{"id":"fd83ff71-efaf-46e5-bbac-b4d8865e1bb2","arxiv_id":"2504.20383","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A hybrid disparity compensation module, combining shifted cost volumes with normalized cross-attention, improves neural stereo video compression by up to 55% bitrate over MV-HEVC on driving benchmarks.","lead":"The paper presents a neural stereo video compression system that combines explicit disparity shifts with an attention-based soft alignment to reduce redundant information between left and right views. It reports consistent bitrate reductions over prior stereo codecs on driving and general scenes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 26.81% average gain over LLSS depends on numbers from their papers, not a common-protocol rerun; protocol mismatch could materially change the comparison.","rationale":"The reader's weakest assumption is exactly the load-bearing issue: the headline empirical superiority over prior neural stereo codecs is not established under a common protocol. I found no more serious internal inconsistency in the HDC formulation; the shift equation can be read as a centered disparity representation, and the internal ablations (Figs. 6-8) support the module's contribution. However, because LSVC/LLSS numbers are cited from their original papers, protocol differences in I-frame codec, crop, GOP, and PSNR color space could materially change the relative gains. The paper's explicit statement of this limitation is transparent, but it means the SOTA claim remains conditional. The reader's CONDITIONAL verdict is appropriate, so no verdict change is needed.","tokens_in":18961,"tokens_out":14068,"duration_ms":146297,"concrete_test":"Using the same frame preprocessing, GOP (30/21/32), ELIC I-frame coder, and RGB-PSNR BD-rate anchor as the paper, run the released LSVC and LLSS models on Cityscapes and KITTI (and if possible Nagoya), then recompute Table I. If the gap to LLSS narrows by more than about 10 percentage points on KITTI 2012, the claimed 26.81% average gain is partly a protocol artifact rather than an HDC benefit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-C states: 'The results of LSVC [4] and LLSS [3] are reported from their original papers.' The central claim of outperforming previous neural SOTA thus rests on cross-paper BD-rate values, whose evaluation conditions (I-frame codec, cropping, GOP, lambda range, color-space PSNR) are not demonstrated to match this pipeline. The suspicious pattern in Table I—Ours vs LLSS is -53.13 vs -49.44 on Cityscapes but -55.97 vs -18.18 on KITTI 2012—suggests the cited LLSS numbers may have been generated under a less favorable protocol. Even the single-view DCVC-TCM* baseline beats LLSS by roughly 30 points on KITTI, so the reported 26.81% relative gain could stem largely from protocol differences rather than the HDC modules. A common-protocol rerun is required to validate the SOTA claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid disparity compensation (HDC) strategy for neural stereo video compression, combining explicit horizontal feature shifting with an implicit cross-attention mechanism. HDC is instantiated in two modules: HDC-FER, used in the encoders/decoders of the motion and context compression branches to align and fuse cross-view features, and HDC-EM, used in the entropy model to condition slice-wise autoregressive coding on aligned cross-view priors. The framework is built on a lightweight variant of DCVC-TCM. Experiments on Cityscapes, KITTI 2012/2015, and two Nagoya sequences report average BD-rate savings of 54.61% over MV-HEVC and 26.81% over the prior neural method LLSS on driving datasets, with ablations isolating the contributions of the two modules and of the explicit/implicit components.","tokens_in":19167,"tokens_out":6415,"duration_ms":67149,"significance":"If the reported gains hold under a common evaluation protocol, the work is a solid engineering contribution: the HDC design is clearly specified, the ablations are well structured (HDC-EM alone, HDC-FER alone, removal of the attention score, removal of the shift, replacement with Stereo/Mutual Attention, and a second backbone), and the complexity overhead relative to the single-view backbone is moderate. The extension to general-scene multi-view data (Nagoya) is a useful addition to a literature that has focused on driving scenes. However, the headline claim of outperforming prior neural stereo codecs rests on numbers cited from the original LSVC/LLSS papers rather than on a common-protocol re-run, and the paper's own Table I contains a pattern that casts doubt on protocol compatibility. The central architecture appears sound, but the evaluation needs to be reworked before the state-of-the-art claim can be accepted.","major_comments":[{"comment":"The core claim of surpassing the previous neural stereo SOTA (LLSS, LSVC) relies on BD-rate numbers cited from their original papers: the manuscript states in §IV-C, 'The results of LSVC [4] and LLSS [3] are reported from their original papers.' The present work uses the ELIC model of He et al. [58] for I-frames, a specific cropping protocol, a specific GOP, and a specific lambda schedule, and it is not demonstrated that the original LLSS/LSVC evaluations used the same conditions. A concrete red flag is Table I on KITTI 2012: the single-view backbone DCVC-TCM* already achieves -48.61% versus LLSS's -18.18%, a 30-point gap that is difficult to attribute