{"id":"a5f978c9-46a4-413a-a616-77d56f8bc919","arxiv_id":"2504.19938","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Mesh-Learner learns spherical-harmonic textures on each mesh triangle via a rasterization pipeline, achieving high-quality novel-view synthesis that is natively compatible with standard graphics software.","lead":"This paper presents Mesh-Learner, a method that stores a scene's appearance as spherical-harmonic textures on mesh triangles and trains them end-to-end with standard rasterization. The goal is photorealistic novel-view rendering that can be dropped directly into game engines and robotics simulators, without conversion steps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim is contradicted by the paper's own tables: Mesh-Learner's PSNR is below named baselines on FAST-LIVO2 and on Replica extrapolation, so the central claim is not supported as stated.","rationale":"The Reader's verdict is CONDITIONAL, and I agree that a conditional verdict is appropriate: the pipeline is coherent and the qualitative results are promising, but the headline SOTA claim is not fully substantiated. My primary load-bearing concern differs from the Reader's stated weakest_assumption. The Reader focused on the dependence on a high-quality input mesh, which is a genuine limitation. My concern is more direct: the paper's own tables contradict the strongest claim as written, because PSNR is worse than a named state-of-the-art method in two key comparisons (FAST-LIVO2 interpolation and Replica extrapolation). This is not a matter of engineering detail or external assumptions; it is an internal inconsistency between the abstract's claim and the reported numbers. The Reader's rationale did note that the SOTA claim is not fully supported by the tables, so there is partial overlap, but the weakest_assumption field emphasizes mesh dependency rather than the metric inconsistency. For the verdict, I would keep it CONDITIONAL: the work may still be a useful system, but the central claim needs to be narrowed and the code/tables need to be verified before the stronger assertion is accepted.","tokens_in":12622,"tokens_out":6803,"duration_ms":72849,"concrete_test":"With the released code, reproduce Table I and Table II using the exact train/test splits and mesh inputs. Then compute per-scene paired PSNR differences between Mesh-Learner and the best baseline (M2-Mapping on Replica; 3DGS/Textured-GS on FAST-LIVO2), and report average and variance across scenes. If Mesh-Learner does not show a PSNR advantage in a majority of scenes for both interpolation and extrapolation, the abstract's 'state-of-the-art' claim should be replaced by a claim of competitiveness with best SSIM/LPIPS on selected settings. Also run an ablation with a deliberately incomplete mesh to quantify how much of the advantage relies on the external mesh quality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is that Mesh-Learner achieves state-of-the-art performance on interpolation and extrapolation rendering for Replica and FAST-LIVO2, citing 3DGS and M2-Mapping. The paper's own quantitative results do not support this. In Table I (FAST-LIVO2), the average PSNR of Ours is 25.624, below 3DGS at 25.825 and Textured-GS at 25.856; only SSIM (0.829) and LPIPS (0.219) are best. In Table II (Replica extrapolation), the average PSNR of Ours is 37.915, below M2-Mapping at 38.226, even though Ours has the best SSIM (0.974 vs 0.967) and LPIPS (0.047 vs 0.078). Since PSNR is the first-listed fidelity metric and the abstract specifically names M2-Mapping and 3DGS, the SOTA assertion depends on selectively privileging SSIM/LPIPS while ignoring PSNR deficits. In addition, the abstract states that results cover interpolation and extrapolation on both Replica and FAST-LIVO2, but Section IV.C reports only FAST-LIVO2 interpolation; no FAST-LIVO2 extrapolation table is provided. The central claim therefore rests on inconsistent metric selection and an unstated input advantage (a high-quality mesh supplied a priori), not on a consistent quantitative superiority over the named baselines.