{"id":"263a2791-3586-4e48-87d8-1648e2fb93e8","arxiv_id":"2411.11199","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The BVI-CR dataset contributes 18 multi-view RGB-D human captures with textured meshes, plus a benchmark where INR codecs beat the MPEG TMIV anchor by up to 38.5% BD-rate.","lead":"A new open dataset, BVI-CR, adds 18 multi-view RGB-D human action sequences with textured 3D meshes to a field that lacked compression-oriented volumetric content. Benchmarks on it show two neural-representation codecs beat the MPEG TMIV anchor by up to about 38 percent in average BD-rate PSNR.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Compression benchmark may be asymmetric: depth and bitrate accounting for INR codecs is unspecified, so the 38.5% BD-rate gain over TMIV is not fully verified.","rationale":"The reader's weakest assumption points to the accuracy of the reconstructed meshes from the TSDF pipeline. That is a legitimate concern for downstream mesh-based tasks, but it is not the most load-bearing assumption for the paper's central compression claim. The Section III benchmark operates on multi-view RGB-D data, not on the supplied meshes; mesh geometry errors do not directly affect the BD-rate comparison because both codecs are evaluated on the same source views and synthesized views. The load-bearing assumption is that the rate-distortion evaluation treats TMIV and the INR codecs symmetrically, including how depth is used for view synthesis and how bitrate is defined. The paper's silence on these details, coupled with the explicit internal inconsistencies in FPS and depth resolution, makes this asymmetry a concrete and testable risk. I do not claim the reported gains are false; the text is plausible and the results are consistent across sequences. However, without code or a precise protocol description, the 38.5% BD-rate figure cannot be independently verified. This supports the reader's CONDITIONAL verdict, but for a different reason than the one highlighted in the weakest_assumption field, hence partial agreement. The appropriate next step is to request the release of code/data and a detailed evaluation protocol, not to reject the paper, since no evidence of fraud or an obviously incorrect calculation is present.","tokens_in":7893,"tokens_out":10903,"duration_ms":109175,"concrete_test":"Once BVI-CR data and benchmark code are released, re-run the Section III comparison on a subset of sequences (e.g., sequences 1, 7, 16) with a strict protocol: feed the TMIV view synthesizer with the decoded depth maps generated by each codec (not original depth), compute INR bpp including all model parameters and latent codes for both texture and depth, and calculate BD-rate over at least four RD points. If MV-HiNeRV's BD-rate gain remains near -38.5% under these conditions, the central claim is supported; if the gain shrinks substantially, the reported comparison is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section III and Table II, that MV-HiNeRV reduces BD-rate by 38.5% over TMIV, depends entirely on the rate-distortion comparison being symmetric. The paper does not state whether the depth maps used for the four synthesized evaluation views are the decoded depth produced by each codec or the original uncompressed depth. Section III only says synthesized views are rendered 'given the decoded texture attributes (of ten source views),' with no mention of decoded depth, and it does not define how INR bitrate is counted (e.g., whether model parameters and latent codes for depth are included, and how the background mask affects bpp). Internal inconsistencies reinforce this concern: the abstract and Section II-A say 30 FPS and 640×576 depth, while Table I says 25 FPS and Section II-B says 512×424 depth. If the INR codecs effectively use uncompressed depth, or if their bitrate omits depth or excludes background pixels while TMIV is charged for full-resolution depth and background, the reported gains are inflated and the benchmark comparison is invalid.