to the stereo modules alone and suggests the LLSS numbers were obtained under a different test protocol (e.g., a different I-frame codec or color-space PSNR). Please re-run LSVC and LLSS under this paper's exact evaluation protocol, or provide a detailed, quantitative protocol-equivalence argument (I-frame codec, crop, GOP, lambda set, color space, and test frames).","section":"§IV-C, Table I; §IV-B"},{"comment":"The reported '26.81% average BD-rate reduction relative to LLSS' is computed as the simple arithmetic mean of the per-dataset differences in BD-rate percentage points: ((49.44-53.13) + (18.18-55.97) + (15.76-54.72))/3 = -26.81. This is not a standard relative BD-rate saving, because BD-rate is already a percentage relative to the MV-HEVC anchor. For example, on KITTI 2012 the actual relative saving of Ours over LLSS is (18.18 - 55.97)/(100 - 18.18) × 100 ≈ 46.2%, not 37.8 percentage points. Please report per-dataset relative BD-rate calculations using a standard formula, and recompute the average accordingly.","section":"§IV-C, Table I"},{"comment":"The backbone ablation claims that integrating HDC into the FVC backbone yields 'superior performance over prior SOTA approaches under identical backbone settings.' However, the comparison in Fig. 8 is made against LSVC and LLSS using their originally reported numbers; it is not established that those methods share the FVC backbone used for the Ours(FVC) variant. The 'identical backbone' statement is therefore only valid for the FVC-to-Ours(FVC) comparison, not for the comparison to LSVC/LLSS. Please either provide a controlled comparison in which LSVC and LLSS are re-run on the same backbone (e.g., by removing the HDC modules), or rephrase the claim to avoid implying that the cited LSVC/LLSS numbers were produced with the FVC backbone.","section":"§IV-E, Fig. 8"}],"minor_comments":[{"comment":"In the introduction, 'contexture coding' appears to be a typo for 'context coding' or 'contextual coding'; please correct it.","section":"§I"},{"comment":"The text contains an orphan fragment 'K L2.' in the description of intermediate features; it should read 'K^L' or be removed.","section":"§III-C1"},{"comment":"The figure captions contain visible 'uni00A0' artifacts (e.g., 'KITTI/uni00A02012'), which appear to be a PDF/LaTeX conversion issue; please ensure the final version renders proper spaces.","section":"Fig. 4 captions"},{"comment":"LSVC and LLSS have no entries for Kendo and Balloons; the claim that the method outperforms neural SOTA on general scenes is therefore unsupported by a direct neural baseline on those sequences. Please add such baselines or soften the claim.","section":"Table I"},{"comment":"The maximum disparity D is set to 192 based on KITTI's reported maximum disparity, but the Nagoya sequences have different resolutions (1024×768) and camera baselines; the paper does not discuss whether D=192 is appropriate for those sequences. A brief sensitivity analysis or justification is requested.","section":"§IV-B"},{"comment":"Table I reports BD-rate in RGB color space, but the original LLSS/LSVC papers may report in YUV/PSNR or a different color space; please verify and state explicitly that the cited numbers were computed in the same color space as this work.","section":"§IV-C"}],"recommendation":"major_revision","confidential_remarks":"The core architecture and ablations are sound, but the state-of-the-art comparison is not yet trustworthy. The KITTI 2012 discrepancy (single-view DCVC-TCM* at -48.61% vs LLSS at -18.18%) is a strong signal of protocol mismatch, and the paper itself discloses that the LLSS/LSVC numbers were not re-run. I would advise the editor to require a common-protocol re-evaluation of the neural baselines before considering acceptance, in addition to correcting the relative BD-rate calculation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's real contribution is the hybrid disparity compensation (HDC) design, and that part is solid. The module-level ablations are clean and the gains are consistent. The headline SOTA claim is weaker, because the LSVC and LLSS numbers are quoted from their original papers rather than rerun under this pipeline.\n\nThe mechanism itself is a sensible middle ground: build a horizontal-shift cost volume, compute a normalized dot-product similarity, and use that as attention weights for implicit cross-view alignment. Applying this in both feature extract/reconstruction and entropy modeling is new for stereo video, and the ablations make the case. On Cityscapes, HDC-EM alone gives a 26.6% BD-rate reduction, HDC-FER another 6.0%, and the two together 32.2%; removing the attention score or the shift operation costs 13.8% and 9.1% respectively. Those are internally valid numbers, and the latency comparison against Stereo and Mutual Attention is useful.\n\nThe soft spots are mostly the usual ones: no code, no error bars, and the general-scene evaluation is only two sequences (Kendo and Balloons). The bigger issue is the external baselines. The paper states plainly that LSVC and LLSS results come from their own papers. Table I shows a pattern that is hard to explain otherwise: DCVC-TCM*, the single-view backbone, beats LLSS by about 30 points on KITTI 2012 but is worse than LLSS by about 25 points on Cityscapes. That smells like protocol mismatch, not method behavior. A common-protocol rerun of at least LLSS is needed before I would put much weight on the 26.8% relative-gain claim. But the method does not need that claim to stand; the internal ablations and the comparison against its own single-view baseline carry the paper.