{"id":"9db26529-1a20-46ea-952b-8819772b9232","arxiv_id":"2508.09543","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"IVF-AStereo fuses correlation and concatenation cost volumes in two phases to keep stereo disparity accurate when camera views differ in resolution or color.","lead":"This paper presents a stereo-matching algorithm that handles pairs of camera images with different resolution or color quality, a common situation in tele-wide multi-camera systems. It works by combining two complementary matching volumes in sequence, and the authors show it stays accurate even when one view is heavily degraded.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The method's core mitigation (degrading the left view to match the right) assumes a known degradation operator, which real tele-wide/RCCB/unknown-degradation systems do not provide; the paper's experiments only validate known synthetic degradations, leaving the central claim overbroad.","rationale":"The reader's weakest_assumption identifies exactly the concern I consider most load-bearing: the deliberate degradation of the high-quality view presupposes knowledge of the degradation operator corrupting the other view. My independent reading of the manuscript confirms this is not a peripheral detail but the foundation of the correlation-volume mitigation. The paper's contribution explicitly avoids image restoration, yet it replaces it with a degradation synthesis step that requires the same operator knowledge a restoration method would need. The experiments never deviate from exact known degradation: every asymmetric input is generated by the same bilinear downsampling and grayscale conversion used to create I'_L, so the correlation volume is always computed from perfectly matched synthetic views. Real tele-wide systems, RCCB/RCCC sensors, or even different camera models introduce degradation that is not known a priori and not trivially synthesizable. Without an estimation step or an evaluation under operator mismatch, the central claim is scope-limited. I see no internal inconsistency in the proposed network under the tested conditions; the ablations, especially Table IV and V, provide reasonable support for the design choices. The missing comparison with prior asymmetric-stereo methods is also a weakness, but the known-degradation concern is more fundamental because it challenges the general applicability of the method, not just its relative performance. I therefore agree with the reader and would keep the verdict CONDITIONAL, requiring the authors to either restrict their claims to known-degradation scenarios or demonstrate robustness to unknown/mismatched degradations. My proposed concrete test—mismatched degradation evaluation—would directly settle whether the concern lands. If the method remains effective under mismatched operators, the reader's concern would be partially alleviated; if not, the paper's framing must be revised.","tokens_in":10229,"tokens_out":5128,"duration_ms":64417,"concrete_test":"Train IVF-AStereo on Scene Flow with right views degraded by 4x bilinear downsampling plus grayscale conversion. Evaluate the trained model on right views degraded by a different operator that preserves the same resolution, e.g., 4x nearest-neighbor downsampling or Gaussian blur with σ matched to the 4x downsampling's cutoff. Compare EPE against (a) the same model evaluated with the matching 4x bilinear degradation, (b) a no-degradation baseline where the correlation volume uses upsampled IR (as in Table IV row 1), and (c) MC-Stereo+A under the same mismatched degradation. If the full method's EPE under mismatched degradation increases substantially relative to the matching-degradation case and approaches or exceeds the no-degradation baseline, the concern that the method depends on knowing the exact degradation operator is confirmed; if it remains close, the method may be robust to oper","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-C.1 states that 'IL is deliberately degraded into I_L according to the degradation of IR' and that 'degradation is simpler than restoration and introduces less bias.' This is the linchpin of the correlation-volume mitigation: the correlation volume in Eq. (2) is computed from features of I'_L and IR, which are symmetric only if the degradation operator applied to IL exactly matches the unknown degradation that produced IR. In all experiments, the degradation is synthetic and fully known (bilinear downsampling plus grayscale conversion), so I'_L can be synthesized perfectly. For the tele-wide, RCCB/RCCC, or any real asymmetric system cited in the Introduction, the actual degradation operator is not given; it may involve spatially varying blur, different spectral responses, or other unmodeled effects. If I'_L is not actually symmetric