{"id":"4fe8b0a3-8cbd-40bf-a1e0-dfbd2a81467f","arxiv_id":"2508.07083","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"TeSO represents a point cloud scene as textured surfels on an octree and compresses it with a learned entropy model, reporting lower LPIPS at equal bit-rates than G-PCC, P2ENet, and B2P baselines.","lead":"This paper introduces TeSO, a 3D scene representation that stores small textured patches on flat surface pieces arranged in a tree, so smooth geometry can stay coarse while texture stays sharp. The authors report better rendering quality at lower bit rates than point-cloud and 3D Gaussian baselines, which matters for streaming 3D video to VR headsets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"R-D advantage may be an artifact of the Poisson-mesh ground truth and the sole use of LPIPS; an independent GT and metric are needed.","rationale":"The central claim of the paper is the rate-distortion superiority of TeSO over point-cloud and 3D-Gaussian baselines. The reader's weakest assumption singles out the evaluation premise: ground truth is a Poisson surface reconstruction and the metric is LPIPS. I agree this is the most load-bearing concern. TeSO's design is essentially a compact surface approximation with texture patches; the GT is also a surface approximation of the same source points, so TeSO is structurally aligned with the GT in a way that the baselines are not. This is not a claim about author intent or integrity; it is a structural risk in the experimental design. The paper itself acknowledges the LPIPS assumption (Sec. 7), but does not address the GT bias. The proposed check—varying the GT and the metric—directly tests whether the reported advantage is a property of the representation or of the specific evaluation scaffold. Since the reader's verdict is already CONDITIONAL and this concern is a sharpened articulation of that condition, no change to the verdict is needed; the condition should be stated as a requirement for acceptance.","tokens_in":18732,"tokens_out":5811,"duration_ms":62472,"concrete_test":"Recompute the R-D curves of Fig. 15 using an alternative ground truth that is not a Poisson reconstruction of the same point cloud—e.g., a mesh from ball-pivoting or screened Poisson with different depth/scale—and additionally report a second perceptual metric (DISTS or LPIPS with a different backbone) and PSNR/SSIM after alignment, for TeSO, G-PCC+P2ENet, and B2P at matched bpp. If TeSO remains on the Pareto frontier under both GTs and multiple metrics, the concern is resolved; if its margin shrinks or the ordering changes, the headline claim must be conditioned on the evaluation setup.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim is operationalized by Fig. 15, where rendering quality is measured by LPIPS against views rasterized from a Poisson surface reconstruction of the source point cloud (Sec. 6.1.1). TeSO is itself a smooth, planar-surface approximation with texture patches derived from the same point cloud; its rendered images are therefore likely to conform more closely to a Poisson-smoothed GT than are point splats or 3D Gaussians, which are not surface-reconstruction-based. This is not a small implementation detail: the GT embodies the same 'surface from points' assumption that TeSO is explicitly designed to exploit, creating a potential bias in the evaluation. The authors acknowledge the LPIPS-assumption in Sec. 7, but the GT bias is an additional, unaddressed confound. Because no code or data is released, the reported R-D curves cannot be independently verified. If the ranking reverses under a different GT or metric, the central claim reduces to 'TeSO fits a Poisson-rasterized GT well', not 'TeSO renders point-cloud scenes better at lower bitrates.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Textured Surfel Octree (TeSO), a 3D representation constructed from point clouds: an octree whose leaf nodes carry cube-bounded planar surfels, each with an attached texture patch. It describes a CUDA-accelerated construction algorithm, a rasterization-based renderer with soft blending, and a compression scheme that losslessly codes the octree geometry with a learned entropy model while coding texture either by packing patches into AV1 images or by rasterizing the TeSO into a colored point cloud and coding colors with G-PCC. The central claim is that TeSO plus this compression achieves lower LPIPS at equal or lower bit-rates than G-PCC+OpenGL, G-PCC+P2ENet, and B2P on the 8iVFB dataset, at both 1024x1024 and 1920x1920.","tokens_in":18962,"tokens_out":6734,"duration_ms":68996,"significance":"If the rate-distortion claim holds, TeSO is a practically interesting representation: it combines explicit surface geometry with high-frequency texture, is directly renderable without decoder-side surface reconstruction, and is compressible with a mixture of learned and standard codecs. The construction-speed and decoding-latency numbers are also useful engineering results, and the ablation in Table 1 gives some insight into the entropy model. However, the current evidence is not yet sufficient to establish the headline superiority: the evaluation rests on a single perceptual metric against a Poisson-derived ground truth, on only four sequences without variance reporting, and on baselines that include two of the authors' own prior systems.","major_comments":[{"comment":"The R-D claim is measured only as LPIPS against views rasterized from a Poisson surface reconstruction of the source point cloud. TeSO is itself a smooth-surfel representation built from that