{"id":"2e433593-a033-4780-aebe-dd1794ffbb70","arxiv_id":"2505.16709","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A single-encoder dual-decoder learned codec jointly compresses point cloud geometry and color, reporting 75.0% geometry, 32.6% color, and 33.2% overall bitrate savings over G-PCC.","lead":"This paper presents SEDD-PCC, a learned point cloud compressor that uses one shared encoder and two decoders to handle both 3D geometry and color in a single pass. A reader might care because it reports large bitrate savings over the G-PCC standard while using a much smaller model than other learned codecs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 2.2 sends a G-PCC-octree-coded thumbnail geometry C_z as side information, but Eqs. (1)-(3) optimize only the feature rate R_zhat and no ablation isolates C_z; if C_z's bits are excluded from Table 1, the 75.0% D1 BD-BR claim is not established.","rationale":"The reader correctly identifies that a three-channel attribute-only latent carrying enough geometry is a necessary condition for the architecture to work as advertised. I agree that no analysis is given for that condition. My stress-test focuses one step earlier: Section 2.2 explicitly routes a thumbnail geometry C_z through G-PCC's octree codec as a losslessly coded side channel. This is not mentioned in the abstract or the reader's summary, and it changes what has to be true for the headline claim. If C_z's bits are part of the rate, then the reported BD-BR is not produced by the learned latent alone; if they are not part of the rate, the comparison with G-PCC is unfair. The training losses in Eqs. (1)-(3) refer only to R_zhat, so the paper is at least internally unclear about how C_z bits enter the RD calculation. The Table 3 ablations test teacher and transform module, but not the existence of C_z, so they cannot attribute the D1-PSNR gain to the shared encoder. This is a concrete accounting and attribution gap, not a style objection. It does not prove the method is wrong; a clean bitrate split and a recomputed Table 1 could resolve it. Therefore the reader's CONDITIONAL verdict stands, with an added explicit condition.","tokens_in":9174,"tokens_out":5601,"duration_ms":49457,"concrete_test":"On the six test sequences, report the per-sequence bit counts for C_z (G-PCC octree) and F_z, and recompute the BD-BR numbers in Table 1 using total bits = bits(C_z) + bits(F_z). If the re-computed average D1-PSNR BD-BR saving is materially below 75.0%, the headline claim fails; if it is stable, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 states that the thumbnail point cloud geometry C_z is losslessly encoded with the G-PCC octree codec while the feature F_z is quantized and entropy coded. This is an explicit, non-learned geometry side channel. The paper's central contribution (abstract, Section 3.2) is that one attribute-only encoder extracts shared geometric and attribute features and that the geometry decoder reconstructs geometry from the unified latent space. That claim is only load-bearing if geometry reconstruction is not simply inherited from the G-PCC-coded C_z. The training losses in Eqs. (1)-(3) all define R as R_zhat, i.e. the rate term is the quantized feature rate; no term includes the octree rate of C_z. It is therefore unclear whether the reported RD points in Table 1 and Fig. 4 include the C_z bitstream. If they do not, the BD-BR savings against G-PCC are computed against an incomplete total bitrate. If they do, the paper still does not report the bitrate split, so the reader cannot tell whether the learned decoders or the G-PCC side information drive the geometry metric. The ablation in Table 3 tests the transform module and teacher, but never removes or varies C_z, so the contribution of the side channel is unknown. This is a concrete, verifiable gap in the central argument, not merely a missing baseline.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SEDD-PCC, a unified learned point cloud codec with a single encoder and two decoders for geometry and attributes. Instead of representing the input as a four-channel occupancy-plus-color voxel grid, the encoder takes only the three RGB attribute channels, and a transform module plus a geometry decoder reconstruct the point positions from the shared latent. Training is done in three stages: attribute coding, geometry coding with knowledge distillation from PCGCv2, and joint fine-tuning. On 8iVFB and Owlii sequences, the method reports BD-BR savings of 75.0% (D1-PSNR), 32.6% (Y-PSNR), and 33.2% (1-PCQM) relative to G-PCC v23, with a 32.6 MB model and lower complexity than sequential