with IR, the correlation volume is computed from mismatched views, and the claimed 'mitigation' fails. The paper provides no method to estimate the degradation from the input pair, nor any evaluation under mismatch between the assumed and actual degradation. Consequently, the central claim that 'the key solution to asymmetric stereo matching lies in the comprehensive utilization of distorted cost volumes' (Section I) is only supported for the favorable case where the degradation is exactly known. The broad framing in the Introduction and abstract overstates the applicability to real asymmetric systems, and the two-phase fusion argument in Section III-D rests on the effectiveness of this degradation step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses stereo matching under asymmetric visual properties (e.g., one view downsampled or grayscale-converted), arguing that the primary cause of performance degradation is distortion of the cost volume. It proposes IVF-AStereo, a two-phase iterative network that combines a correlation volume (computed from a deliberately degraded left view, to restore symmetry) and a concatenation volume (computed from the original high-quality left view and upsampled right view). Two GRU branches are used: the first refines the correlation volume with multi-peak lookup, and the second optimizes disparity, first without and then with the refined correlation volume. Experiments on Scene Flow, KITTI, and Middlebury under synthetic right-view degradation report clear gains over existing methods at 4x-8x downsampling, plus ablations of the input alignment, multi-peak group count, and fusion scheme.","tokens_in":10549,"tokens_out":3000,"duration_ms":31017,"significance":"If the claims hold, the paper provides a useful and simple architectural insight: correlation and concatenation cost volumes degrade differently under asymmetric inputs, and deliberately degrading the high-quality view can mitigate correlation-volume distortion. The proposed method achieves strong results under large synthetic resolution/color asymmetry, and the zero-shot Middlebury results are impressive. The work also isolates a concrete and falsifiable design principle (degradation for correlation volume, preservation for concatenation volume) that could transfer to other iterative stereo architectures. However, the significance is currently limited by the reliance on known synthetic degradations and by the absence of comparisons with prior asymmetric-stereo methods; no code is mentioned, so reproducibility is not verifiable from the manuscript.","major_comments":[{"comment":"The empirical study motivating the central claim is under-specified. It is unclear how many KITTI images were used, how the matching costs c_ideal and c_asym were aggregated, and how the percentages in Figure 2(d) were computed and averaged across pixels/images. No confidence intervals or error bars are given. Since the entire architecture is justified by the claim that 'correlation volume consistently favors case (c) whereas concatenation volume exhibits progressively diminished performance,' the reader needs enough detail to evaluate whether this is a robust trend or an anecdotal illustration. Please specify the protocol precisely and ideally report per-image variability.","section":"Section III-A, Figure 2"},{"comment":"The proposed correlation-volume mitigation assumes the degradation operator is known exactly: 'IL is deliberately degraded into I_L according to the degradation of IR.' In the experiments the degradation is synthetic (bilinear downsampling plus grayscale conversion), so I'_L can be made perfectly symmetric with IR. For the real tele-wide, RCCB/RCCC, or unknown-degradation systems cited in the Introduction, this operator is not given, and the paper provides no degradation-estimation method and no robustness evaluation under mismatch between assumed and actual degradation. The manuscript's broad framing, that the 'key solution to asymmetric stereo matching' is comprehensive utilization of distorted cost volumes, is therefore only supported for the favorable known-degradation case. Either restrict the claim to this setting or add an experimental protocol with unknown/mismatched degradation.","section":"Section III-C.1, Eq. (2)"},{"comment":"No prior asymmetric-stereo methods are compared. The Related Work discusses Chen et al. (CVPR 2022), Liu et al. (CVPR 2020), and Song et al. (CVPR 2023) as directly addressing asymmetric stereo, but none appear in the quantitative comparisons on Scene Flow or Middlebury. Without these baselines, the claim that the method is 'superior' and 'state-of-the-art' in asymmetric scenarios is not established against the most relevant prior work. Please include comparisons under the same degradation protocol, even if those