same point cloud, so it may be systematically favored by this ground truth over point-splat and Gaussian baselines. The acknowledgment in Sec. 7 about LPIPS does not address this ground-truth confound. I request additional evidence: (i) a reference that is not derived from surface reconstruction (e.g., original multi-view captures if available, or a high-quality EWA/surfel rendering of the original point cloud), (ii) at least one additional distortion metric (PSNR/SSIM or a small subjective study), and (iii) per-scene R-D curves. Without this, the reported advantage could be an artifact of the evaluation protocol rather than a property of the representation.","section":"Sec. 6.1.1, Fig. 15"},{"comment":"The headline comparison is averaged over only four 8iVFB sequences with no error bars or per-sequence breakdown. Two of the three baselines (P2ENet and B2P) are the authors' own prior methods, which increases the need for transparent reporting. Please provide per-sequence rates and distortions, standard deviations or confidence intervals, and the exact operating points that form the convex hulls. The introduction promises open-source code, but no code or data link is provided; releasing the implementation and bitstreams is important for independent verification of the R-D curves.","section":"Sec. 6.4 / Fig. 15"},{"comment":"The description says that after the TeSO geometry is coded, 'we only need to transmit the bits for the color attributes in the point cloud bit-stream.' Since G-PCC TMC13 normally encodes geometry and attributes in a joint bitstream, it is not clear how attribute-only decoding with pre-existing geometry is performed. If a modified or non-default configuration is used, the exact software configuration and bitstream-size accounting must be described; otherwise the reported bpp for the G-PCC-texture variant may omit geometry-related overhead. Please clarify this step and, if possible, provide a bitrate breakdown by geometry and texture.","section":"Sec. 5.3, 'Coding as Colored Point Cloud'"},{"comment":"The P2ENet baseline is reported as trained for 1024x1024 rendering, yet it is also evaluated at 1920x1920. If the learned renderer is resolution-dependent, this may disadvantage it at the higher resolution and inflate the apparent gap. Please state whether P2ENet was retrained or adapted for 1920x1920, or discuss its resolution generalization. The same level of detail should be provided for B2P's configuration and training data.","section":"Sec. 6.1.2 / Fig. 15"}],"minor_comments":[{"comment":"The decision function f is described as comparing D1-PSNR between grid points on the surfel and the point set, but the number and placement of grid points are not specified. Please make this precise for reproducibility.","section":"Sec. 3.3"},{"comment":"The notation Sgn is undefined; use sign(·) or define it at first use.","section":"Eq. (4)"},{"comment":"Typo: 'geoemetry' should be 'geometry'. Also, the table reports decoding times but not the standard deviation or the exact machine configuration beyond CPU/GPU model.","section":"Table 2"},{"comment":"References [49] and [50] appear to be the same paper with identical titles; one should be removed or corrected. Reference [1] is missing a closing period.","section":"References"},{"comment":"The caption does not state how many R-D points are averaged, what the curve construction procedure is, or whether shaded regions (if any) denote variance. Please clarify.","section":"Fig. 15"}],"recommendation":"major_revision","confidential_remarks":"The core representation is original and likely of interest, but the evaluation protocol is the main weakness. I do not see a fundamental flaw in the construction or compression pipeline; the issues are about evidence and fairness of comparison. I would encourage the editor to require code/data release as part of the revision, and to ask the authors to include an independent baseline and ground-truth source. The reliance on the authors' own baselines is not disqualifying by itself, but it raises the bar for transparent evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: TeSO is a legitimate new 3D representation — cube-bounded surfels on an octree, each carrying a texture patch — and the paper shows the combination works for streaming: direct rasterization, no decoder-side surface reconstruction, and texture detail that doesn't collapse when geometry is simplified. The R-D curves in Fig. 15 look good, and construction is fast (0.3s for 1M points). This is more than a repackaging of known pieces; the specific design of textured, cube-bounded surfels on an octree with a learned geometry entropy model and separately coded texture patches is genuinely new, to my knowledge.\n\nWhat the paper does well: it identifies a real limitation — point clouds and 3D Gaussians struggle to carry high-frequency texture at low primitive counts — and attacks it directly. The sequential conditioning in the geometry codec (offsets → normals → radius) gives measurable compression gains (Table 1), and the two texture coding options (G-PCC-based vs AV1 packetization) are a sensible speed/quality trade-off. The authors are also honest in Sec. 7 about relying on LPIPS and about the representation struggling with hair/transparency.