{"id":"f7783763-db71-4491-bbf2-1ab6fb6c625d","arxiv_id":"2504.14240","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A dual-layer point cloud codec with ROI masking and detection-aware training reports better detection on compressed indoor scenes at high bitrates, but evaluation caveats weaken the claim.","lead":"Researchers present a point cloud compression method that spends extra bits on object regions, guided by a segmentation network, and that trains the codec together with an object detector. They report better geometry compression and higher detection accuracy on indoor datasets, which matters for autonomous driving and augmented reality where compressed 3D data must stay machine-readable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 10% detection gain is not established because the evaluation uses oracle ROI masks on SUN RGB-D and jointly optimizes RPCGC with the detector while baselines use an off-the-shelf detector; the gain may come from protocol, not the codec.","rationale":"The paper's distinct contribution is the machine-vision gain, and that gain is claimed in the abstract as '10% gain' in detection accuracy. The evaluation protocol is the load-bearing support for that claim. Two explicit decisions in the paper undermine the comparison. First, using oracle segmentation labels on SUN RGB-D means the ROI mask is not predicted but taken from ground truth; a practitioner would not have those labels at inference time. Second, only RPCGC is trained jointly with the detection loss and detector, while baselines are evaluated with an off-the-shelf detector. These two choices are exactly the variables that the paper's ablation studies show influence detection performance: adding detection loss and mask weighting. Therefore the observed mAP advantage cannot be attributed to the compression method itself. A matched evaluation would either fix the detector and training procedure across all codecs, or jointly optimize every codec with the detector. Neither is done. The compression fidelity results in Table 1 and Fig. 3 are less affected, but the paper's headline claim is the machine-vision gain, and that claim is not supported as stated. The reader's weakest assumption identifies the same issue, and I agree with the REJECT verdict: the central claim fails under the reported protocol, and the paper should not be accepted as evidence for the claimed detection improvement.","tokens_in":16422,"tokens_out":2226,"duration_ms":21426,"concrete_test":"Re-run the Table 2 comparison with a matched protocol: (1) fine-tune or train the Group-Free detector on reconstructions from each baseline codec (G-PCC, AVS, PCGCv2, SPCGC) using the same training data, iterations, and hyperparameters; (2) replace oracle segmentation labels with RPN predictions on SUN RGB-D; (3) report mAP@0.25 and mAP@0.5 at matched bpp. If RPCGC still exceeds every baseline by roughly the same margin, the concern is resolved; if the gap collapses, the headline detection gain is not supported.","verdict_should_be":"REJECT","load_bearing_attack":"The central machine-vision claim rests on Table 2, but the comparison does not isolate the ROI-guided compression mechanism. Section 4.1 explicitly states that on SUN RGB-D the original segmentation labels are used instead of RPN predictions because the RPN mIoU is only 60%, so the mask is oracle information not available to a deployed codec. Section 4.2 then states that baseline codecs are paired with an off-the-shelf Group-Free detector, while RPCGC uses joint optimization with detection loss. This means the reported mAP gap could reflect detector fine-tuning and oracle mask supervision rather than a compression advantage. The ablation in Fig. 5(d) supports this worry: adding detection loss markedly improves detection performance, so the joint loss alone may explain most of the apparent gain. No matched protocol is reported in which every codec is paired with an equally trained detector and in which predicted masks are used for RPCGC.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an ROI-guided point cloud geometry compression framework (RPCGC) that combines a G-PCC-based base layer with a learned enhancement layer. An ROI prediction network (RPN) and ROI searching network (RSN) generate per-point masks that weight both the residual features and a Chamfer-distance distortion term, and a detection loss from Group-Free is added to the rate-distortion objective. The authors claim exceptional compression performance and a roughly 10% detection accuracy improvement over learning-based methods at high bitrates on ScanNet and SUN RGB-D. The compression results are reported via RD curves and BD-PSNR, and detection results via mAP@0.25 and mAP@0.5.","tokens_in":16636,"tokens_out":7979,"duration_ms":73466,"significance":"If the central claim were supported, the ROI-guided