methods require their own training or adaptation.","section":"Experiments, Tables I and III"},{"comment":"The gain from the proposed two-phase fusion over the 'all iterations' fusion (Scheme 4 vs. Scheme 5) is marginal: EPE is 0.53 in both cases and the >3px metric improves only from 2.62% to 2.59%. The text claims the two-phase scheme 'further enhances the >3px' and 'refining initially poor correlation volume,' but this difference is likely within noise. Report multiple runs with variance, or temper the claim. The same concern applies to Table V, where K=3 differs from K=2 only by 0.01 EPE and 0.03 percentage points in >3px.","section":"Table VI"}],"minor_comments":[{"comment":"The KITTI comparison is not apples-to-apples: 'Ours(A)' is evaluated on asymmetric inputs while the reference methods are evaluated on symmetric inputs. The authors acknowledge this, but the text should state more explicitly that this comparison cannot establish superiority over these methods under the same conditions.","section":"Table II and Section IV-B"},{"comment":"The notation DD(c_asym) = D_KL(softmax(c_asym) || softmax(c_ideal)) should be defined more carefully: are c_asym and c_ideal the same cost-volume location after alignment/disparity warping? How is the KL divergence computed over the disparity dimension only? Please clarify.","section":"Section III-A, Eq. (1)"},{"comment":"Minor typos and formatting issues: 'Boldand underlined' in Table I footnote, 'V olume' in references, and inconsistent use of I_L versus I'_L in the text around Section III-C.1.","section":"General"},{"comment":"The diagram is dense and the flow of C'_R in Phase 1 vs Phase 2 is hard to follow. A clearer annotation showing which tensors are zeroed in Phase 1 would help the reader map Eq. (10) to the figure.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's central contribution is a sensible engineering combination of existing components (MC-Stereo's multi-peak lookup and IGEV-Stereo's concatenation volume) with a two-phase fusion scheme. The main scientific novelty is the empirical observation about differential distortion of the two volume types, but that observation is currently presented with insufficient experimental detail. The lack of comparison with prior asymmetric-stereo methods and the strong dependence on a known degradation operator are the biggest risks. The paper could become acceptable if these are addressed, but the current scope of the claims exceeds what the experiments support."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth a serious look if you work on asymmetric stereo. The central idea is that correlation and concatenation volumes carry complementary distortions when the two views differ in quality, and that deliberately degrading the high-quality view to match the other one makes the correlation volume more reliable. The zero-shot Middlebury numbers at 4x–8x downsampling are the strongest evidence: EPE 3.34 at 8x versus 12.57 for IGEV-Stereo and 8.58 for IGEV++. Applying the same degradation trick to MC-Stereo (MC-Stereo +A) also helps, which supports the diagnosis. The two-phase fusion schedule, where the concatenation volume drives early updates before the refined correlation volume is mixed in, is a sensible design, and the ablation points in the right direction.\n\nThe soft spots are real but not disqualifying. The paper never compares against the prior asymmetric-stereo methods it cites [4,18,22], so we don't know whether existing restoration or self-similarity approaches already achieve similar accuracy. The bigger issue is the known-degradation assumption: the method degrades the left view according to the right view's degradation, but there is no estimation procedure and no experiments with mismatch between assumed and actual degradation. For tele-wide or RCCB systems the degradation is not given as ground truth; it has to be calibrated or estimated. The stress-test note is fair, though I'd frame it as a scoping limit rather than a deal-breaker, since the paper explicitly restricts experiments to resolution and color asymmetry and says other distortions are future work. Still, the abstract and intro claim more generality than the experiments back.\n\nThe diagnostic study in Section III-A is also under-specified, and the ablation gains for both K=3 and the two-phase scheme are small (e.g., >3px from 2.62 to 2.59 for the fusion scheme). No code is released. These are all fixable, but they leave the verification incomplete.