approaches.","tokens_in":9405,"tokens_out":5278,"duration_ms":39437,"significance":"If the reported numbers are sound, SEDD-PCC offers a compelling demonstration that a single attribute-centric encoder can support both geometry and attribute decoding in a lightweight, end-to-end learned pipeline, avoiding the recoloring overhead of sequential codecs. The three-stage training and the use of a transform module with knowledge distillation are sensible engineering choices, and the ablation study (Table 3) provides initial evidence for their contribution. However, the manuscript currently omits critical experimental details—most importantly the bitrate accounting for the G-PCC-coded thumbnail geometry C_z—and several comparisons are incomplete, so the central claim of large BD-BR savings is not yet established. The paper also lacks absolute rate-distortion points and error bars, which are standard expectations in learned compression papers.","major_comments":[{"comment":"The encoding process losslessly encodes a thumbnail point cloud geometry C_z with the G-PCC octree codec, yet the rate R in all training losses is defined as R_zhat, the rate of the quantized feature F_z. The paper never states whether the bitrate of C_z is included in the reported BD-BR numbers of Table 1. If it is excluded, the claimed 75.0% bitrate reduction for D1-PSNR over G-PCC is not a fair end-to-end comparison, since G-PCC's bitrate includes full geometry. If it is included, the paper must report the bitrate split between C_z and F_z, otherwise the reader cannot determine whether the geometry-quality gains come from the learned geometry decoder or from the G-PCC side information. Because the central contribution is that a single attribute-only encoder extracts shared features for geometry reconstruction, this missing accounting is load-bearing.","section":"Section 2.2, Eqs. (1)-(3), Table 1"},{"comment":"Table 1 reports only BD-BR values and no absolute rate-distortion operating points. Without the underlying RD points, the BD-BR computations cannot be verified, and the reader cannot assess whether the average savings are dominated by a particular rate region. The authors should provide the per-sequence RD data (e.g., for D1-PSNR, Y-PSNR, 1-PCQM) in supplementary material.","section":"Section 3.1, Table 1"},{"comment":"In Table 1, the JPEG Pleno columns contain dashes for five of the six test sequences, with numbers only for Soldier, yet the final row reports an 'Average' of -62.9, -51.5, -42.3 for D1, Y, 1-PCQM. This average appears to be computed from the single Soldier row, not from an average over sequences. Reporting a single-sequence result as an 'Average' is misleading and should be corrected; the comparison to JPEG Pleno is otherwise incomplete. Similarly, the Unicorn method is mentioned in Section 3.2 but no results for it appear in Table 1 or Fig. 4, so the claim of 'competitive performance against other learning-based methods' is not fully supported.","section":"Table 1, JPEG Pleno column"},{"comment":"The paper claims that using only three attribute channels in the shared encoder is sufficient because 'attributes inherently encapsulate richer information than the geometry occupancy,' but it provides no evidence that the latent representation retains the geometric structure needed for accurate D1-PSNR reconstruction. Since the geometry decoder operates on the latent F_z, and the paper does not isolate the contribution of F_z versus the G-PCC-coded C_z, the role of the shared latent in geometry reconstruction is unclear. An ablation that removes or corrupts C_z (or compares with a pure G-PCC geometry side stream) would directly address this concern.","section":"Section 2.1, Section 2.2"}],"minor_comments":[{"comment":"The hyperparameter lists for λ_A, λ_G, λ_t in stage 3 are given as six values but it is unclear how many rate points are generated and whether they correspond to the six columns of Fig. 4; please clarify the mapping between λ values and rate points.","section":"Section 3.1"},{"comment":"The term ℒ_BCE2 is described as 'multi-scale BCE' but the multi-scale formulation is not specified; please define it consistently with the main BCE loss.","section":"Eq. (2)"},{"comment":"The encoding/decoding times are missing for several methods with only dashes shown; please either provide the values or state that they were not measured.","section":"Table 2"},{"comment":"The generation of the thumbnail geometry C_z is not described (e.g., voxel size, whether it is derived from the input coordinates or from the latent); please provide this detail.","section":"Section 2.2"},{"comment":"Reference [5] contains a typo 'J, -J. Chiu' and should be 'J.