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces BVI-CR, a multi-view volumetric human dataset consisting of 18 sequences (15 single-person, 3 multi-person) captured with 10 synchronized RGB-D cameras, together with background masks and textured polygonal meshes generated by a TSDF fusion and non-rigid registration pipeline. The authors also benchmark one conventional codec (TMIV) and two implicit-neural-representation (INR) codecs (MV-HiNeRV and MV-IERV) on this dataset, reporting BD-rate savings that reach 38.5% for MV-HiNeRV relative to the TMIV anchor. The paper positions BVI-CR as an open resource for volumetric video compression, reconstruction, and quality assessment.","tokens_in":8112,"tokens_out":3303,"duration_ms":34856,"significance":"If the dataset is released as described and the benchmark is fair, BVI-CR would fill a real gap: existing volumetric human datasets either lack multi-view depth, lack textured meshes, have limited multi-person content, or are not oriented toward compression evaluation. The inclusion of 10 synchronized RGB-D views plus meshes and masks, with diverse actions and occlusions, is genuinely useful. The benchmark also provides evidence that INR-based codecs can substantially outperform the MPEG test model on multi-view volumetric video, which is a significant and timely result. The paper's explicit parallelization with MIV CTC is a strength, as is the use of multiple metrics including IV-PSNR. However, the current manuscript does not yet make the dataset accessible or verifiable, and the benchmark protocol is insufficiently specified for the headline claim to be fully trusted.","major_comments":[{"comment":"The central claim that MV-HiNeRV achieves a 38.5% average BD-rate reduction over TMIV depends on a symmetric rate-distortion comparison, but the paper does not specify how depth is handled in the rendering and how INR bitrate is counted. The text states that synthesized views are rendered 'given the decoded texture attributes (of ten source views)' but does not say whether the depth maps input to the TMIV view synthesizer are the decoded depth produced by each codec or the original uncompressed depth. Since TMIV encodes both texture and depth with VVenC, any asymmetry here would directly inflate or deflate the reported gains. In addition, the bitrate definition for MV-HiNeRV and MV-IERV is not given: the reader cannot tell whether model parameters, latent codes, and background masks are included, and whether bpp is computed over the full frame or only the foreground region. Please provide the exact rendering input per codec and a complete, equation-level definition of bitrate for all three codecs.","section":"Section III, Table II"},{"comment":"The paper contains internally inconsistent capture specifications that affect both the dataset documentation and the benchmark calculations. Section II-A states that the cameras capture 2560x1440 RGB at 30 FPS and that the ToF depth sensor produces 640x576 depth images at the same frame rate, while Section II-B and Table I report 25 FPS and 512x424 depth resolution, and the abstract says 30 FPS. These discrepancies propagate to the stated frame counts (300-450 frames) and to any bitrate or BD-rate computation. The authors should clarify the native capture resolution and frame rate, and describe any resampling or cropping applied before releasing the RGB and depth data.","section":"Section II-A, Section II-B, Table I"},{"comment":"The central deliverable of the paper is the dataset itself, but the provided link (https://github.com/fan-aaron-zhang/bvi-cr) is not shown to contain any data or code, and no benchmark code is released. This makes it impossible for a reader to verify the dataset's content, format, or the reproducibility of the reported benchmark. Moreover, the volumetric reconstruction pipeline (TSDF fusion, non-rigid registration, marching cubes, UV mapping) is described only qualitatively, with no quantitative measure of reconstruction accuracy against an independent reference. Since all downstream compression results inherit the geometry and texture quality of these meshes, the paper should include either a data release with actual download instructions, dataset statistics and validation metrics, or an explicit, measured discussion of reconstruction error and its limitations.","section":"Section II, dataset availability and validation"}],"minor_comments":[{"comment":"There are several typographical errors in the text, including 'for for' in the second paragraph, 'mutli-camera' in Section II-A, 'a a set' in Section II-A, and 'persumably' in Section III; these should be corrected.","section":"Introduction"},{"comment":"The RD curves in Figure 4 and the qualitative comparison in Figure 5 do not identify which sequence or sequences they correspond to; since the paper reports per-sequence variation, please state the sequence index in the captions.","section":"Figure 4 and Figure 5"},{"comment":"The statement that the benchmark 'follow[s] the MPEG MIV Common Test Conditions' is vague; the paper does not specify the