\n\nTwo minor things: the abstract says the method outperforms all methods, while the intro says it is comparable to MV-HEVC on general scenes; and the FVC-backbone experiment is called 'identical backbone settings' even though LSVC and LLSS are not built on FVC, so that phrase overstates the control.\n\nBottom line: this deserves peer review, but the reviewers should request a shared-protocol rerun of the prior neural baselines and ideally code release. I would cite it for the HDC module design and the ablation approach.","headline":"A well-built hybrid disparity compensation module with genuinely clean ablations, but the headline SOTA numbers lean on other papers' baselines and need a common-protocol rerun before I'd trust them.","tokens_in":19673,"tokens_out":4346,"would_cite":true,"duration_ms":42972,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A hybrid disparity compensation method that combines explicit pixel shifts with implicit cross-attention reports a 54.61% average bitrate reduction over MV-HEVC on driving benchmarks and a 15.07% saving on general multi-view sequences.","keywords":["stereo video compression","neural video compression","disparity compensation","cross-attention","entropy modeling","rate-distortion optimization","multi-view video coding","autoregressive coding"],"falsifier":"Re-run the previous neural stereo methods and this one on the same preprocessed frames with the same I-frame compression code and the same rate-distortion operating points, then recompute BD-rate; if the gap over the previous method falls well below 26.81% or reverses, the paper's central comparative claim is not supported.","tokens_in":18795,"feed_emoji":"🎥","tokens_out":10634,"duration_ms":92293,"temperature":0.7,"pith_summary":"The paper tries to show that stereo video compression can be improved by hybridizing the two main ways of removing cross-view redundancy: explicit horizontal pixel shifts, which provide a cheap structural prior, and implicit cross-attention, which captures a wider range of matches. The proposed HDC strategy builds feature volumes by shifting each view horizontally, fuses them with a dot product to form a similarity map, normalizes that map into an attention score, and uses the score to warp one view toward the other. On autonomous-driving benchmarks the full system reports an average 54.61% bitrate reduction over MV-HEVC and a 26.81% reduction over the previous neural stereo method; on general multi-view sequences it reports a 15.07% bitrate saving over MV-HEVC. If these results hold, stereo and multi-view video could be stored or transmitted at a fraction of the current bitrate without sacrificing reconstruction quality.","feed_headline":"Cut stereo video bitrate by 54.61% vs MV-HEVC","feed_subtitle":"Explicit disparity shifts plus learned cross-attention removes cross-view redundancy in driving and general scenes.","key_machinery":"The load-bearing mechanism is the HDC module: a hybrid block that first shifts feature maps horizontally to build 4D cost volumes, then fuses the two volumes with an element-wise dot product into a similarity map $F = V^{L} \\odot V^{R}$, normalizes $F$ into attention scores via Softplus and Tanh, and finally uses those scores to weight and 3D-convolve the opposite view's volume into a reference feature that is refined and added back to the original feature. Two instantiations carry the system: HDC-FER applies this block to intermediate encoder and decoder features in both the motion and context compression paths, and HDC-EM applies a simplified version to channel-wise slices of quantized latents inside an alternating left-right autoregressive entropy model. The explicit shift supplies the geometric prior that makes optimization easy, while the attention score supplies adaptive weighting that lets matching extend beyond the local shift window.","core_discovery":"The paper's central discovery is that one disparity compensation block can serve both cross-view feature extraction and reconstruction and cross-view entropy modeling, and that the block works best when it keeps the explicit shift as a prior rather than replacing it with pure attention. HDC forms disparity volumes $V^{L}$ and $V^{R}$ by horizontally shifting each view's features over a disparity range up to $D=192$, takes their element-wise product to obtain a similarity map, normalizes it with Softplus and Tanh to obtain attention scores, and performs a weighted soft-warp with a 3D convolution to align features across views. In the entropy model, the same idea is applied to channel-wise slices of quantized latents, with left and right views encoded in alternating order so that each new slice is predicted from all previously coded slices of both views. The paper reports that HDC-FER alone lowers BD-rate by 6.03% and HDC-EM alone by 26.59% on the driving benchmark, with the combination reaching 32.17%, indicating that most of the gain comes from better entropy modeling rather than from feature alignment alone.","pith_inferences":["If the protocol-mismatch concern is put aside, the ablation predicts that transplanting the cross-view entropy model into other neural stereo codecs could transfer most of the reported gain, which is a directly testable extension.","The framework assumes rectified stereo pairs because the