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Mesh-Learner, a view-synthesis framework that learns spherical-harmonic (SH) textures on a given triangle mesh using a rasterization-based training pipeline. The method defines per-triangle SH textures with resolution set by a 'SH Density', renders via deferred shading plus a hybrid interpolation scheme (bilinear in interior regions and inverse-distance weighting at edges/corners), applies a world-space elliptical weighted average filter for distant views, and adaptively adjusts SH density per mesh during training. Experiments on Replica and FAST-LIVO2 compare interpolation and extrapolation rendering against Instant-NGP, 3DGS, Textured-GS, MonoGS, and M2-Mapping. The paper claims state-of-the-art performance and emphasizes compatibility with standard rasterization tools such as OpenGL and Blender.","tokens_in":13049,"tokens_out":6407,"duration_ms":60601,"significance":"The engineering contribution is timely: a mesh-texture representation that trains and renders natively in the rasterization pipeline could reduce conversion overhead for robotics and simulation applications, and the CPU/GPU texture streaming design is a practical response to memory limits. The method description is detailed and the qualitative results, particularly on Replica interpolation, are strong. However, the quantitative evidence does not establish the claimed state-of-the-art status: on FAST-LIVO2 the method is below 3DGS and Textured-GS in PSNR, and on Replica extrapolation it is below M2-Mapping. The comparison also does not isolate the benefit of the supplied high-quality mesh. The strengths are real, but the claims need to be scaled back and the missing experiments added before the paper can be accepted.","major_comments":[{"comment":"On the FAST-LIVO2 interpolation split, the average PSNR of Ours is 25.624 dB, below 3DGS (25.825 dB) and Textured-GS (25.856 dB). Since the abstract explicitly names 3DGS as a comparison and PSNR is a standard fidelity metric, the 'state-of-the-art performance' claim is not supported; the authors must either establish that the PSNR gap is not statistically significant or revise the claim to be metric-specific.","section":"IV.C, Table I"},{"comment":"On the Replica extrapolation split, Ours achieves average PSNR 37.915 dB, below M2-Mapping's 38.226 dB, and loses to M2-Mapping in five of the eight scenes. The abstract's claim that extrapolation results achieve state-of-the-art performance is therefore not supported by the reported numbers.","section":"IV.B, Table II"},{"comment":"The abstract states that interpolation and extrapolation sequences in both Replica and FAST-LIVO2 are evaluated, but Section IV.C reports only interpolation for FAST-LIVO2 and no FAST-LIVO2 extrapolation table is given. The missing experiment is directly relevant to the central claim and should either be added or the abstract revised.","section":"IV.C and Abstract"},{"comment":"Mesh-Learner requires a high-quality mesh a priori, whereas 3DGS, MonoGS, and M2-Mapping reconstruct geometry from RGB or RGB-D input; the reported comparison does not control for this additional input. The Limitations section acknowledges mesh incompleteness as a source of artifacts, but no experiment quantifies the effect of mesh quality on rendering accuracy, so the comparison conflates representation quality with the advantage of a supplied mesh.","section":"III.A and Limitations"},{"comment":"No quantitative ablation is reported for the three proposed components (hybrid interpolation, world-space EWA filter, adaptive SH density); Figures 5 and 6 are qualitative only. Without ablations, the contribution of each component to the final numbers cannot be assessed, and the novelty claims in Section I are not substantiated.","section":"IV.D"}],"minor_comments":[{"comment":"Algorithm 1 contains typos ('patiance', 'T_patience') and uses ¯Ai, ¯Vi, and ¯ρi without defining their initial values; please clarify the initialization and the exact termination condition.","section":"III-F, Algorithm 1"},{"comment":"The inverse-distance weighting exponent of 0.9 is introduced without rationale or sensitivity analysis; please add a brief justification or an ablation.","section":"III-D, Eq. (1)"},{"comment":"The 15-degree coplanarity threshold for the world-space EWA filter is stated without supporting evidence; please provide a sensitivity study or a citation.","section":"III-E"},{"comment":"The baselines are not described as being evaluated under the same protocol (e.g., how training and test frames are selected for each method on FAST-LIVO2); please report the exact split and any filtering of test poses for all methods.","section":"IV.A"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid