{"id":"d7862331-b76b-4819-a3be-ad5d37163b13","arxiv_id":"2508.01275","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"An unsupervised stereo matching method uses predicted relative depth to filter confident disparity estimates and guide training losses, claiming state-of-the-art KITTI results, but the key confidence equation is internally inconsistent.","lead":"This paper proposes a new way to train stereo depth algorithms without labeled data, using a neural network's own depth ranking predictions as a guide. The authors report top accuracy on the KITTI benchmarks among unsupervised methods, but the paper's central formula for computing disparity confidence is self-contradictory as written.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (3) is self-contradictory: the two Θ factors require |ΔD|≤1 and |ΔD|≥σ>1 simultaneously, so F_VC≡0 and DDCV outputs an all-zero confidence map, disabling the LDR mechanism.","rationale":"The reader's weakest assumption exactly matches the most load-bearing flaw. Eq. (3) is the first concrete mechanism by which relative depth priors enter the framework, and it is self-contradictory: no pair of scalar values can satisfy both |ΔD|≤1 and |ΔD|≥σ>1, nor both |ΔeD|≥γσ and |ΔeD|≤γ with σ>1. Therefore F_VC is identically zero and DDCV collapses to an all-zero confidence map, which removes the basis for the LDR loss's top-k reliable-correspondence selection. No amount of retuning γ, σ, or λ_i can repair this as written; the central contribution is not reproducible from the manuscript. The paper has useful empirical pieces: an ablation table, benchmark numbers, and a plausible plug-and-play confidence idea, but those cannot be validated until the equation is corrected and the code confirms the intended behavior. Because the submitted algorithm as written cannot produce the claimed results, the reader's REJECT verdict stands without adjustment.","tokens_in":17069,"tokens_out":5254,"duration_ms":56246,"concrete_test":"Recompute the truth conditions of Eq. (3) for any σ>1 and any positive γ: the two Θ factors impose contradictory bounds on |ΔD|, proving F_VC≡0. Then execute the authors' released code on one KITTI training stereo pair and print the maximum and histogram of C. If max(C)>0, the implemented formula differs from Eq. (3) and the manuscript must be corrected. If max(C)=0, the DDCV component cannot support the LDR loss as described, and the reported results would require a different implementation than the one specified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that DDCV (Eqs. (1)–(6)) produces a meaningful confidence map that gates the LDR correspondences. This fails as written. Eq. (3) defines F_VC as the product of two Θ(·) factors. The left factor is 1 only if |ΔeD| ≥ γσ and |ΔD| ≤ 1. The right factor is 1 only if |ΔeD| ≤ γ and |ΔD| ≥ σ. Since σ ∈ (1,+∞), the conditions on |ΔD| are disjoint and the conditions on |ΔeD| are also disjoint whenever σ > 1; hence F_VC(p,q)=0 for every neighboring pair. Consequently v_pq = FRC(p,q) × FVC(p,q) = 0, so C(p)=0 for every pixel by Eq. (6). The confidence map used by the proposed LDR loss is therefore all zeros, and the top-k selection in Eq. (10) chooses among equally low-confidence points; the claimed quasi-dense correspondences with reliable disparities and the accuracy gains in Table I and Table IV cannot be produced by the mechanism as described. The implementation paragraph also says “α = 2 in (3)” although Eq. (3) contains σ, which hints at a transcription slip, but the submitted formula is not merely ambiguous; it is unsatisfiable. This concern is independent of the benchmark comparison and of hyper-parameter choices.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an unsupervised stereo matching framework, Un-ViTAStereo, built on the authors' earlier ViTAStereo network. The three contributions are: (i) DDCV, a plug-and-play disparity confidence estimation algorithm that checks local coherence consistency between disparities and relative depths; (ii) a local depth ranking (LDR) loss that builds quasi-dense correspondences using only high-confidence disparities to transfer relative-depth ranking knowledge; and (iii) a dual disparity smoothness (DDS) loss that aligns disparity gradients with relative-depth gradients. Experiments on SceneFlow, KITTI 2012/2015, Middlebury, and ETH3D report state-of-the-art results among unsupervised stereo