-C. Chiu'.","section":"References"},{"comment":"The axes and units of the rate-distortion plot are not legible in the text; please ensure the figure is high resolution with labeled axes.","section":"Fig. 4"},{"comment":"The text states that without the transform module 'the bitrate is reduced by 7.1%' but Table 3 reports -7.1 for D1, which indicates a bitrate saving; the wording is ambiguous, as it could be misinterpreted as an increase; please rephrase to 'bitrate is reduced (improved) by 7.1%'.","section":"Section 3.4.2 and Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the potential omission of the side-information bitrate; this is not a minor transparency matter but affects the validity of the headline result. The authors should be asked to provide the bitrate split and recompute all BD-BR numbers with the total bitrate. The claim of 'first unified model' may also need qualification given JPEG Pleno and other works. The paper's scope is appropriate for the journal but requires substantial revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is worth a look: a single attribute-oriented encoder feeding two specialized decoders, with a three-stage training protocol and knowledge distillation. That combination is not in the cited prior work, and the ablations suggest both the transform module and the teacher actually help. The model is also light at 32.6 MB and the paper is clearly written. Credit where due: this is a sensible engineering contribution, not a repackaging of an existing system.\n\nThe soft spot is the one the stress test flags, and it lands. Section 2.2 explicitly says the thumbnail geometry C_z is losslessly encoded with the G-PCC octree codec. But the losses in Eqs. (1)–(3) all define rate as R_zhat, the quantized feature rate, with no term for C_z's bits. Table 1 reports BD-BR against G-PCC, yet nowhere does the paper give the bitrate split or even confirm that C_z's bits are included in the total. If they are excluded, the 75.0% D1-PSNR savings are computed against an incomplete bitrate. If they are included, the reader still cannot tell how much of the geometry quality comes from the learned decoder versus the side channel. The ablation in Table 3 never varies or removes C_z, so the load-bearing contribution of that channel is unknown. This is not a minor omission; it directly affects the central claim of joint end-to-end compression.\n\nOther gaps are smaller but real: the Unicorn baseline named in the text does not appear in Table 1, several cells are dashes while averages are reported, there are no absolute rate-distortion points, no error bars or multiple seeds, and no code or checkpoints. None of these refute the idea, but together they make independent verification impossible from the preprint.\n\nMy take: the architecture is novel and the training strategy is thoughtful, but the side-channel accounting is a load-bearing flaw that must be fixed before the headline results can be trusted. A serious referee should see this paper, because the problem is fixable and the approach may influence the AI-PCC direction. But the verdict should be major revision: clarify the C_z bitrate, report absolute RD curves and bitrate splits, complete the baseline table, and release code/data or at least provide enough detail for exact reproduction.\n\nI would not cite this as a reference result in the next year until the bitrate accounting is resolved. It is worth bringing to a reading group, though, as a case study in how side information can quietly undermine an end-to-end claim.","headline":"The architecture is genuinely new, but the reported bitrate savings may not include the G-PCC-coded thumbnail geometry side channel, so the headline numbers are not yet established.","tokens_in":10035,"tokens_out":1553,"would_cite":false,"duration_ms":14518,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single encoder can map a point cloud's color and geometry into one shared latent space, and two decoders reconstruct both, beating G-PCC by 75.0% in geometry bitrate.","keywords":["learned point cloud compression","joint geometry and attribute coding","single encoder dual decoder","shared latent space","knowledge distillation","sparse convolution","BD-BR","PCQM"],"falsifier":"Train SEDD-PCC on point clouds whose RGB values are replaced by a constant while keeping the same coordinates, and compare D1-PSNR on the 8iVFB test set. If geometry quality is maintained, the geometric information is coming from the sparse coordinates rather than from the attribute-derived latent, undercutting the claim that the shared latent carries geometry; if