quantization parameter ladders, rate points, or the exact version of TMIV and VVenC used, which are needed to reproduce the results.","section":"Section III"},{"comment":"The description of the depth-to-RGB alignment is incomplete: the paper lists different resolutions for RGB and depth, but does not explain how they are registered or whether the released depth maps are upsampled to the RGB resolution or kept native.","section":"Section II-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is from a group with a strong track record in video compression, and the dataset concept is valuable. However, as a dataset paper, the lack of an accessible or even described data release is unusual and blocks validation of the central contribution. The benchmark asymmetry concern (depth handling and INR bitrate accounting) is serious and must be resolved before acceptance. I would also suggest that the editor or reviewers ask for a direct comparison of the INR codecs' bitrate components (texture, depth, model, masks) to ensure the headline BD-rate numbers are apples-to-apples. The fact that one of the benchmarked codecs, MV-HiNeRV, is from the same group should be explicitly disclosed in the paper, though this is not itself a reason for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper is that the dataset is real and useful: 18 sequences of 10-view 1080p RGB-D capture with textured meshes, background masks, and calibration, built for compression work. That combination is not available elsewhere, and the paper does a solid job describing the capture rig, the reconstruction pipeline, and the content. The benchmark itself is routine — application of MIV common test conditions to three codecs — but that is fine for a dataset paper. The BD-rate table is internally consistent and the qualitative examples line up with the numbers.\n\nThe soft spots are real but mostly fixable. First, the dataset is not accessible: the preprint points to a GitHub link with no actual data, no checksums, no preview. Since the dataset is the central artifact, this alone forces a conditional verdict. Second, the benchmark's rate-distortion comparison is under-specified in exactly the way that matters. The text says synthesized views are rendered from decoded texture attributes but never says whether depth is decoded or original for the INR codecs, and it never defines how INR bitrate is counted — model parameters? latent codes? background masks? If MV-HiNeRV and MV-IERV are charged for texture only while TMIV is charged for full-resolution texture plus depth, the reported 38.5% gain is inflated. I read the paper trying to resolve this and could not. That omission is load-bearing, not a nitpick. Third, there are internal inconsistencies: 30 FPS in the abstract and Section II-A, 25 FPS in Table I; 640×576 depth in II-A, 512×424 in II-B. These are easy to correct but currently undermine confidence.\n\nThe lack of an independent reconstruction accuracy check is a fair concern, but for a dataset paper it is acceptable at this stage because the raw RGB-D is what most users will work from and the meshes are a bonus. It should still be addressed before final release.\n\nWho is this for? Anyone doing volumetric video compression, neural rendering, or dataset-driven 3D quality assessment. It is a within-subfield resource, not a breakthrough, but it fills a gap.\n\nMy recommendation: send it to peer review with a request for the dataset to be made available, the benchmark details to be specified (especially depth handling and INR bitrate accounting), and the internal inconsistencies fixed. The core idea is sound and the resource is needed. I would not desk-reject it, but I would not accept it in its current form either.","headline":"A genuinely new multi-view RGB-D dataset that deserves a serious referee, but the benchmark claims need more detail and the data must actually be released before the contribution can be verified.","tokens_in":771,"tokens_out":1916,"would_cite":true,"duration_ms":28622,"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":"The paper introduces BVI-CR, an 18-sequence multi-view RGB-D human dataset with textured meshes, and benchmarks three codecs under MPEG MIV conditions, finding that the neural codec MV-HiNeRV cuts bitrate by an average 38.5% in PSNR…","keywords":["volumetric video","multi-view RGB-D dataset","textured mesh","implicit neural representation","video compression","BD-rate","immersive video","free-viewpoint video"],"falsifier":"A reader could take a few BVI-CR sequences, compare the provided