systems contribution, but the abstract and introduction overstate the results. I would ask the authors to add the missing FAST-LIVO2 extrapolation experiments, provide quantitative ablations, and either control for the mesh input or tone down the state-of-the-art claim. The underlying method appears sound and the writing is generally clear, so I do not see the need for rejection; major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this paper does something real. It learns view-dependent SH textures per mesh entirely inside a rasterization pipeline, with hybrid interpolation across neighboring textures and a world-space EWA filter. That combination is new, and it gives a practical path from lidar-built meshes to photorealistic rendering in engines like Blender or Unity without converting NeRF or 3DGS. The system is well engineered, the ideas are clearly explained, and the qualitative results on Replica and FAST-LIVO2 look strong, especially on LPIPS and SSIM. The adaptive density heuristic is a sensible engineering answer to a real problem.\n\nBut the central claim is oversold. The abstract says state-of-the-art on both datasets for both interpolation and extrapolation, but the tables don't say that. On FAST-LIVO2, your average PSNR (25.624) is below 3DGS (25.825) and Textured-GS (25.856). On Replica extrapolation, your PSNR (37.915) is below M2-Mapping (38.226). You win on SSIM and LPIPS, which matters, but you can't call it SOTA without acknowledging the PSNR deficit. Also, the abstract implies FAST-LIVO2 extrapolation results, but Section IV.C only gives interpolation. That's a missing experiment, not a small omission.\n\nThe bigger limitation is the input assumption: you need a high-quality mesh, typically from lidar mapping. The paper states this openly in the Limitations section, and it's honest, but it changes the comparison. 3DGS and M2-Mapping build geometry from images or RGB-D; you start with ground-truth-quality geometry. That's an advantage you don't fully account for in the SOTA framing.\n\nWhat's not soft: the framework itself is coherent, the math is straightforward, and the rasterization compatibility story is compelling for robotics simulation where meshes already exist. The paper is honest about its main weakness. I'd like to see code and a few missing details (training time, GPU memory, and the actual FAST-LIVO2 extrapolation split), but these are fixable.\n\nBottom line: this deserves a serious referee and likely publication after revision. The SOTA claim needs to be qualified, the missing experiments added, and the mesh dependency made prominent. If the code ships, I'd cite it.\n\nSend it to peer review. The core contribution is worth the referees' time.","headline":"Mesh-Learner is a genuinely useful step toward rasterization-native view synthesis, but its SOTA claim is not backed by the paper's own tables.","tokens_in":13498,"tokens_out":1431,"would_cite":true,"duration_ms":16208,"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":"A mesh whose textures store spherical-harmonic coefficients can be trained end-to-end to render photorealistic novel views, performing on par with or better than state-of-the-art point-based and implicit representations while remaining…","keywords":["spherical harmonics","mesh textures","novel view synthesis","rasterization pipeline","view-dependent rendering","anti-aliasing","EWA filter","radiance fields"],"falsifier":"Take a scene with a foreground object whose mesh has a deliberate hole while the background mesh stays intact, train Mesh-Learner, and render a view behind the object: the paper's limitation statement predicts background surfaces will show foreground colors. A second check: place two adjacent triangles at a dihedral angle just above and just below 15 degrees and render a distant view; the EWA filter should suddenly stop borrowing texels across the boundary, producing a sharp aliasing seam at that threshold.","tokens_in":12467,"feed_emoji":"🎨","tokens_out":6249,"duration_ms":60294,"temperature":0.7,"pith_summary":"Mesh-Learner sets out to show that photorealistic novel-view synthesis does not require point splats or neural fields: an ordinary triangle mesh, textured with maps whose entries are spherical-harmonic coefficients, can be learned end-to-end and rendered through the standard rasterization pipeline. The authors claim this mesh-native representation achieves