\n\nWhere the soft spots are: the evaluation's ground truth is a rasterization of a Poisson surface reconstruction (Sec. 6.1.1). That's a confound. TeSO is a smooth, planar-surface representation; it is expected to align more closely with a Poisson-smoothed GT than point splats or unconstrained Gaussians. The paper acknowledges the LPIPS assumption but not this GT bias. The reported R-D advantage could shrink or reverse under a different GT (e.g., real multiview captures or raw point splats as reference). Also: four sequences, no variance bars; two of three baselines (P2ENet and B2P) are the authors' own prior work; P2ENet is evaluated at 1920×1920 though it was trained for 1024×1024; and the texture patches — the core innovation — are never ablated (what happens if you render with per-surfel average color?). No code or data is released, so the numbers can't be independently checked.\n\nNone of these are fatal, and the central claim — for this tested regime, TeSO beats the baselines — likely holds up. But the magnitude of the advantage in Fig. 15 is less certain than it looks.\n\nThis paper deserves a serious referee. The representation is worth engaging with, and the evaluation weaknesses are addressable in a revision. I'd ask the authors for a fairer GT, more scenes with error bars, an ablation of the texture patch, and released code before trusting the headline R-D numbers.","headline":"TeSO is a genuine new representation with measured R-D gains, but the Poisson-mesh ground truth and sole LPIPS metric keep me from fully trusting the headline numbers.","tokens_in":19487,"tokens_out":2493,"would_cite":true,"duration_ms":23131,"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 new 3D representation, the textured surfel octree, is claimed to deliver higher rendering quality at lower bit-rates than existing point-cloud and 3D-Gaussian pipelines.","keywords":["3D representation","surfel octree","point cloud compression","texture mapping","rate-distortion","3D Gaussian splatting","learned entropy model","real-time rendering"],"falsifier":"Render the same bitstreams for the 8iVFB scenes at 1920x1920 and run a human pairwise-preference study between TeSO and G-PCC+P2ENet at matched bpp; if viewers do not prefer TeSO, or if a different objective metric such as DISTS reverses the ordering, the claim of higher rendering quality at lower bit-rates would be refuted. Alternatively, re-run the R-D comparison with ground truth taken from a high-quality multi-view capture rather than Poisson reconstruction; if TeSO's advantage shrinks or reverses, the result depends on the ground-truth choice.","tokens_in":18570,"feed_emoji":"📦","tokens_out":7210,"duration_ms":68435,"temperature":0.7,"pith_summary":"The paper proposes TeSO, a 3D representation that stores a scene's geometry as flat, cube-bounded surfels arranged on an octree, with a separate small texture patch attached to each surfel. The central claim is that by using large surfels for smooth regions and preserving fine detail in texture maps, TeSO can render high-quality views at lower bit-rates than point-cloud or 3D-Gaussian-based baselines, without any decoder-side surface reconstruction. The paper also introduces a compression scheme that codes the octree geometry losslessly with a learned entropy model and codes the texture patches using G-PCC or AV1. On the four-scene 8iVFB test set, TeSO is reported to achieve lower LPIPS than G-PCC+OpenGL, G-PCC+P2ENet, and B2P at matched bit-rates, at both 1024x1024 and 1920x1920 resolutions.","feed_headline":"Textured surfel octree beats point clouds for 3D streaming","feed_subtitle":"Geometry simplifies in flat areas while texture stays dense, improving rate-distortion on 8iVFB.","key_machinery":"The central object is the textured surfel octree: each cube-bounded surfel is a plane with a center offset, normal, and radius, clipped to its octree cube; a soft semi-transparent border extends beyond the cube and blends with neighbors to hide cracks. Each surfel carries a small M×M texture patch in its tangent frame, computed from the original point colors. The compression chain quantizes the geometry attributes, losslessly codes octree occupancy, leaf flags, and attributes with a learned sparse-convolutional entropy model conditioned on previously coded levels and attributes, then codes texture patches either as a packed image with AV1 or as point colors with G-PCC. This machinery keeps t","core_discovery":"The paper's core claim is that decoupling geometry from texture in a hierarchical surfel representation improves the rate-distortion frontier for streaming 3D scenes. A scene is encoded as a set of cube-bounded planes — each with an offset, normal, and radius — stored on the leaves and selected internal nodes of an octree, with a small texture patch per surfel. Construction starts from a point cloud: points are grouped by cube at coarse levels, and a cube is kept as one surfel if a sampled grid on the surfel reconstructs the points with D1-PSNR above a threshold; otherwise the cube splits. On the 8iVFB benchmark, the compressed representation is reported to yield lower LPIPS at equal bits-pe","pith_inferences":["Because TeSO rendering is differentiable, the hand-tuned splitting threshold and quantizer steps could be replaced by a learned rate-distortion optimization, a direction the paper names but does not pursue.","The ground-truth evaluation uses Poisson-reconstructed meshes, which are smooth; if real multi-view captures are used instead, the comparison with point-based renderers might shift, so the reported margin is likely tied to the smoothness of the reference.","The soft-blending width is fixed to the finest cube width; scenes mixing very different object scales might need an adaptive blending range, which the current method does not address.","Since texture coding via G-PCC outperforms AV1 image packing, a learned texture