matching methods on the KITTI benchmarks, with extensive ablations and a runtime of 0.9 s for the full pipeline.","tokens_in":17390,"tokens_out":9788,"duration_ms":113418,"significance":"If the described mechanisms are correct, the work is a meaningful contribution to unsupervised stereo matching. The DDCV confidence estimator is novel in using relative depth priors, is parameter-free and real-time, and the LDR loss addresses a known weakness of prior knowledge-transfer methods by filtering noisy correspondences. The paper is also strong in scope: it provides public code and supplementary material, evaluates on multiple benchmarks, and includes careful ablations. However, the central formula for DDCV, Eq. (3), does not implement the behavior described in the text, and the reported performance gains are therefore not accounted for by the manuscript as written. This is a load-bearing issue that must be resolved before the claims can be accepted.","major_comments":[{"comment":"The claim that F_VC is identically zero is not correct, because Θ(0)=1, so each factor in Eq. (3) evaluates to 1 whenever its inner condition is not met. However, the formula as written implements the opposite of the stated intention. For a pair with stable relative depth (|ΔeD| ≤ γ), the first factor is 1 but the second factor is 1 only if |ΔD| ≥ σ, i.e., it votes positively only for large disparity variations. For a pair at a depth discontinuity (|ΔeD| ≥ γσ), the second factor is 1 but the first factor is 1 only if |ΔD| ≤ 1, i.e., it votes positively only for small disparity variations. Thus DDCV assigns high confidence to disparity pairs that are inconsistent with the relative depth prior and low confidence to consistent pairs. Since the LDR loss in Eq. (10) selects the top-k highest-confidence disparities as reference points, the mechanism described cannot produce the behavior claimed in Sections I and III. The two Θ arguments in Eq. (3) appear to be interchanged relative to the description; please correct the formula and re-verify all experiments that depend on the DDCV confidence map.","section":"III-A, Eq. (3)"},{"comment":"The implementation states \"We also set α = 2 in (3)\", but Eq. (3) contains the symbol σ, not α. This is a transcription error that makes the reported hyper-parameter setting ambiguous. Please clarify whether σ = 2, and ensure that all symbols in the manuscript are consistent between the equations, the text, and the implementation details.","section":"IV-A, implementation details"},{"comment":"The relative depth prior is generated by the Depth Anything V2 model, whose vision Transformer encoder is shared with the stereo matching network ViTAStereo, as shown in Fig. 2. The depth prior is therefore not fully independent of the features used for stereo matching; it is a frozen, pretrained representation, but the same encoder feeds both tasks. The paper should discuss the implications of this architectural sharing for the claimed \"knowledge transfer\" and the novelty of the approach. Ideally, an ablation with an alternative relative-depth source (e.g., a different monocular depth network) would demonstrate that the benefit is not an artifact of the shared representation.","section":"III-A and Fig. 2"}],"minor_comments":[{"comment":"Section II-A contains two consecutive, essentially identical paragraphs on unsupervised stereo matching; the duplication should be removed.","section":"II-A"},{"comment":"The notation Θ(p, r) in Eq. (10) is undefined. Please define it explicitly (presumably as a step function on the consistency between ΔeD and ΔD, analogous to Eq. (1)).","section":"III-B1, Eq. (10)"},{"comment":"The denominator in Eq. (10) can be zero when all correspondences in the neighborhood are ranking-consistent. Please add a small epsilon or define the loss term as zero in that case to avoid division by zero.","section":"III-B1, Eq. (10)"},{"comment":"The axes in Fig. 5 are not labeled; the horizontal axis should be labeled with the number of reference points k and the vertical axis with EPE (pixel).","section":"Fig. 5"},{"comment":"Table II compares DDCV with supervised confidence estimation methods (CCNN, LGC, SEDNet), but DDCV additionally uses a relative depth map as input. This should be stated in the caption or text so that the comparison is not misread as head-to-head on identical inputs.","section":"IV-D, Table