bit allocation and joint detection loss would be a practically valuable contribution to human-and-machine point cloud compression. The paper has clear strengths: a complete dual-branch architecture with residual analysis/synthesis, a Multi-Scale Feature Extraction Module, a Semantic-aware Attention Module, ablation studies, and generalization experiments on MPEG datasets. The RD comparisons on ScanNet against learned codecs are informative. However, the central machine-vision claim is not established by the experiments as reported, because the detection evaluation uses oracle ground-truth masks on SUN RGB-D and jointly trains the detector for RPCGC while using an off-the-shelf detector for all baselines. The compression claims are also somewhat overstated relative to the reported BD-PSNR values. As a result, the main novelty—the ROI-guided mechanism—is not isolated or validated by the current protocol.","major_comments":[{"comment":"The SUN RGB-D experiments use the original segmentation labels instead of RPN-predicted masks because the RPN mIoU is only 60%. This is an oracle setting: the encoder is given object-location information that would not be available to a deployed codec. The claimed detection gain on SUN RGB-D is therefore an upper bound on the proposed pipeline, not its actual performance. Please report detection accuracy with predicted RPN masks or explicitly reframe these experiments as an oracle study and adjust the abstract and conclusion claims accordingly.","section":"Section 4.1"},{"comment":"The detection comparison is confounded by asymmetric detector training. Baselines are evaluated with an off-the-shelf Group-Free detector, while RPCGC's compression network and the detector are jointly optimized using the detection loss in Eq. (10). Figure 5(d) shows that adding detection loss alone markedly improves detection performance, so the reported mAP gap may be due to detector adaptation rather than to ROI-guided compression. A matched protocol—either a fixed detector for all codecs or the same fine-tuning procedure on compressed outputs for every codec—is needed to attribute the gain to the proposed mechanism.","section":"Section 4.2, Machine Vision Evaluation"},{"comment":"The detection table reports rows without bpp labels, stating that 'each algorithm computes detection accuracy at a similar bpp' without specifying the operating points. The claimed 'roughly 10%' improvement cannot be traced from the table, and several rows contradict it (e.g., ScanNet row 3 at mAP@0.25: RPCGC 0.616 vs G-PCC 0.629; SUN RGB-D row 1 at mAP@0.25: RPCGC 0.032 vs G-PCC 0.099). Please provide bpp values for each row and report results at explicitly matched bitrates.","section":"Table 2"},{"comment":"The compression performance claims are overstated. On SUN RGB-D, Table 1 shows RPCGC is worse than G-PCC in D2 PSNR (-0.185 dB) and worse than SPCGC in D1 PSNR (-0.126 dB); on ScanNet, D2 PSNR is -0.084 dB relative to G-PCC. The text should qualify 'exceptional compression performance' to the specific bitrate ranges and distortion metrics where RPCGC is actually superior.","section":"Abstract, Table 1, Figure 3"}],"minor_comments":[{"comment":"The information-bottleneck notation is inconsistent: \\bar{x} and t are not defined, and the variables v and t are not used consistently in the derivation.","section":"Section 3.2, Eq. (2)"},{"comment":"After describing the entropy encoder for the enhancement layer, the text says 'obtaining the BL bitstream'; this appears to be a typo for 'EL bitstream'.","section":"Section 3.4"},{"comment":"The caption says 'RPCGC-base represents a scenario where no optimization strategies are added'; please clarify which strategies are omitted (mask weighting, detection loss, or both) and whether the architecture is otherwise unchanged.","section":"Figure 3 caption"},{"comment":"The average encoding time for SPCGC is listed as 1.327 seconds, which is inconsistent with the ScanNet and SUN RGB-D values (0.738 and 0.589 seconds); please verify the average calculation.","section":"Table 3"},{"comment":"The claim of being 'the first to introduce' an ROI-guided point cloud compression scheme should be supported by a more detailed comparison with existing task-oriented point cloud compression works, such as [19], [21], [39], and [49], which also target machine vision.","section":"Section 1"}],"recommendation":"reject","confidential_remarks":"The paper's own text discloses both the oracle-mask protocol and the asymmetric joint detector training, so these are not hidden flaws. They are, however, load-bearing and invalidate the main comparison supporting the central claim. The evaluation would need to be redesigned with predicted masks and a matched detector-training protocol for all codecs; even then, the current claims might not survive. I recommend rejection of the current version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper has a real engineering contribution, but the central machine-vision claim is overstated because the detection comparison is not apples-to-apples. I would send it to peer review, but the authors need to redo the detection evaluation before the headline result can be trusted.