codec that respects the 3D adjacency of patches would likely improve the rate-distortion curve further."],"forward_implications":["Directly renderable streaming: a receiver decodes TeSO into surfel primitives and rasters them immediately, eliminating the surface-reconstruction or learned-rendering step that point-cloud pipelines require.","Bit-rate can be spent on texture instead of geometry in smooth areas, so scenes with flat regions and high-frequency color keep their appearance even when geometry is aggressively simplified.","The same octree structure supports partial and viewport-adaptive streaming: texture patches are local, so a client can fetch only the surfels and patches that are visible.","The geometry bitstream is lossless once attributes are quantized, so coding distortion is controlled entirely by geometry quantization, splitting threshold, and texture quantization."],"supporting_citations":[{"why":"Provides the G-PCC geometry codec used for the base octree, the texture-color point cloud coding, and the main comparison baseline.","marker":"[2]"},{"why":"Supplies the four 8iVFB test scenes used for the rate-distortion evaluation.","marker":"[10]"},{"why":"Supplies the UVG-VPC training sequences for the learned entropy model.","marker":"[13]"},{"why":"The P2ENet learned splatting baseline that converts decoded points to 3D Gaussians; the advanced rendering comparison target.","marker":"[19]"},{"why":"The B2P end-to-end compression-to-Gaussian-rendering baseline that must be beaten on rate-distortion.","marker":"[20]"},{"why":"Screened Poisson surface reconstruction used to generate the ground-truth mesh and views for distortion evaluation.","marker":"[24]"},{"why":"3D Gaussian Splatting whose soft-blending mechanism the rendering adopts, and whose single-color-per-primitive limitation motivates the texture patches.","marker":"[25]"},{"why":"The G-PCC white paper defining the standard the geometry coding builds on and compares against.","marker":"[39]"},{"why":"The original surfel primitive concept that the cube-bounded surfel extends.","marker":"[41]"},{"why":"LPIPS metric used to measure rendering distortion, chosen for perceptual alignment and insensitivity to small boundary shifts.","marker":"[55]"}],"fun_headline_variants":["Surfel octree beats point clouds and 3D Gaussians for streaming","Decoupling geometry and texture in surfel octrees improves 3D streaming","TeSO: surfel octree with texture patches cuts 3D streaming bitrate","Surfel octree with texture patches beats point cloud baselines"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The headline improvement is measured only with LPIPS against ground-truth views that are rasterizations of a Poisson surface reconstruction of the source point cloud; if LPIPS does not match human perception for these artifacts, or if Poisson reconstruction systematically favors smooth parametric surfaces, the reported rate-distortion advantage may not carry over to real captures or to human viewers.","fun_headline_variants_meta":{"raw":{"variants":["Surfel octree beats point clouds and 3D Gaussians for streaming","Decoupling geometry and texture in surfel octrees improves 3D streaming","TeSO: surfel octree with texture patches cuts 3D streaming bitrate","Surfel octree with texture patches beats point cloud baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000655,"raw_usage":{"total_tokens":2855,"prompt_tokens":782,"completion_tokens":2073,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":2003}},"tokens_in":526,"tokens_out":2073,"duration_ms":15231,"temperature":1.0,"reasoning_tokens":2003,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:20:24.855882+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the same bitstreams for the 8iVFB scenes at 1920x1920 and run a human pairwise-preference study between TeSO and G-PCC+P2ENet at matched bpp; if viewers do not prefer TeSO, or if a different objective metric such as DISTS reverses the ordering, the claim of higher rendering quality at lower bit-rates would be refuted. Alternatively, re-run the R-D comparison with ground truth taken from a high-quality multi-view capture rather than Poisson reconstruction; if TeSO's advantage shrinks or reverses, the result depends on the ground-truth choice.","supporting_citations":[{"cited_title":"MPEG G-PCC TMC13","cited_arxiv_id":null,"evidence_quote":"Provides the G-PCC geometry codec used for the base octree, the texture-color point cloud coding, and the main comparison baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the four 8iVFB test scenes used for the rate-distortion evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the UVG-VPC training sequences for the learned entropy model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The P2ENet learned splatting baseline that converts decoded points to 3D Gaussians; the advanced rendering comparison target."},{"cited_title":"Bits-to-Photon: End-to-End Learned Scalable Point Cloud Compression for Direct Rendering","cited_arxiv_id":"2406.05915","evidence_quote":"The B2P end-to-end compression-to-Gaussian-rendering baseline that must be beaten on rate-distortion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Screened Poisson surface reconstruction used to generate the ground-truth mesh and views for distortion evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The G-PCC white paper defining the standard the geometry coding builds on and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The original surfel primitive concept that the cube-bounded surfel extends."}],"review_version":1}