\n\nMy bottom line: the architecture is coherent, the empirical gains on the tested setting are clear, and the work advances the subfield. It needs a substantial revision that adds asymmetric-stereo baseline comparisons, tests with mismatched degradation, and more detail on the cost-distortion analysis. I'd send it to review with the expectation of major revision rather than desk rejection.","headline":"A coherent and promising method for asymmetric stereo under known resolution/color degradation, with striking zero-shot gains, but the known-degradation assumption and missing comparisons keep it from being the last word.","tokens_in":11030,"tokens_out":3558,"would_cite":true,"duration_ms":38240,"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":"Asymmetric stereo matching fails mainly because visual asymmetry distorts the cost volume; fusing a deliberately symmetrized correlation volume with a robust concatenation volume in two phases recovers accuracy.","keywords":["asymmetric stereo matching","cost volume distortion","correlation volume","concatenation volume","iterative disparity optimization","multi-peak lookup","tele-wide cameras","resolution asymmetry"],"falsifier":"Take a pair where the right view is downsampled 4x and converted to grayscale, as in the paper. Compute the Eq. (1) KL divergence of the correlation cost before and after deliberately degrading the left view with the same operator. If the deliberate degradation does not reduce the distortion, or if a variant that fuses GRU 1 from the first iteration matches or beats the two-phase gate on EPE, the central claim fails.","tokens_in":10112,"feed_emoji":"📏","tokens_out":9604,"duration_ms":84788,"temperature":0.7,"pith_summary":"This paper argues that when the two views of a stereo pair differ in resolution or color, the main failure point is the cost volume, and that the two standard cost-volume constructions fail in opposite ways. It proposes IVF-AStereo, an end-to-end network that deliberately degrades the high-quality view to restore symmetry for a correlation volume, keeps a concatenation volume that is more tolerant of asymmetry, and fuses the two through separate GRU branches in a two-phase schedule. The paper reports that this design stays accurate under strong resolution and color asymmetry and generalizes to unseen data without a separate image-restoration step. If the claim holds, asymmetric multi-camera rigs such as tele-wide and RCCB/RCCC systems could be used for depth estimation directly, without adding a symmetric stereo pair.","feed_headline":"Fusing two cost volumes fixes depth under asymmetric views","feed_subtitle":"Deliberately degrading the sharp view and fusing two cost volumes cuts depth error under resolution and color mismatch.","key_machinery":"The central object is the pair of cost volumes plus the two-phase gate. The correlation volume $C_{\\mathrm{corr}}(d,x,y)=\\langle f^{\\mathrm{cor}}_l(x,y), f^{\\mathrm{cor}}_r(x-d,y)\\rangle$ is computed from features of the deliberately degraded left view and the right view, making the pair symmetric; the concatenation volume $C_{\\mathrm{cat}}$ uses group-wise concatenation from the original high-quality left and upsampled right and is regularized by a 3D network. Two ConvGRUs (convolutional gated recurrent units) carry the iterative update: GRU 1 refines $C^{i-1}_R$ using multi-peak lookup from the top-$K$ correlation peaks, and GRU 2 updates disparity using a single-peak local cost from $C_{\\","core_discovery":"The paper claims that asymmetric visual quality breaks stereo matching chiefly at the cost volume, and that the two standard cost volumes break in complementary ways. Measuring distribution distortion with KL divergence between the softmax of asymmetric matching costs and the softmax of ideal symmetric costs, the authors find that a correlation volume's costs drift and become multi-peaked under asymmetry, while a concatenation volume stays comparatively stable but loses fine-detail information through 3D downsampling. On this basis IVF-AStereo degrades the high-quality left view to match the degraded right view for correlation, builds the concatenation volume from the original high-quality l","pith_inferences":["Not pursued in the paper: the KL-distortion diagnostic could be reused for blur, noise, or cross-spectral mismatch, predicting that any mitigation lowering correlation-cost distortion will lower disparity error.","Not pursued in the paper: for unknown real-world degradation, the deliberate degradation step could be learned end-to-end before feature extraction instead of being hand-specified.","Not pursued in the paper: the K=3 multi-peak sweet spot suggests confidence-weighted peak selection might extend the benefit to more peaks."],"forward_implications":["Depth estimation on tele-wide, RCCB, and RCCC camera rigs becomes possible with a single matching network, with no restoration stage before stereo.","Correlation-only stereo systems can be made more asymmetric-resistant by degrading the high-quality view into symmetry, as the paper's +A variant demonstrates.","Concatenation-based iterative