\n\nWhat is actually new: the system itself—a dual-branch codec where the enhancement layer weights residual features by an instance-mask, and the RD loss weights Chamfer distance per-point by the same mask, with detection loss folded into training. That combination is not in the cited prior work. The compression RD results on ScanNet at high bitrates are plausible, and the ablation studies are honest: Fig. 5(d) shows that adding detection loss (DL) and the attention module (SAM) each move detection accuracy, which tells you something real about where the gain comes from. The writing is clear and the module descriptions are detailed enough to reproduce.\n\nWhere it falls down: the 10% detection gain in Table 2 comes from a protocol that rewards RPCGC in two separate ways. On SUN RGB-D, the ROI mask is computed from ground-truth segmentation labels because the RPN mIoU is only 60% (Section 4.1). That is oracle information no deployed codec would have. And in Section 4.2, the baselines are paired with an off-the-shelf Group-Free detector while RPCGC jointly optimizes with the same detector in its loss (Eq. 10). The ablation in Fig. 5(d) shows that just adding detection loss improves detection markedly—so the reported mAP gap could be mostly detector fine-tuning, not a compression benefit. The paper even notes this indirectly: RPCGC's rows at the lowest bitrate are below G-PCC and AVS, which is consistent with the codec not preserving enough geometry and the detector being trained to work around that. Table 2 also lacks explicit bpp labels for each row; the text says 'similar bpp' but does not show the numbers, so the 10% gain is not traceable.\n\nThe compression-fidelity part is less affected and is a reasonable contribution on its own. But the paper's stated novelty is the machine-vision gain, and that gain is not demonstrated by the evaluation as run. A matched protocol—same detector training for all codecs, predicted masks for RPCGC—could shrink or erase the gap.\n\nWho is this for: people working on learned point cloud compression for downstream tasks. The architecture ideas (mask-weighted residuals, per-point weighted CD) are worth studying even if the evaluation is currently not convincing. I would not cite the 10% claim, but I would cite the compression results after a careful re-read.\n\nRecommendation: send to peer review, but with the requirement that the detection comparison be redone fairly—equal detector training, predicted masks, and a table with exact bpp per row. Without that, the headline result should not stand.","headline":"A genuinely new ROI-guided point cloud codec with a plausible compression story, but the headline 10% detection gain is not supported by the evaluation as run.","tokens_in":17172,"tokens_out":1052,"would_cite":false,"duration_ms":11088,"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":"The paper proposes RPCGC, a point cloud geometry codec whose enhancement layer spends bits according to an ROI mask and whose loss includes 3D detection error, and claims about 10% higher detection mAP at high bitrates on ScanNet and SUN…","keywords":["point cloud geometry compression","region of interest","machine vision","3D object detection","rate-distortion optimization","residual coding","human and machine vision","learned compression"],"falsifier":"A controlled experiment would settle it: run RPCGC and the baselines on SUN RGB-D with masks predicted by the RPN (mIoU about 60%) rather than ground-truth labels, give every codec the same joint detection-loss fine-tuning, and compare mAP at matched bitrates; if the detection gain over baselines falls below the reported roughly 10%, the evaluation protocol rather than the ROI mechanism is carrying the claim.","tokens_in":16225,"feed_emoji":"🎯","tokens_out":7004,"duration_ms":58197,"temperature":0.7,"pith_summary":"Point cloud compression usually optimizes only human-visible fidelity, so the semantic detail that a downstream detector needs is the first thing lost at low bitrates. This paper proposes RPCGC, a dual-layer geometry codec in which a base layer stores a coarse G-PCC reconstruction and an enhancement layer codes residuals, with both layers steered toward object regions by an ROI mask produced by a segmentation-style prediction network. The mask enters twice: it multiplies the residual features before entropy coding, and it weights a Chamfer-distance distortion term per point, so bits are spent preferentially on foreground objects. A detection loss from a 