II"}],"recommendation":"major_revision","confidential_remarks":"The reader's 'identically zero' concern about Eq. (3) is technically incorrect, but the formula has a different and equally serious flaw: it inverts the intended voting logic. This is not a minor typo; it is the core mechanism of the paper. The authors should be asked to correct Eq. (3) and, if the reported results were obtained with the corrected formula, to state so explicitly and likely re-run the ablations. The shared encoder between the stereo network and the depth prior also deserves more critical discussion. Given the paper's strong experimental scope and the plausibility of the overall idea, I support a major-revision decision rather than outright rejection, provided the authors can resolve the formula inconsistency and provide evidence that the reported gains are reproducible from the corrected specification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this paper is a decent within-subfield improvement to unsupervised stereo matching, and the reviewer concern about Eq. (3) being self-contradictory does not hold up. The stress-test misreads the outer Θ: when the inner Θ evaluates to 0, the outer Θ of 0 is 1, so the two terms in F_VC are not mutually exclusive. For relative-depth differences between γ and γσ, both factors are satisfied regardless of disparity difference. DDCV can produce a meaningful confidence map. So the central mechanism is plausible.\n\nWhat is actually new: DDCV as an explicit confidence measure that combines rank consistency and variation consistency between disparities and a relative depth map; the LDR loss that selects top-k confident reference points to build quasi-dense correspondences; and the DDS loss that respects both smoothness and discontinuities from the depth prior. The ablations on SceneFlow show each piece contributes, and the KITTI benchmark numbers are state-of-the-art among unsupervised methods. That is earned credit.\n\nWhere it is softer: The text says “α = 2 in (3)” but Eq. (3) uses σ; that is a minor typo but should be fixed. More substantively, the relative depth prior comes from Depth Anything V2, and the stereo network ViTAStereo shares that same frozen encoder. So the “prior” is not fully external; it is the same feature backbone reused. This is disclosed, but it weakens the “knowledge transfer” framing a bit—it is more like self-distillation from a shared representation. The hyper-parameters are numerous (λs, k, window sizes, σ) and mostly hand-picked, though the k ablation exists. No code was available to me to verify, but the paper promises it.\n\nThe paper is a solid engineering contribution, not a breakthrough. If you work on unsupervised stereo or confidence estimation, it is worth a careful read and probably worth citing. I would send it to peer review; the authors should be asked to fix the α/σ inconsistency, clarify the Θ conditions (maybe with a small truth table), and release code. The earlier reject verdict seems too harsh because it was based on a misreading of the equation.\n\nRecommendation: engage with it; send a serious referee.","headline":"A solid incremental unsupervised-stereo paper whose apparent fatal flaw in Eq. (3) evaporates on close reading; the real caveats are a typo and the shared-encoder prior.","tokens_in":17938,"tokens_out":3701,"would_cite":true,"duration_ms":38299,"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 claims that an unsupervised stereo matching framework, combining disparity confidence voting with two relative-depth-prior losses, reaches state-of-the-art accuracy on the KITTI Stereo benchmarks among all unsupervised methods.","keywords":["unsupervised stereo matching","disparity confidence estimation","relative depth prior","knowledge transfer","depth ranking loss","disparity smoothness loss","KITTI Stereo","vision foundation models"],"falsifier":"Compute Eq. (3) on any real disparity and depth pair with the stated hyper-parameters: the first factor demands $|\\Delta\\tilde D_{pq}|\\ge\\gamma\\sigma$ while the second demands $|\\Delta\\tilde D_{pq}|\\le\\gamma$, and with $\\sigma>1$ no value satisfies both, so $C(p)=0$ for every pixel. Replacing DDCV's output by an all-zero confidence map while keeping LDR and DDS unchanged should reproduce the same EPE on SceneFlow if the confidence map contributes nothing.","tokens_in":16873,"feed_emoji":"🎯","tokens_out":10054,"duration_ms":103075,"temperature":0.7,"pith_summary":"The