{"id":"4cb52115-5c35-4423-aba1-53937592155a","arxiv_id":"2608.11845","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"ResPCC trains an end-to-end learned point cloud codec to survive packet loss by conditioning on loss rate, interleaving latent features, and restoring damaged entries with graph and dictionary modules.","lead":"ResPCC is a neural point cloud compression codec that keeps 3D reconstructions usable even when 5 to 30 percent of transmitted packets are lost. It conditions encoding on the expected loss rate, scrambles the latent data so losses become scattered instead of channel-wide, and repairs the damaged latent features before decoding.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No error bars and unreported Gilbert-Elliott burst parameters leave the core claim of practical burst-loss resilience under-tested: training uses element-wise i.i.d. erasures while deployment uses packet-level bursty loss.","rationale":"The paper's main contribution is not simply a new architecture but the claim of intrinsic resilience to packet loss. If the training erasure model is too far from packet-level burst loss, the method may be overfit to the simulation rather than robust to real networks. The reader's weakest_assumption identifies the same gap: the load-bearing premise is that packet loss can be modeled as binary erasure controlled by a single perceived loss rate. The paper does include G-E evaluation, which is a step in the right direction, but without G-E parameters and burst-length sensitivity, the evidence is incomplete. Other issues, such as the duplicated Table 1 row and the strengthened backbone confound, are important but secondary; the loss-model mismatch is the one that would invalidate the central generalization claim if it lands. The reader's conditional verdict remains appropriate because the concern is addressable with an additional experiment rather than a fundamental flaw.","tokens_in":21070,"tokens_out":14990,"duration_ms":166519,"concrete_test":"Run the released checkpoint on ShapeNet at a fixed average PLR of 20% under three Gilbert-Elliott parameterizations with mean burst lengths of 1, 4, and 10 packets (same average loss rate), and report D2-PSNR. If D2 drops by more than about 1 dB as burst length increases, the resilience is tied to the specific burstiness used in the paper; also retrain with packet-level G-E masks in training and compare to the i.i.d.-masked model at the same test settings, which directly isolates the training/test mismatch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a single model conditioned on perceived loss rate stays robust under realistic packet loss. The training-time corruption in Sec. 5.5 samples the mask m_i over the interleaved H×W×C tensor \"according to p\", which for a scalar p is an element-wise independent Bernoulli erasure. The test-time corruption in Sec. 6.1 is instead generated by a Gilbert-Elliott model at the packet level, and after inverse SCI a lost packet maps to a structured pattern rather than to independent elements. The paper never reports the G-E transition probabilities or mean burst length, so the reader cannot tell whether the tested burstiness resembles real networks or whether the method is sensitive to it. If the default burst length is short, training and test may nearly coincide; if it is long, MGLR may have been evaluated on a regime the i.i.d.-trained model was never optimized for. Since the abstract promises \"diverse packet loss conditions\" and \"practical networks\", this missing sensitivity analysis is the load-bearing gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ResPCC, an end-to-end learned point cloud codec built on a DPCC-style backbone, with four additions for packet-loss resilience: CALM conditions the encoder on a perceived loss rate; SCI permutes latent entries so packet-level/channel-wise erasures become scattered element-wise misses; MGLR restores missing latent entries using a K-NN graph with a loss mask and reliability weighting; and DBR refines restored features against a learned codebook. The encoder and decoder are trained jointly with simulated erasures at sampled loss rates, and evaluated on ShapeNet and SemanticKITTI at 5% to 30% packet loss against G-PCC, DPCC, PCGCv2, and OctAttention. The reported results show consistent R-D gains, and the appendix provides architecture, training, and additional qualitative details.","tokens_in":21276,"tokens_out":7217,"duration_ms":75552,"significance":"If the empirical claims hold, ResPCC addresses a practically important gap: most learned PCC methods assume lossless transport, and the proposed mechanism (one model conditioned on loss rate plus restoration) is more attractive than per-rate retraining or external concealment. The paper also provides useful analysis of why interleaving helps (turning channel-wise extinction into scattered misses) and ships a code link and detailed appendices, which support reproducibility. The significance is currently contingent, however, on correcting internal inconsistencies in the ablation and on demonstrating that the i.i.d.