3D object detector is added to the rate-distortion objective, jointly optimizing compression and perception. The paper reports that at high bitrates on ScanNet and SUN RGB-D this scheme beats several learning-based codecs on both rate-distortion and detection accuracy, with roughly 10% higher mAP.","feed_headline":"Object-aware point cloud codec lifts 3D detection by 10%","feed_subtitle":"Dual-layer compression weights residuals by ROI masks, keeping fidelity while helping detectors at high bitrates.","key_machinery":"The load-bearing object is the ROI mask produced by the pair (RPN, RSN): the RPN is a sparse-convolution U-Net that outputs a per-point probability of instance label, and the RSN converts that into a foreground/background mask matched to residual features by nearest-neighbor search, with foreground weighted doubly. This mask is the mechanism that carries the argument in two places: it multiplies the enhancement-layer residual feature map, written $x'_{enh} = x_{enh} * (1 + x_m)$, and it enters the distortion term as the Mask-weighted Chamfer Distance $D_{RW-CD}$, in which each point's squared distance is weighted by its predicted mask value. The same mask logic is reflected in the total loss $L_{total} = \\alpha D_{RW-CD} + \\beta R + \\gamma L_{detection}$, coupling rate, fidelity, and detection. A semantic-aware attention module and a multi-scale feature extraction module sit inside the residual codec and are shown by ablation to improve compression performance; the mask and detection loss are what move detection accuracy.","core_discovery":"On the paper's own terms, the central discovery is that ROI-guided bit allocation makes point cloud geometry compression serve both humans and machines: treating object regions as a mask and weighting both the residual features and the distortion metric by that mask yields a reconstruction that is better for 3D object detection without sacrificing visual fidelity. The mechanism is a two-layer encoding: the base layer uses G-PCC to code quantized coordinates, and the enhancement layer codes the residual between original and de-quantized coordinates, processed by residual analysis and synthesis networks. A sparse-convolution U-Net (the RPN) predicts per-point instance probabilities, and the RSN converts them into a foreground/background mask that is applied to residual features as strong supervision and to a mask-weighted Chamfer distance as weak supervision. Detection loss from the Group-Free detector is folded into the rate-distortion objective. The experiments show the gain is confined to higher bitrates: around 0.5-3 bpp the detection mAP rises by roughly 10% over the compared learning-based codecs, while at low bitrates contour information is lost and G-PCC/AVS detect better.","pith_inferences":["Editorial inference: because the SUN RGB-D evaluation uses ground-truth segmentation labels rather than the RPN's predicted masks (the paper reports RPN mIoU of only about 60%), the practical gain depends on how well the mask predictor does at inference; a fair field test should use predicted masks and report the drop.","Editorial inference: the baselines are not jointly optimized with the detector, so part of the 10% detection gap may come from joint optimization rather than from the ROI mechanism itself; retraining all codecs with the same detection loss would isolate the ROI contribution.","Editorial inference: the mask-weighted distortion idea is naturally a rate-allocation rule, so it could be extended to other downstream tasks such as classification and segmentation, and to variable-rate point cloud codecs, which the authors list as future directions.","Editorial inference: if the RPN could be made lightweight and run entirely on the encoder side, the same architecture would apply to LiDAR point clouds from autonomous driving, where the high-bitrate regime is common; this is a testable extension the paper does not carry out."],"forward_implications":["If RPCGC is right, point cloud codecs can be built with a two-layer structure where the enhancement layer's bit budget is explicitly steered by object masks, and this pays off specifically in the high-bitrate regime used by machine-analysis applications.","At high bitrates on ScanNet, RPCGC reports BD-PSNR gains of 9.129 dB over PCGCv2 and 1.211 dB over SparsePCGC on D1 PSNR, and detection mAP@0.25 rises to 0.562-0.616 compared with about 0.52-0.63 for the baselines, with similar average gains on SUN RGB-D.","The method is not a low-bitrate win: the paper states that at lower bitrates G-PCC and AVS detect better, and RPCGC is likely better suited for applications that require high bitrates specifications.","The ablation results imply that each component contributes differently: detection loss and mask weighting mainly lift detection, while SAM and MSFEM mainly lift reconstruction