paper tries to show that unsupervised stereo matching can be made substantially more accurate by filtering disparity estimates with a confidence map derived from a monocular relative-depth prior, then transferring depth-ranking knowledge only through confident quasi-dense correspondences. It proposes a plug-and-play confidence estimator, DDCV, that votes on local coherence between neighboring disparities and their relative depths, plus two losses: a local depth ranking (LDR) loss and a dual disparity smoothness (DDS) loss. On the KITTI Stereo 2015 benchmark the resulting Un-ViTAStereo model reports D1-all of 5.03%, which the paper claims is state of the art among unsupervised stereo matching methods and a 22.7% relative improvement over the next-best unsupervised method. If true, the practical payoff is that an unsupervised stereo network can exploit the same frozen vision-foundation depth priors that have driven supervised methods.","feed_headline":"Depth priors plus confidence voting lift unsupervised stereo on KITTI","feed_subtitle":"Reliable disparity estimates feed depth-ranking losses, cutting KITTI 2015 D1-all to 5.03% among unsupervised methods.","key_machinery":"The central objects are three. First, DDCV: a parameter-free voting scheme that computes, for each pixel $p$, $C(p)=\\frac{1}{m}\\sum_{q\\in\\mathcal{N}_p} v_{pq}$ with $v_{pq}=F_{RC}(p,q)F_{VC}(p,q)$; $F_{RC}$ checks whether $(\\tilde D(p)-\\tilde D(q))(D(p)-D(q))$ is nonnegative, and $F_{VC}$ is defined to require mild disparity variations in stable-depth regions and large disparity variations at depth discontinuities. Second, the LDR loss: for each pixel, the top-$k$ highest-confidence disparities in a dilated window act as anchors, and the loss penalizes a target disparity only when its metric-depth ranking contradicts the relative-depth ranking of its anchor, with weights $\\omega(\\Delta\\tilde D_{pr})$ favoring pairs at substantially different depths. Third, the DDS loss replaces image-gradient weighting in the standard smoothness loss with relative-depth-gradient weighting and adds a symmetric term that penalizes smooth disparities where the depth map jumps. The depth prior itself comes from a frozen vision Transformer (Depth Anything V2) that is shared, through ViTAStereo, with the stereo network.","core_discovery":"On its own terms, the paper's discovery is that the reliability of a disparity estimate can be judged by whether the local sign and magnitude changes in the disparity map agree with the local sign and magnitude changes in an off-the-shelf relative-depth map, and that using only high-confidence disparities as anchors for depth-ranking supervision removes the noise that plagues random-correspondence knowledge transfer. The authors package this into DDCV, whose confidence for a pixel is the average of binary votes from neighboring pixels; a positive vote requires both ranking consistency (Eq. 1) and variation consistency (Eq. 3), and Eq. 5 multiplies the two signals. The confidence map then selects top-k reference points for LDR, and DDS adds a depth-gradient-aware smoothness term that penalizes disagreement between disparity discontinuities and depth discontinuities. With these components, Un-ViTAStereo reports lower EPE and D1 than all published unsupervised methods on the KITTI Stereo benchmarks, and DDCV alone reaches AUC scores competitive with supervised confidence networks such as SEDNet.","pith_inferences":["A plausible correction of Eq. (3) that treats large-depth-change and stable-depth-change as two separate cases rather than simultaneous inequalities would make the variation-consistency vote nonzero; re-running the ablations with that corrected formula is the direct way to test whether the reported gains come from DDCV or from the two losses alone.","Because DDCV needs no labels and no training, it can be dropped into supervised stereo pipelines as a free confidence head; this would let future work decide whether its value in Table II, comparable AUC to SEDNet, translates into downstream error-detection gains.","The same rank-by-relative-depth, filter-by-confidence recipe could apply to other pixel-correspondence tasks, such as optical flow or video depth, wherever a monocular depth prior is available; the paper does not explore this.","If the KITTI results hold with a corrected confidence term, the practical