-trained model is actually robust under the bursty packet-loss model used at test time.","major_comments":[{"comment":"Table 1 lists the identical configuration (DBR=no, CALM=no, MGLR=yes, Linear=no, SCI=yes) twice: Category I, third row, reports Bpp 2.653 and D2-PSNR 69.152 at PLR=5%, while Category II, first row, reports Bpp 2.325 and D2-PSNR 68.850 for the same configuration at the same PLR. Since the text states every configuration is re-trained end-to-end, these two sets of numbers cannot both describe the same model. This internal inconsistency affects the ablation conclusions for MGLR and SCI, which are key to the paper's central claim.","section":"Table 1"},{"comment":"Training simulates loss by sampling an element-wise binary mask m_i 'according to p' on the interleaved H×W×C tensor, i.e., independent Bernoulli erasures. Testing instead uses a Gilbert-Elliott model at packet level. The paper never reports the G-E transition probabilities or mean burst length, nor does it specify how packet-level losses map to entries of the interleaved tensor after SCI. Without these details and a sensitivity analysis over burst lengths, the central claim of resilience under realistic bursty loss is not established; the model may have been optimized for a corruption distribution very different from the one tested.","section":"Sec. 5.5 and Sec. 6.1"},{"comment":"All reported PSNR and Bpp values in Fig. 4 and the tables appear to be single point estimates. Since packet losses are random, D1/D2 PSNR at a given PLR is a random quantity; no error bars, confidence intervals, or multiple-loss-realization statistics are reported. The claims of 'consistently delivers superior stability and R-D performance' therefore lack statistical support. At minimum, the authors should report mean±std over multiple loss patterns and over multiple training runs, or provide a significance test.","section":"Sec. 6.2 and Fig. 4"},{"comment":"Appendix E states that 'the main paper included relatively successful cases' for OctAttention and that the supplementary figures highlight its high variance. This is an explicit admission of potentially selective qualitative reporting in the main text. The authors should clarify whether the quantitative curves in Fig. 4 are averages over all patches and loss realizations and, if qualitative examples are chosen, present them without selection or state the selection criterion.","section":"Appendix E"}],"minor_comments":[{"comment":"ResPCC's Bpp accounting for the FEC-protected anchors and hyperprior is not specified; baselines include anchor FEC overhead in Bpp, so the comparison should state explicitly whether the same overhead is included for ResPCC.","section":"Sec. 5.1 and Sec. 6.1"},{"comment":"The G-E parameters used to generate test loss sequences should be reported, and the code release should include the exact scripts that produce both the Bernoulli training masks and the Gilbert-Elliott test masks.","section":"Sec. 6.1"},{"comment":"The caption says all variants are evaluated without the DBR module; this should also be stated in the main text to avoid confusion with Table 1.","section":"Fig. 6"},{"comment":"The symbol m_i is used both for the packet-loss mask and for entries of that mask; a single notational convention (e.g., bold M for the mask tensor and scalar entries m_c) would improve readability.","section":"Sec. 5.3"},{"comment":"The claim of being 'the first end-to-end neural point cloud codec with intrinsic resilience' should be carefully qualified in light of prior intrinsic-resilience work in other modalities and existing point-cloud error-concealment methods; if no intra-frame point-cloud baseline exists, this should be stated explicitly rather than implied.","section":"Abstract and Sec. 1"}],"recommendation":"major_revision","confidential_remarks":"The duplicate row in Table 1 is the most serious internal error; before resubmission the authors need to rerun the ablations or correct the table. The missing Gilbert-Elliott parameters and lack of error bars are also essential to the central claim. If these issues are addressed, the paper could be a solid fit for ACM MM. I would also ask the editors to have the authors double-check the 'first' claim against GRACE and related learned codecs."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper up front. First, it is a genuine engineering contribution: a single neural point cloud codec that conditions on the packet loss rate, interleaves the latent to spread channel-wise loss into scattered missing entries, and restores the corrupted latent with a mask-aware graph network. The combination is new to learned PCC, the appendix is unusually thorough, and the code is public. Second, the empirical case has a concrete hole: the ablation table lists the same configuration twice with different numbers, and the paper never reports the Gilbert-Elliott burst parameters used at test time, even though training simulates only independent element-wise erasures.\n\nWhat the paper does well: the CALM conditioning idea is simple but clearly effective in the ablation; SCI is a cheap pre-packetization permute that shows a real independent gain; MGLR beats the linear interpolation baseline by a healthy margin at 30% loss. The qualitative figures are convincing—baselines fall apart under severe loss while ResPCC keeps recognizable geometry. The internal ablations compare against a strengthened DPCC backbone, so the proposed modules carry their own weight even if the headline gains over stock DPCC are partly capacity-driven.