it collapses, the attribute-centric encoder is genuinely load-bearing.","tokens_in":8877,"feed_emoji":"📉","tokens_out":7789,"duration_ms":64014,"temperature":0.7,"pith_summary":"The paper is trying to establish that point cloud geometry and attributes can be compressed jointly by one encoder and two decoders, rather than by separate codecs. It claims a three-channel, attribute-only input still leaves a shared latent that supports accurate geometry reconstruction, eliminating the recoloring step used by sequential pipelines. The headline numbers are BD-BR savings of 75.0% for D1-PSNR, 32.6% for Y-PSNR, and 33.2% for 1-PCQM versus G-PCC on MPEG test sequences, with a model of 32.6 MB. A sympathetic reader cares because this is a simpler, lighter architecture than sequential joint codecs, and it goes end-to-end.","feed_headline":"Single encoder cuts point-cloud bitrate by 75% versus G-PCC","feed_subtitle":"One shared latent space carries geometry and color; codec is end-to-end, avoids recoloring, and runs on a 32.6 MB model.","key_machinery":"The core mechanism is the shared latent space $z$ obtained by an attribute-oriented sparse-convolution encoder built with Inception-Residual layers. The encoder sees a sparse voxel tensor with coordinates $C$ and RGB features $F$, so the latent jointly represents appearance and structure. A transform module adapts $z$ into geometry-specific features for the geometry decoder, which uses a Top-k+1 classification mechanism to reconstruct point positions, while the attribute decoder follows the Sparse-PCAC process. The three-stage training protocol, attribute coding, geometry coding with teacher-model knowledge distillation, and joint fine-tuning, resolves the conflicting objectives, and the loss is a Lagrangian rate-distortion objective with lambda weighting.","core_discovery":"SEDD-PCC establishes that a single sparse-convolution encoder, fed only the three color channels of a voxelized point cloud, can produce a shared latent representation from which a transform module plus a geometry decoder reconstructs point positions while an attribute decoder reconstructs colors. The paper claims this is the first single-encoder, dual-decoder architecture for joint geometry and attribute compression, and that it avoids the recoloring stage of sequential pipelines. Compared with G-PCC, it reports average BD-BR reductions of 75.0% for D1-PSNR, 32.6% for Y-PSNR, and 33.2% for 1-PCQM, while staying competitive with learned sequential codecs.","pith_inferences":["The paper leaves implicit that the same single-encoder design could serve point clouds without color by setting attributes to a constant; a natural test is whether geometry-only BD-BR still beats G-PCC under that input.","Because bit allocation is set by the loss weights instead of a rate-allocation algorithm, the reported average BD-BR may shift along the geometry/attribute trade-off frontier; comparing SEDD-PCC against an optimal allocation frontier would show how much headroom remains.","The claimed independence from recoloring implies that geometry errors no longer leak into attribute quality; a direct test is to compare Y-PSNR of SEDD-PCC against a sequential codec at equal geometry distortion."],"forward_implications":["SEDD-PCC removes the separate recoloring stage, so attribute quality no longer inherits geometry reconstruction errors through a recoloring step.","A single attribute-centric encoder suffices for both tasks, so joint codecs can be built with fewer parameters than two separate codecs; the model is 32.6 MB.","Knowledge distillation from a geometry teacher and a dedicated transform module each provide roughly 7% bitrate savings, making them effective components for shared-latent joint compression.","The design offers an end-to-end learned alternative to G-PCC with an average 75.0% D1-PSNR bitrate reduction and 33.2% 1-PCQM reduction on the evaluated test sequences.","Because geometry and attributes share one latent, rate-distortion optimization can be done jointly in a single loss, avoiding separate bit allocation between two codecs, though the paper notes this allocation may not yet be optimal."],"supporting_citations":[{"why":"Supplies the geometry decoder structure and the teacher model whose feature maps are used for knowledge distillation in Stage 2.","marker":"[3]"},{"why":"Supplies the sparse-convolution attribute codec that the shared encoder and attribute decoder are built on.","marker":"[9]"},{"why":"Provides the four-channel single encoder-decoder joint model that SEDD-PCC contrasts with by using only three channels.","marker":"[14]"},{"why":"Sequential joint codec