explicit prior is a pure horizontal shift; extending HDC to non-rectified or multi-camera arrays would require replacing the fixed shift with a learned displacement or depth-based warp that the attention score could then refine.","A head-to-head re-run using the same I-frame codec, cropping, and rate-distortion operating points would settle whether the 26.81% advantage over the previous neural method is intrinsic to HDC or partly caused by evaluation differences."],"forward_implications":["If the reported bitrate reductions hold, autonomous-driving and VR pipelines could transmit the same stereo quality at roughly half the bitrate of MV-HEVC, or higher quality at the same bitrate.","The alternating left-right sliced entropy coding means the decoder can condition on every previously decoded slice from both views, a pattern other stereo codecs could adopt independently of the alignment module.","The ablation places the larger gain in entropy modeling (26.59% alone) rather than feature alignment (6.03% alone), suggesting cross-view entropy priors are the highest-value target for further stereo compression research.","The system adds only a moderate amount of computation to its lightweight single-view backbone, so the approach stays in a practical complexity range for real-time or embedded use."],"supporting_citations":[{"why":"Provides the previous neural stereo method whose published BD-rate is the main comparison for the claimed 26.81% reduction.","marker":"[3]"},{"why":"Provides the earlier neural stereo codec used as a second comparison baseline.","marker":"[4]"},{"why":"Supplies the single-view video compression backbone that the framework extends with stereo modules.","marker":"[50]"},{"why":"Supplies the stereo attention interaction strategy against which HDC is compared in the cross-view interaction ablation.","marker":"[13]"},{"why":"Supplies the mutual attention strategy against which HDC is compared in the cross-view interaction ablation.","marker":"[67]"},{"why":"Supplies the learned image compression model used to code the initial frames in both views.","marker":"[58]"},{"why":"Defines the BD-rate metric used to compute all reported average bitrate savings.","marker":"[61]"},{"why":"Defines the MV-HEVC anchor codec against which every BD-rate in the main table is measured.","marker":"[2]"},{"why":"Supplies the general-scene multi-view sequences on which the 15.07% saving over MV-HEVC is measured.","marker":"[21]"}],"fun_headline_variants":["Entropy model does the heavy lifting in stereo video codec","Hybrid disparity compensation wins 54.6% bitrate cut","One disparity block: feature warp plus entropy coding","Stereo video compression: entropy model drives 32% gain","Explicit shift plus attention trims stereo video bits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison relies on numbers reported in other papers rather than re-running those methods under identical conditions, so if the test setups differ even slightly, part of the claimed bitrate advantage could come from that difference.","fun_headline_variants_meta":{"raw":{"variants":["Entropy model does the heavy lifting in stereo video codec","Hybrid disparity compensation wins 54.6% bitrate cut","One disparity block: feature warp plus entropy coding","Stereo video compression: entropy model drives 32% gain","Explicit shift plus attention trims stereo video bits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00027,"raw_usage":{"total_tokens":1657,"prompt_tokens":1011,"completion_tokens":646,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":564}},"tokens_in":627,"tokens_out":646,"duration_ms":6618,"temperature":1.0,"reasoning_tokens":564,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:30:36.364331+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the previous neural stereo methods and this one on the same preprocessed frames with the same I-frame compression code and the same rate-distortion operating points, then recompute BD-rate; if the gap over the previous method falls well below 26.81% or reverses, the paper's central comparative claim is not supported.","supporting_citations":[{"cited_title":"Lsvc: A learning- based stereo video compression framework,","cited_arxiv_id":null,"evidence_quote":"Provides the earlier neural stereo codec used as a second comparison baseline."},{"cited_title":"Sasic: Stereo image compression with latent shifts and stereo attention,","cited_arxiv_id":null,"evidence_quote":"Supplies the stereo attention interaction strategy against which HDC is compared in the cross-view interaction ablation."},{"cited_title":"Bidirectional stereo image compression with cross-dimensional entropy model,","cited_arxiv_id":null,"evidence_quote":"Supplies the mutual attention strategy against which HDC is compared in the cross-view interaction ablation."},{"cited_title":"Calculation of average psnr differences between rd- curves,","cited_arxiv_id":null,"evidence_quote":"Defines the BD-rate metric used to compute all reported average bitrate savings."},{"cited_title":"Overview of the multiview and 3d extensions of high efficiency video coding,","cited_arxiv_id":null,"evidence_quote":"Defines the MV-HEVC anchor codec against which every BD-rate in the main table is measured."},{"cited_title":"Nagoya university sequences,","cited_arxiv_id":null,"evidence_quote":"Supplies the general-scene multi-view sequences on which the 15.07% saving over MV-HEVC is measured."}],"review_version":1}