state-of-the-art quality on interpolation and extrapolation benchmarks from the Replica and FAST-LIVO2 datasets, comparing favorably with 3D Gaussian Splatting, M2-Mapping, and texture-based 3DGS variants. The payoff is practical: the trained output is a mesh plus texture maps, importable into existing rasterization-based engines and robotics and simulation stacks without conversion. The framework also trains large scenes under moderate GPU memory by moving only the SH textures inside the camera frustum to the GPU.","feed_headline":"Painting meshes with SH textures rivals Gaussian splatting","feed_subtitle":"Mesh-Learner renders novel views straight from a rasterizer, matching or beating 3DGS and M2-Mapping on Replica and FAST-LIVO2.","key_machinery":"The load-bearing object is the SH Texture — a texture map attached to each triangle mesh in which every texel stores spherical-harmonic coefficients rather than a fixed color, so shading becomes a per-view linear combination of coefficients. The machinery around it is a per-triangle resolution rule ('SH Density', the world-space distance between adjacent texel centers), a hybrid interpolation policy (bilinear inside the triangle; inverse-distance weighting over neighboring meshes' texels on edges and corners), a world-space EWA filter that averages all SH texels falling inside an ellipse around the sampling point, and an adaptive density scheduler that raises a mesh's SH Density while PSNR improves and stops once metrics stop improving.","core_discovery":"The paper's central claim is that view-dependent radiance can be stored, optimized, and rendered directly in SH-valued textures on scene meshes, and that this representation reaches state-of-the-art rendering quality without sacrificing rasterizer compatibility. Each triangle gets a dedicated SH texture, with resolution set by a world-space SH Density; rendering rasters geometry once into offscreen buffers (texture coordinates, view directions, world-space sample positions) and then a CUDA pass looks up and interpolates SH coefficients at every pixel. Interpolation is hybrid: bilinear inside the triangle interior, inverse-distance weighting that pulls in texels from neighboring meshes at edges and corners, and a world-space EWA anisotropic filter for distant pixels that averages all in-range SH texels within an ellipse, up to a 15-degree normal-disparity cutoff between coplanar neighbors. On Replica, the reported averages beat all compared methods on interpolation PSNR and LPIPS and on extrapolation SSIM and LPIPS; on FAST-LIVO2, the reported SSIM and LPIPS averages beat 3DGS and Textured-GS while PSNR is close.","pith_inferences":["If the quality gap holds at scale, lidar-based mapping systems that already output meshes could output a directly renderable representation, making the mesh itself the final asset rather than an intermediate that must be converted from Gaussians or neural fields.","A natural extension is to reformulate the adaptive SH-density rule as a constrained optimization that, for a fixed total texture budget, picks per-triangle densities to maximize a perceptual index rather than per-mesh PSNR.","The 15-degree normal cutoff in the world-space EWA filter is a binary heuristic; a continuous weight that decays with the angle between normals would likely remove visible seams at the cutoff.","Since the pipeline renders through deferred rasterization, it could be coupled with differentiable rasterizers that jointly optimize mesh vertices and SH textures, directly addressing the incomplete-mesh limitation the paper identifies."],"forward_implications":["Trained SH textures can be loaded directly into existing rasterization-based renderers (for example Blender) and used for rendering without conversion, because inference is just texture sampling and interpolation.","Large scenes can be trained with moderate GPU memory, since only the SH textures inside the current camera frustum are resident on the GPU; the rest stay in CPU RAM.","The same trained representation can feed downstream tasks that already consume rasterized meshes, such as robot simulation and reinforcement learning environments.","The adaptive density scheduler and world-space EWA filter shift texture memory toward detail-rich regions and suppress aliasing in distant views, so quality does not require a single worst-case texture resolution."],"supporting_citations":[{"why":"3D Gaussian