baseline that uses G-PCC plus enhancement; SEDD-PCC compares BD-BR against it.","marker":"[16]"},{"why":"Learned sequential geometry-plus-attribute baseline using sparse convolution and self-attention, used as a comparison.","marker":"[17]"},{"why":"JPEG Pleno learning-based joint PCC standard baseline whose BD-BR is compared on Soldier.","marker":"[19]"},{"why":"Conditional joint codec representing the sequential approach SEDD-PCC aims to supersede.","marker":"[21]"},{"why":"Defines the Inception-Residual Network layers used inside the shared encoder and decoders.","marker":"[23]"},{"why":"Supplies the Top-k+1 classification mechanism the geometry decoder uses when upscaling points.","marker":"[24]"}],"fun_headline_variants":["One encoder, dual decoders: 75% bitrate cut vs G-PCC","Shared latent space shrinks point-cloud bitrate by 75%","Joint geometry-color coding beats G-PCC by 75% BD-BR","SEDD-PCC: one encoder, two decoders, 75% savings","Single encoder + dual decoders = 75% BD-BR gain vs G-PCC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the three attribute channels, after passing through one shared encoder, leave a latent representation rich enough for the transform module and geometry decoder to reconstruct point positions accurately; if attribute-oriented features wash out geometric detail, the reported D1-PSNR gains collapse.","fun_headline_variants_meta":{"raw":{"variants":["One encoder, dual decoders: 75% bitrate cut vs G-PCC","Shared latent space shrinks point-cloud bitrate by 75%","Joint geometry-color coding beats G-PCC by 75% BD-BR","SEDD-PCC: one encoder, two decoders, 75% savings","Single encoder + dual decoders = 75% BD-BR gain vs G-PCC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000578,"raw_usage":{"total_tokens":2687,"prompt_tokens":872,"completion_tokens":1815,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":1709}},"tokens_in":488,"tokens_out":1815,"duration_ms":12202,"temperature":1.0,"reasoning_tokens":1709,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:56:12.502477+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SEDD-PCC on point clouds whose RGB values are replaced by a constant while keeping the same coordinates, and compare D1-PSNR on the 8iVFB test set. If geometry quality is maintained, the geometric information is coming from the sparse coordinates rather than from the attribute-derived latent, undercutting the claim that the shared latent carries geometry; if it collapses, the attribute-centric encoder is genuinely load-bearing.","supporting_citations":[{"cited_title":"Multiscale point cloud geometry compression,","cited_arxiv_id":null,"evidence_quote":"Supplies the geometry decoder structure and the teacher model whose feature maps are used for knowledge distillation in Stage 2."},{"cited_title":"Sparse tensor -based point cloud attribute compression ,","cited_arxiv_id":null,"evidence_quote":"Supplies the sparse-convolution attribute codec that the shared encoder and attribute decoder are built on."},{"cited_title":"Towards neural network approaches for point cloud compression,","cited_arxiv_id":null,"evidence_quote":"Provides the four-channel single encoder-decoder joint model that SEDD-PCC contrasts with by using only three channels."},{"cited_title":"YOGA: Yet another geometry-based point clou d compressor,","cited_arxiv_id":null,"evidence_quote":"Sequential joint codec baseline that uses G-PCC plus enhancement; SEDD-PCC compares BD-BR against it."},{"cited_title":"DeepPCC: Learned Lossy Point Cloud Compression,","cited_arxiv_id":null,"evidence_quote":"Learned sequential geometry-plus-attribute baseline using sparse convolution and self-attention, used as a comparison."},{"cited_title":"The JPEG Pleno Learning-based Point Cloud Coding Standard: Serving Man and Machine,","cited_arxiv_id":null,"evidence_quote":"JPEG Pleno learning-based joint PCC standard baseline whose BD-BR is compared on Soldier."},{"cited_title":"A Versatile Point Cloud Compressor Using Universal Multiscale Conditional Coding - Part II: Attribute,","cited_arxiv_id":null,"evidence_quote":"Conditional joint codec representing the sequential approach SEDD-PCC aims to supersede."},{"cited_title":"Inception-v4, inception -resnet and the impact of residual connections on learning,","cited_arxiv_id":null,"evidence_quote":"Defines the Inception-Residual Network layers used inside the shared encoder and decoders."},{"cited_title":"SparsePCGCv1 Update: Improvements on dense/sparse/LiDAR point clouds,","cited_arxiv_id":null,"evidence_quote":"Supplies the Top-k+1 classification mechanism the geometry decoder uses when upscaling points."}],"review_version":1}