meshes to a high-precision independent 3D scan of the same actor and pose (e.g., a structured-light or laser scan), and compute per-vertex or surface-to-surface error; if the reconstruction error is comparable to or larger than the compression quality differences between the codecs, the benchmark conclusions would be undermined.","tokens_in":7720,"feed_emoji":"🎥","tokens_out":4719,"duration_ms":37860,"temperature":0.7,"pith_summary":"This paper presents BVI-CR, a public dataset of 18 multi-view RGB-D human sequences with corresponding textured polygonal meshes, designed to support volumetric video compression research. Using this dataset, the authors benchmark three codecs under MPEG MIV Common Test Conditions: the MPEG test model TMIV and two implicit neural representation (INR) codecs, MV-HiNeRV and MV-IERV. The central result is that MV-HiNeRV achieves the best rate-distortion performance across all metrics and sequences, with an average 38.5% BD-rate reduction in PSNR and 34.3% in IV-PSNR relative to TMIV, while MV-IERV also outperforms TMIV by 29.5% and 26.9% respectively. If this holds, the dataset offers a credible public platform for developing and comparing volumetric video codecs, and strengthens the case for INR-based compression as a practical alternative to conventional multi-view coding.","feed_headline":"Neural codecs beat MPEG test model by 38.5% on new volumetric dataset","feed_subtitle":"18 multi-view RGB-D human sequences plus textured meshes for testing volumetric video codecs.","key_machinery":"The central object is the BVI-CR dataset itself, built from 10 synchronized Azure RGB-D cameras arranged on a hemisphere around a 2.5m capture area, with volumes reconstructed via TSDF fusion with non-rigid registration, meshes extracted by marching cubes, and textures generated through triangle-based UV mapping. The benchmark machinery is the MPEG MIV common test condition: all ten source views are encoded and decoded, four additional front-facing synthesized views are rendered by the TMIV default view synthesizer, and rate-distortion performance is compared using PSNR, SSIM, IV-PSNR, and Bjøntegaard $\\Delta$ Rate (BD-rate) relative to the TMIV anchor. The two INR codecs, MV-HiNeRV and MV-IERV, are the competing neural representations whose reconstruction quality is evaluated against these metrics.","core_discovery":"On the paper's own terms, the discovery is that the newly released BVI-CR dataset, comprising 18 diverse multi-view RGB-D captures with corresponding textured meshes, enables a fair benchmark of volumetric video codecs, and under this benchmark the INR-based codec MV-HiNeRV outperforms the MPEG MIV anchor TMIV by an average of 38.5% in BD-rate measured by PSNR and 34.3% measured by IV-PSNR, with MV-IERV also showing large gains of 29.5% and 26.9%. The claim is that neural, coordinate-based representations are a promising and competitive alternative to conventional multi-view coding on this data.","pith_inferences":["The reported gains may partly depend on the reconstruction pipeline's output being treated as ground truth; if those meshes were compared against an independent high-resolution 3D reference, the compression-quality conclusions could shift.","Because BD-rate is computed with TMIV as the single anchor, the results may not generalize to other conventional codecs or to end-to-end volumetric rendering quality; a natural extension is to add subjective quality assessment on rendered views.","The dataset could be extended with audio tracks, as the authors plan, enabling joint audiovisual volumetric compression research.","A testable extension is to use BVI-CR to pretrain generalizable neural codecs across sequences and measure cross-content generalization, which would indicate how well the reported gains transfer to unseen captures."],"forward_implications":["A public dataset of 18 sequences lets researchers train and validate volumetric compression and reconstruction methods without needing to build their own multi-camera capture rig.","The reported BD-rate reductions suggest that INR-based codecs can deliver the same quality at roughly 38% lower bitrate than TMIV on this content, pointing toward practical neural volumetric video delivery.","The inclusion of multi-person interactions and occluded actions provides stress cases where the relative gains of INR codecs are smaller, indicating where further codec improvements are needed.","The availability of camera extrinsics, depth maps, background masks, and textured meshes supports downstream tasks beyond