Splatting is the main point-based baseline the paper compares against and must beat to support the state-of-the-art claim.","marker":"[6]"},{"why":"The EWA filter is the anisotropic filtering method the paper extends from local texture space into world space for anti-aliasing.","marker":"[15]"},{"why":"Plenoxels supplies the idea of storing spherical-harmonic coefficients in a grid and interpolating them, which the SH Texture adapts from 3D grids to 2D mesh textures.","marker":"[18]"},{"why":"Textured-GS is the texture-based 3DGS baseline that represents the closest competing approach using textures attached to Gaussians.","marker":"[19]"},{"why":"M2-Mapping provides the Replica evaluation trajectories and is a baseline method for neural surface reconstruction and rendering.","marker":"[23]"},{"why":"FAST-LIVO2 is the real-world dataset used for evaluation and is also cited as a source of high-precision meshes.","marker":"[14]"},{"why":"The Replica dataset is the synthetic indoor benchmark used for both interpolation and extrapolation rendering tests.","marker":"[27]"},{"why":"MonoGS is an RGBD-based Gaussian Splatting baseline included in the Replica comparisons.","marker":"[29]"},{"why":"ImMesh is cited as a lidar-based meshing method that can supply the high-precision input meshes the framework depends on.","marker":"[13]"}],"fun_headline_variants":["Rasterizer-friendly SH textures beat 3DGS on Replica","Mesh-Learner: SH textures make meshes rival 3D Gaussian","Spherical harmonic textures: mesh radiance without splatting","Mesh-Learner renders SH textures via plain rasterization","SH-Texel meshes: new SOTA, still Blender-compatible"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline takes a good scene mesh as given; if that mesh is incomplete, foreground colors are painted onto background surfaces and appear as artifacts in novel views (the paper states this limitation explicitly), and the world-space EWA filter further assumes adjacent triangles lie within 15 degrees of coplanar.","fun_headline_variants_meta":{"raw":{"variants":["Rasterizer-friendly SH textures beat 3DGS on Replica","Mesh-Learner: SH textures make meshes rival 3D Gaussian","Spherical harmonic textures: mesh radiance without splatting","Mesh-Learner renders SH textures via plain rasterization","SH-Texel meshes: new SOTA, still Blender-compatible"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000505,"raw_usage":{"total_tokens":2518,"prompt_tokens":1055,"completion_tokens":1463,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":671,"completion_tokens_details":{"reasoning_tokens":1369}},"tokens_in":671,"tokens_out":1463,"duration_ms":11741,"temperature":1.0,"reasoning_tokens":1369,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:38:47.950121+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a scene with a foreground object whose mesh has a deliberate hole while the background mesh stays intact, train Mesh-Learner, and render a view behind the object: the paper's limitation statement predicts background surfaces will show foreground colors. A second check: place two adjacent triangles at a dihedral angle just above and just below 15 degrees and render a distant view; the EWA filter should suddenly stop borrowing texels across the boundary, producing a sharp aliasing seam at that threshold.","supporting_citations":[{"cited_title":"Creating raster omnimax images from multiple perspective views using the elliptical weighted average filter,","cited_arxiv_id":null,"evidence_quote":"The EWA filter is the anisotropic filtering method the paper extends from local texture space into world space for anti-aliasing."},{"cited_title":"Plenoxels: Radiance fields without neural networks,","cited_arxiv_id":null,"evidence_quote":"Plenoxels supplies the idea of storing spherical-harmonic coefficients in a grid and interpolating them, which the SH Texture adapts from 3D grids to 2D mesh textures."},{"cited_title":"Fast- livo2: Fast, direct lidar–inertial–visual odometry,","cited_arxiv_id":null,"evidence_quote":"FAST-LIVO2 is the real-world dataset used for evaluation and is also cited as a source of high-precision meshes."},{"cited_title":"ImMesh: An Immediate LiDAR Localization and Meshing Framework","cited_arxiv_id":"2301.05206","evidence_quote":"ImMesh is cited as a lidar-based meshing method that can supply the high-precision input meshes the framework depends on."}],"review_version":1}