\n\nSoft spots, in proportion. The duplicated ablation row (Table 1, Category I row 3 and Category II row 1: same DBR/CALM/MGLR/Linear/SCI settings, different Bpp and PSNR) is a factual error that any referee will catch; it undercuts confidence in the table until fixed. There are no error bars or significance tests anywhere, which is normal for this venue but worth a sentence in revision. The bigger conceptual gap is the one the stress-test flags: training samples the mask element-wise i.i.d., while testing uses Gilbert-Elliott bursty loss at the packet level, and no G-E transition probabilities or mean burst lengths are reported. If the default burst length is short, the gap may be small; if long, MGLR may have been evaluated in a regime the model was never optimized for. That is not fatal—the SCI design is explicitly meant to bridge just this gap—but the paper should show a sensitivity sweep over burstiness before claiming “practical networks.” The “first” claim is plausible within the cited literature; I would soften it to “to our knowledge” regardless.\n\nBottom line: this paper deserves a serious referee. It is a solid, reproducible engineering advance with addressable issues. The right outcome is heavy revision, not rejection. I would send it to peer review and ask for the ablation fix, the G-E parameters with a burst-length sensitivity study, and error bars on the main curves.","headline":"Solid loss-resilient point cloud codec with a real, fixable reporting flaw; the burst-loss sensitivity gap is the main thing to probe before believing the strong practical claims.","tokens_in":21841,"tokens_out":2427,"would_cite":true,"duration_ms":27333,"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":"ResPCC claims that point cloud codecs can be made intrinsically resilient to packet loss by conditioning on loss rate and restoring scattered missing latent features; tests on ShapeNet and SemanticKITTI at 5–30% loss support this.","keywords":["point cloud compression","loss resilience","packet loss","latent restoration","error concealment","rate-distortion optimization","interleaving","neural codec"],"falsifier":"Train or test with the same average loss rate but long bursts (e.g., 50–200 consecutive lost packets) and compare reconstruction PSNR against the Gilbert-Elliott results; ResPCC's margin should shrink or vanish if its restoration depends on scattered single-entry gaps rather than true burst resilience. A second check is to corrupt or drop the protected anchor and hyperprior packets and observe whether reconstruction collapses entirely.","tokens_in":20847,"feed_emoji":"☁️","tokens_out":4850,"duration_ms":45593,"temperature":0.7,"pith_summary":"ResPCC is a learned point cloud codec that treats packet loss as part of the coding problem rather than an afterthought. The paper claims it is the first end-to-end neural codec with intrinsic loss resilience for intra-frame geometry: a single model, conditioned on an estimated loss rate, keeps reconstruction quality high when 5% to 30% of packets are lost. The design scatters loss damage across the latent representation, restores missing entries with graph and dictionary priors, and trains under simulated erasures. If the claim holds, point cloud delivery over lossy networks no longer needs per-loss-rate retraining or heavy forward error correction on the main geometry bitstream.","feed_headline":"One neural codec holds point clouds together through 30% packet loss","feed_subtitle":"A single loss-rate-aware model scatters the damage, restores missing latent features, and beats standard codecs on ShapeNet and…","key_machinery":"The load-bearing mechanism is the pairing of Spatial-Channel Interleaving (SCI) with Mask-Aware Graph-based Latent Restoration (MGLR). SCI reorders Morton-sorted latent features in 2x2x4 blocks so one lost packet no longer wipes out an entire channel for every anchor; instead the decoder sees scattered missing entries, each surrounded by surviving neighbors. MGLR then builds a K-nearest-neighbor graph over anchor coordinates, weights edges by mask reliability and spatial distance, and writes corrections only into masked positions, leaving received entries untouched. Two supporting modules complete the chain: Condition-Adaptive Latent Modulation (CALM) injects the loss rate into each downsampling stage through gating factors, and Dictionary-based Refinement (DBR) adds a learned cross-attention residual that pulls restored features toward canonical latent patterns.","core_discovery":"The central claim is that a point-cloud codec can be made intrinsically loss-resilient by conditioning encoding on the perceived packet loss rate and by restoring corrupted latent features before reconstruction. Concretely, ResPCC modulates latent generation with the loss rate, interleaves the latent tensor so channel-wise packet erasures become scattered element-wise gaps, restores those gaps with a mask-aware graph network using nearby anchors, and refines the result against a learned dictionary. Evaluated under Gilbert-Elliott bursty loss at 5% to 30%, it reports consistently higher D1/D2 PSNR than G-PCC, DPCC, PCGCv2, and OctAttention, with gains that grow as loss