fidelity, so the design can be tuned per task.","The generalization experiments on MPEG/MVUB samples indicate the ROI and loss modules transfer to voxelized point clouds outside ScanNet and SUN RGB-D, beating traditional and early learned codecs on D2 PSNR."],"supporting_citations":[{"why":"Group-Free detector that supplies the detection loss and evaluation protocol; RPCGC is jointly optimized with it.","marker":"[22]"},{"why":"G-PCC codec used for the base-layer bitstream and as the main traditional baseline for rate-distortion and rate-detection comparisons.","marker":"[26]"},{"why":"PCGCv2 is a learning-based baseline that RPCGC is compared against for BD-PSNR and detection mAP.","marker":"[41]"},{"why":"SparsePCGC is a sparse-tensor learning-based baseline used in the same comparisons.","marker":"[40]"},{"why":"ScanNetv2 provides the indoor scene point clouds, instance labels, and bounding boxes used for training and testing.","marker":"[5]"},{"why":"SUN RGB-D provides the single-view indoor point clouds, semantic labels, and bounding boxes for the second benchmark.","marker":"[37]"},{"why":"Residual analysis and synthesis coding motivates the coarse-to-fine two-layer coding structure.","marker":"[27]"},{"why":"U-Net architecture is the basis of the ROI prediction network (RPN) used to generate masks.","marker":"[35]"},{"why":"Chamfer distance is the distortion metric that the paper reweights per point with the ROI mask.","marker":"[7]"},{"why":"AVS PCRM is another traditional codec baseline in the rate-distortion and detection comparisons.","marker":"[12]"}],"fun_headline_variants":["ROI masks steer point cloud codec to 10% better detection","Object-aware compression gives detectors a 10% boost","ROI-guided codec preserves detail and lifts detection 10%","Two-layer codec with ROI masks gains 10% mAP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported detection comparison isolates the ROI mechanism: on SUN RGB-D the encoder is given original segmentation labels instead of the RPN's own predictions, and RPCGC is jointly optimized with the detector while the comparison codecs are not, so a practitioner who has to predict masks and tune all codecs equally might see the roughly 10% detection gain shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["ROI masks steer point cloud codec to 10% better detection","Object-aware compression gives detectors a 10% boost","ROI-guided codec preserves detail and lifts detection 10%","Two-layer codec with ROI masks gains 10% mAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000605,"raw_usage":{"total_tokens":2851,"prompt_tokens":1004,"completion_tokens":1847,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":1774}},"tokens_in":620,"tokens_out":1847,"duration_ms":11972,"temperature":1.0,"reasoning_tokens":1774,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:53:38.835176+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment would settle it: run RPCGC and the baselines on SUN RGB-D with masks predicted by the RPN (mIoU about 60%) rather than ground-truth labels, give every codec the same joint detection-loss fine-tuning, and compare mAP at matched bitrates; if the detection gain over baselines falls below the reported roughly 10%, the evaluation protocol rather than the ROI mechanism is carrying the claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Group-Free detector that supplies the detection loss and evaluation protocol; RPCGC is jointly optimized with it."},{"cited_title":"Chou, David Flynn, Maja Krivokuća, Ohji Nakagami, and Toshiyasu Sugio","cited_arxiv_id":null,"evidence_quote":"G-PCC codec used for the base-layer bitstream and as the main traditional baseline for rate-distortion and rate-detection comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PCGCv2 is a learning-based baseline that RPCGC is compared against for BD-PSNR and detection mAP."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ScanNetv2 provides the indoor scene point clouds, instance labels, and bounding boxes used for training and testing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SUN RGB-D provides the single-view indoor point clouds, semantic labels, and bounding boxes for the second benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Residual analysis and synthesis coding motivates the coarse-to-fine two-layer coding structure."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"U-Net architecture is the basis of the ROI prediction network (RPN) used to generate masks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Chamfer distance is the distortion metric that the paper reweights per point with the ROI mask."},{"cited_title":"September, 2022","cited_arxiv_id":null,"evidence_quote":"AVS PCRM is another traditional codec baseline in the rate-distortion and detection comparisons."}],"review_version":1}