that stress the original formulation (degraded views, sparse views, pose inaccuracy, complex light, in-the-wild variability, complex scene configurations, uncertainty, and generalizability). The taxonomy carries the survey's argument by converting every section into a mapping from problem to solution, and it is also what generates the paper's list of open challenges.","core_discovery":"The survey's central claim is that NeRF has grown from a single method into an ecosystem whose progress is best organized around the gap between laboratory assumptions and real-world deployment. It argues that earlier reviews cover NeRF only as one instance of neural rendering, cover only a slice of recent progress, or lack a coherent taxonomy and pay little attention to practical challenges. The paper therefore offers what it calls the first unified taxonomy connecting NeRF's technical developments to real-world deployment challenges and evaluation protocols, and it uses that taxonomy to review fundamentals, eight families of real-world challenges, reconstruction and beyond-reconstruction applications, and available resources. It also states explicitly that Gaussian Splatting is outside the survey's scope because that line has diverged from NeRF methodology and needs its own dedicated treatment.","pith_inferences":["An implication the authors leave implicit is that the challenge axis of the taxonomy could be reused as an evaluation template for any neural scene representation, including Gaussian Splatting; the taxonomy might then outlive NeRF itself.","A testable extension suggested by their gap analysis is a plug-and-play uncertainty module compatible with hash-grid encodings such as Instant NGP, since most current uncertainty methods require retraining or architectural changes.","One could check their forward-looking claims empirically by seeing whether post-survey papers concentrate in the sections the authors label open, such as dynamic in-the-wild scenes and streamable large-scale reconstruction."],"forward_implications":["Readers can locate NeRF variants by the real-world problem they solve instead of by publication timeline or architecture family.","The curated datasets and evaluation protocols give a ready-made benchmarking route for both novel view synthesis and surface reconstruction.","The stated open challenges—4K efficiency, plug-and-play uncertainty quantification, dynamic in-the-wild scenes, and streamable large-scale models—identify concrete targets for follow-up research.","The deliberate exclusion of 3D Gaussian Splatting positions NeRF and splatting as related but separate research lines, which may shape how future surveys divide the field."],"supporting_citations":[{"why":"Defines the NeRF formulation—position-and-direction to color-and-density plus differentiable volume rendering—that the entire survey is organized around.","marker":"[159]"},{"why":"Supplies the scene-contraction and proposal-model machinery used as the reference solution for unbounded and anti-aliased scenes.","marker":"[6]"},{"why":"Introduces per-image appearance and transient embeddings that anchor the survey's in-the-wild and uncertainty sections.","marker":"[151]"},{"why":"Provides the multiresolution hash encoding that anchors the feature-encoding and efficiency discussion.","marker":"[161]"},{"why":"Derives density from a signed distance function, the basis for the surface-reconstruction methods the survey reviews.","marker":"[254]"},{"why":"Introduces score distillation sampling, the foundation for the text-to-3D generation work surveyed in Section 5.3.","marker":"[176]"},{"why":"Supplies the camera-pose estimation that Section 6.1 treats as the standard preprocessing step for most NeRF training.","marker":"[209]"},{"why":"An earlier broad neural-rendering survey that this paper explicitly positions itself against in Table 1.","marker":"[235]"}],"fun_headline_variants":["NeRF survey: one taxonomy for real-world deployment hurdles","Eight real-world challenges unify five years of NeRF fixes","First unified taxonomy links NeRF fixes to real-world failures","NeRF research organized by eight real-world challenge types","Lab-to-real taxonomy maps NeRF progress to deployment problems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's value rests on the accuracy of its summaries and the correctness of its taxonomy; if a meaningful share of the cited papers are misdescribed or assigned to the wrong category, the map would mislead its readers.","fun_headline_variants_meta":{"raw":{"variants":["NeRF survey: one taxonomy for real-world deployment hurdles","Eight real-world challenges unify five years of NeRF fixes","First unified taxonomy links NeRF fixes to real-world failures","NeRF research organized by eight real-world challenge types","Lab-to-real taxonomy maps NeRF progress to deployment problems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1308,"prompt_tokens":800,"completion_tokens":508,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":416,"completion_tokens_details":{"reasoning_tokens":427}},"tokens_in":416,"tokens_out":508,"duration_ms":5426,"temperature":1.0,"reasoning_tokens":427,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:24:57.854901+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of fifty papers cited in Sections 2 and 3, compare each paragraph description and category assignment against the cited paper's own abstract and contributions, and tally mismatches; a substantial mismatch rate would show that the taxonomy does not faithfully represent the field.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Derives density from a signed distance function, the basis for the surface-reconstruction methods the survey reviews."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"An earlier broad neural-rendering survey that this paper explicitly positions itself against in Table 1."}],"review_version":1}