divide between supervised and unsupervised stereo narrows, since the reported D1-all of 5.03% is within reach of some supervised figures on the same public benchmark; quantifying that gap is not attempted in the paper."],"forward_implications":["Any existing stereo network can adopt the framework without architecture or data changes, and at inference time only stereo image pairs are needed; the relative-depth map is required only during training.","Replacing random sparse correspondences with confidence-filtered quasi-dense correspondences should reduce the noise that previously degraded depth-ranking knowledge transfer; the SceneFlow ablation shows LDR alone lowers EPE by about 10.7% relative to the CDR loss.","The relative-depth-guided smoothness loss should improve disparity estimates at object boundaries in texture-less or repeated-pattern regions where image gradients mislead the standard edge-aware smoothness loss.","DDCV provides a label-free, real-time disparity confidence map that is competitive with supervised confidence estimators on KITTI, Middlebury and ETH3D across several disparity inputs."],"supporting_citations":[{"why":"Supplies the ViTAStereo architecture and the shared vision-Transformer encoder that produces both disparity and relative depth; the paper presents itself as the unsupervised version of this work.","marker":"[1]"},{"why":"Defines the CDR confident-depth-ranking loss and the random sparse-correspondence strategy that LDR is designed to replace.","marker":"[2]"},{"why":"Provides the frozen monocular relative-depth prior, Depth Anything V2, used by DDCV, LDR, and DDS.","marker":"[47]"},{"why":"SEDNet is the strongest supervised confidence baseline and also one of the disparity-map input sources used to evaluate DDCV.","marker":"[42]"},{"why":"CCNN is a supervised CNN confidence baseline that DDCV is compared against.","marker":"[39]"},{"why":"OTB is an unsupervised confidence baseline that DDCV is compared against.","marker":"[38]"},{"why":"Flow2Stereo is the prior unsupervised stereo method on KITTI that Un-ViTAStereo is compared with and outperforms.","marker":"[32]"},{"why":"KITTI Stereo 2015 is the benchmark where the state-of-the-art unsupervised claim is evaluated.","marker":"[57]"},{"why":"KITTI Stereo 2012 is the benchmark where the EPE and PEP gains are reported.","marker":"[56]"},{"why":"SceneFlow is the synthetic dataset used for the ablations that isolate the contributions of LDR and DDS.","marker":"[54]"}],"fun_headline_variants":["Disparity confidence via depth agreement prunes noise","Unsupervised stereo SOTA on KITTI from depth-coherence voting","Only confident disparities anchor depth ranking for stereo","Dual depth-guided smoothness sharpens unsupervised stereo","Depth prior agreement predicts disparity reliability, cuts errors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's confidence signal depends on Eq. (3), which as written asks a single vote to require both a large and a small depth change at the same time, so the vote is zero everywhere and the confidence map carries no information; if that formula is not corrected, the reported gains cannot come from DDCV.","fun_headline_variants_meta":{"raw":{"variants":["Disparity confidence via depth agreement prunes noise","Unsupervised stereo SOTA on KITTI from depth-coherence voting","Only confident disparities anchor depth ranking for stereo","Dual depth-guided smoothness sharpens unsupervised stereo","Depth prior agreement predicts disparity reliability, cuts errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1448,"prompt_tokens":975,"completion_tokens":473,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":395}},"tokens_in":591,"tokens_out":473,"duration_ms":6149,"temperature":1.0,"reasoning_tokens":395,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:42:31.322562+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute Eq. (3) on any real disparity and depth pair with the stated hyper-parameters: the first factor demands $|\\Delta\\tilde D_{pq}|\\ge\\gamma\\sigma$ while the second demands $|\\Delta\\tilde D_{pq}|\\le\\gamma$, and with $\\sigma>1$ no value satisfies both, so $C(p)=0$ for every pixel. Replacing DDCV's output by an all-zero confidence map while keeping LDR and DDS unchanged should reproduce the same EPE on SceneFlow if the confidence map contributes nothing.","supporting_citations":[],"review_version":1}