networks gain fine details by fusing a correlation volume, but only after the correlation branch has been refined first; injecting it from the first iteration is worse.","The two-phase gate is a reusable recipe for merging a high-variance, detail-rich signal with a lower-variance, coarse signal in iterative disparity refinement."],"supporting_citations":[{"why":"Supplies the multi-peak lookup and the GRU-1 update rule used to refine the degraded correlation volume.","marker":"[8]"},{"why":"Supplies the concatenation-volume construction, 3D regularization, and the iterative disparity optimization baseline the method extends.","marker":"[28]"},{"why":"Introduces the recurrent ConvGRU disparity-update paradigm and context features used by both branches.","marker":"[17]"},{"why":"Provides the group-wise volume construction used to build the concatenation volume.","marker":"[13]"},{"why":"Defines the resolution-asymmetric stereo problem and gives the degradation-agnostic baseline the paper contrasts with its cost-volume approach.","marker":"[4]"},{"why":"Provides the Scene Flow training data and benchmark used for the main experiments.","marker":"[19]"},{"why":"Provides the Middlebury 2014 zero-shot generalization benchmark.","marker":"[21]"},{"why":"Provide the KITTI test benchmarks used for the representative comparison under asymmetric input.","marker":"[9, 20]"}],"fun_headline_variants":["Fusing two cost volumes fixes depth under asymmetric views","Two cost volumes combine to handle asymmetric stereo","IVF-AStereo fuses two volume types for asymmetric depth","Complementary cost volumes beat asymmetric stereo","Cost volume fusion tackles asymmetric camera setups"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The method assumes the degradation of the lower-quality view is known well enough to create a matching degraded left view; if that synthesized pair is not truly symmetric, the correlation-volume mitigation and the two-phase fusion argument no longer hold.","fun_headline_variants_meta":{"raw":{"variants":["Fusing two cost volumes fixes depth under asymmetric views","Two cost volumes combine to handle asymmetric stereo","IVF-AStereo fuses two volume types for asymmetric depth","Complementary cost volumes beat asymmetric stereo","Cost volume fusion tackles asymmetric camera setups"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000394,"raw_usage":{"total_tokens":1872,"prompt_tokens":681,"completion_tokens":1191,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":425,"completion_tokens_details":{"reasoning_tokens":1119}},"tokens_in":425,"tokens_out":1191,"duration_ms":9345,"temperature":1.0,"reasoning_tokens":1119,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:59:04.657243+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a pair where the right view is downsampled 4x and converted to grayscale, as in the paper. Compute the Eq. (1) KL divergence of the correlation cost before and after deliberately degrading the left view with the same operator. If the deliberate degradation does not reduce the distortion, or if a variant that fuses GRU 1 from the first iteration matches or beats the two-phase gate on EPE, the central claim fails.","supporting_citations":[{"cited_title":"MC-Stereo: Multi-Peak Lookup and Cascade Search Range for Stereo Matching","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-peak lookup and the GRU-1 update rule used to refine the degraded correlation volume."},{"cited_title":"Iterative Geometry Encoding V olume for Stereo Matching","cited_arxiv_id":null,"evidence_quote":"Supplies the concatenation-volume construction, 3D regularization, and the iterative disparity optimization baseline the method extends."},{"cited_title":"RAFT- Stereo: Multilevel Recurrent Field Transforms for Stereo Matching","cited_arxiv_id":null,"evidence_quote":"Introduces the recurrent ConvGRU disparity-update paradigm and context features used by both branches."},{"cited_title":"Group-wise Correlation Stereo Network","cited_arxiv_id":null,"evidence_quote":"Provides the group-wise volume construction used to build the concatenation volume."},{"cited_title":"Degradation-Agnostic Correspon- dence From Resolution-Asymmetric Stereo","cited_arxiv_id":null,"evidence_quote":"Defines the resolution-asymmetric stereo problem and gives the degradation-agnostic baseline the paper contrasts with its cost-volume approach."},{"cited_title":"A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation","cited_arxiv_id":null,"evidence_quote":"Provides the Scene Flow training data and benchmark used for the main experiments."},{"cited_title":"High-Resolution Stereo Datasets with Subpixel-Accurate Ground Truth","cited_arxiv_id":null,"evidence_quote":"Provides the Middlebury 2014 zero-shot generalization benchmark."}],"review_version":1}