compression, including quality assessment and neural rendering."],"supporting_citations":[{"why":"Defines the MPEG immersive video coding standard (MIV) and the TMIV anchor used as the baseline codec and test conditions.","marker":"[7]"},{"why":"Describes MV-HiNeRV, the INR-based codec that achieves the best reported performance.","marker":"[16]"},{"why":"Describes MV-IERV, the other INR-based codec benchmarked against TMIV.","marker":"[14]"},{"why":"Provides the Bjøntegaard Delta Rate (BD-rate) method used to compute average bitrate savings.","marker":"[33]"},{"why":"Defines IV-PSNR, the immersive video quality metric used for evaluation.","marker":"[32]"},{"why":"Supplies the non-rigid registration and TSDF fusion approach on which the volumetric reconstruction pipeline is based.","marker":"[28]"},{"why":"Provides the marching cubes algorithm used to extract meshes from the fused volume.","marker":"[29]"},{"why":"The VVenC encoder used to encode texture and depth in the TMIV anchor configuration.","marker":"[31]"}],"fun_headline_variants":["BVI-CR: Neural codecs beat MPEG by 38.5% in BD-rate","New dataset BVI-CR shows neural codecs outperform MPEG","On BVI-CR, neural compression wins vs MPEG by 38.5%","BVI-CR benchmark: INR codecs gap MPEG by 38.5%","BVI-CR: 38.5% bitrate cut with neural codecs over MPEG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark and dataset utility rest on the assumption that the TSDF fusion with non-rigid registration followed by marching cubes and UV mapping produces accurate ground-truth geometry and texture for all 18 sequences, and no independent quantitative reconstruction accuracy check is reported.","fun_headline_variants_meta":{"raw":{"variants":["BVI-CR: Neural codecs beat MPEG by 38.5% in BD-rate","New dataset BVI-CR shows neural codecs outperform MPEG","On BVI-CR, neural compression wins vs MPEG by 38.5%","BVI-CR benchmark: INR codecs gap MPEG by 38.5%","BVI-CR: 38.5% bitrate cut with neural codecs over MPEG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000922,"raw_usage":{"total_tokens":3967,"prompt_tokens":972,"completion_tokens":2995,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":2880}},"tokens_in":588,"tokens_out":2995,"duration_ms":19612,"temperature":1.0,"reasoning_tokens":2880,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:47:52.966355+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could take a few BVI-CR sequences, compare the provided meshes to a high-precision independent 3D scan of the same actor and pose (e.g., a structured-light or laser scan), and compute per-vertex or surface-to-surface error; if the reconstruction error is comparable to or larger than the compression quality differences between the codecs, the benchmark conclusions would be undermined.","supporting_citations":[{"cited_title":"Mpeg immersive video coding standard,","cited_arxiv_id":null,"evidence_quote":"Defines the MPEG immersive video coding standard (MIV) and the TMIV anchor used as the baseline codec and test conditions."},{"cited_title":"Immersive video compression using implicit neural representations,","cited_arxiv_id":null,"evidence_quote":"Describes MV-HiNeRV, the INR-based codec that achieves the best reported performance."},{"cited_title":"Implicit-explicit Integrated Representations for Multi-view Video Compression","cited_arxiv_id":"2311.17350","evidence_quote":"Describes MV-IERV, the other INR-based codec benchmarked against TMIV."},{"cited_title":"Calculation of average psnr differences be- tween rd-curves,","cited_arxiv_id":null,"evidence_quote":"Provides the Bjøntegaard Delta Rate (BD-rate) method used to compute average bitrate savings."},{"cited_title":"Iv-psnr—the objective quality metric for immersive video applications,","cited_arxiv_id":null,"evidence_quote":"Defines IV-PSNR, the immersive video quality metric used for evaluation."},{"cited_title":"Dynamicfusion: Reconstruction and tracking of non-rigid scenes in real-time,","cited_arxiv_id":null,"evidence_quote":"Supplies the non-rigid registration and TSDF fusion approach on which the volumetric reconstruction pipeline is based."},{"cited_title":"Marching cubes: A high reso- lution 3d surface construction algorithm,","cited_arxiv_id":null,"evidence_quote":"Provides the marching cubes algorithm used to extract meshes from the fused volume."},{"cited_title":"Vvenc: An open and optimized vvc encoder implemen- tation,","cited_arxiv_id":null,"evidence_quote":"The VVenC encoder used to encode texture and depth in the TMIV anchor configuration."}],"review_version":1}