increases. On ShapeNet at 30% loss the reported advantage over DPCC is roughly 14.6% in D1 and 19.3% in D2, and the advantage over OctAttention exceeds 50% in both metrics.","pith_inferences":["The interleaving idea likely transfers to other latent-based codecs, since the ablation shows SCI alone improves PSNR even without a restoration network.","A natural stress test is burst-length generalization: training uses independent erasures while testing uses Gilbert-Elliott bursts, so varying burst length at a fixed loss rate would reveal whether interleaving or restoration is doing the heavy lifting.","The method suggests a middle path between pure source coding and full channel coding: protect only the structural metadata and let a loss-aware decoder absorb moderate latent erasures, which could be combined with unequal error protection in practice."],"forward_implications":["A single ResPCC model trained once can serve a range of loss conditions, since the loss rate is an input rather than a training-scheme parameter.","The codec reduces the need for redundant FEC on latent data; only anchors and hyperprior are protected, and these cost a small fraction of the bitrate.","Performance degrades gracefully: the reported gap over DPCC widens from about 5–9% at 5% loss to about 15–19% at 30% loss on ShapeNet.","The approach transfers across object-like and LiDAR scenes without retraining, as shown by similar trends on ShapeNet and SemanticKITTI."],"supporting_citations":[{"why":"Supplies the DPCC point-based hierarchical autoencoder that ResPCC builds on as its backbone.","marker":"[18]"},{"why":"Establishes the intrinsic resilience idea and the simulated-loss training strategy that ResPCC adapts.","marker":"[7]"},{"why":"Provides the Gilbert-Elliott bursty loss model used to generate test packet loss patterns.","marker":"[17]"},{"why":"Supplies the variational hyperprior used for entropy coding and rate estimation.","marker":"[1]"},{"why":"Defines the MPEG D1/D2 PSNR and Bpp metrics used in evaluation.","marker":"[24]"},{"why":"ShapeNet, the object point cloud dataset used for training and evaluation.","marker":"[5]"},{"why":"SemanticKITTI, the LiDAR scene dataset used to show transfer.","marker":"[2]"},{"why":"OctAttention, the octree baseline whose instability the comparisons highlight.","marker":"[11]"},{"why":"PCGCv2, the voxel-based baseline compared in the evaluation.","marker":"[38]"},{"why":"G-PCC standard reference, the traditional codec baseline.","marker":"[14]"}],"fun_headline_variants":["Loss-resilient point clouds: ResPCC endures 30% packet loss","Neural codec withstands packet loss up to 30%","ResPCC: first end-to-end codec for lossy networks","How to keep 3D point clouds intact under 30% loss","Loss-aware codec scatters damage, restores, survives 30% loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scheme assumes packet loss can be captured by a single rate p that is known to both encoder and decoder and that manifests as independent erasures during training, with anchors and hyperprior protected by FEC; if real loss is burstier, p is badly estimated, or that protection fails, the reported resilience does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Loss-resilient point clouds: ResPCC endures 30% packet loss","Neural codec withstands packet loss up to 30%","ResPCC: first end-to-end codec for lossy networks","How to keep 3D point clouds intact under 30% loss","Loss-aware codec scatters damage, restores, survives 30% loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000848,"raw_usage":{"total_tokens":3724,"prompt_tokens":1016,"completion_tokens":2708,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":2608}},"tokens_in":632,"tokens_out":2708,"duration_ms":20732,"temperature":1.0,"reasoning_tokens":2608,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:25:25.110104+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train or test with the same average loss rate but long bursts (e.g., 50–200 consecutive lost packets) and compare reconstruction PSNR against the Gilbert-Elliott results; ResPCC's margin should shrink or vanish if its restoration depends on scattered single-entry gaps rather than true burst resilience. A second check is to corrupt or drop the protected anchor and hyperprior packets and observe whether reconstruction collapses entirely.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"G-PCC standard reference, the traditional codec baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"OctAttention, the octree baseline whose instability the comparisons highlight."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PCGCv2, the voxel-based baseline compared in the evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DPCC point-based hierarchical autoencoder that ResPCC builds on as its backbone."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the intrinsic resilience idea and the simulated-loss training strategy that ResPCC adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Gilbert-Elliott bursty loss model used to generate test packet loss patterns."},{"cited_title":"2021.MPEG 3D Graphics Coding: Common test condi- tions for G-PCC","cited_arxiv_id":null,"evidence_quote":"Defines the MPEG D1/D2 PSNR and Bpp metrics used in evaluation."}],"review_version":1}