{"id":"144f4b4f-e434-41df-aa3a-e274af85fcb4","arxiv_id":"2501.13387","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A domain-adaptation method transfers image quality knowledge to point clouds, with distortion-weighted alignment and contrastive quality preservation, to predict point cloud quality without point cloud labels.","lead":"This paper presents a method to rate the visual quality of 3D point clouds without using any point cloud quality labels, by borrowing knowledge from image quality databases. It adapts features from distorted images to distorted point clouds, weighting common distortion types and preserving quality-sensitive information during alignment.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's weight matrix C is an outer product of marginals, not Eq. 14's joint distribution; as written it makes the QP objective identically zero and cannot produce the importance weights behind the claimed Table IV gains.","rationale":"The reader correctly identified the distortion-distribution estimation and importance-weight computation as the weakest link in the paper. I partially agree, but the sharper problem is internal: even with a perfect classifier H, Algorithm 1 as written cannot compute meaningful weights. Because C is defined as the outer product of the target-prediction histogram and the source-label histogram, the QP objective in Eq. 16 is zero for every feasible w_y, so the weights are arbitrary. This is not a disagreement with an external consensus; it is a formal inconsistency inside the proposed method. It matters because the DWCE loss in Eq. 11 is the paper's main departure from IT-PCQA, and the ablation explicitly credits this weighting with a measurable improvement. I do not move the verdict to REJECT: the experimental tables could still be produced by a corrected implementation that uses a true source confusion matrix, and other components such as the contrastive disentanglement and adversarial alignment may contribute to the reported performance. But the paper as written is not reproducible, and the authors need to release code, correct Algorithm 1, and report H's target-domain confusion before the central claim can be taken at face value. This reinforces, rather than changes, the reader's CONDITIONAL verdict.","tokens_in":19982,"tokens_out":10243,"duration_ms":74957,"concrete_test":"Reproduce the TID2013-to-SJTU-PCQA setting using Algorithm 1 exactly as printed: form \\hat C as the outer product of the normalized histograms, solve Eq. (16) with the stated constraint, and feed the resulting w_y into Eq. (11). If the QP solution is arbitrary (and w_y=1 is returned), the 'w/ wy' row of Table IV cannot be produced by the published algorithm. Then replace \\hat C with the correct source confusion matrix C_{ij}=P_S(\\hat y_d=i, y_d=j), computed by running H on source images, and recompute the same row. Also report H's per-class accuracy on SJTU point cloud projections. If the +0.045 SROCC gain appears only with the corrected C, the paper's central mechanism depends on an unstated revision rather than on Algorithm 1 as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the estimation of the distribution weights w_y = D_T(y_d)/D_S(y_d) used in the DWCE loss (Eq. 11). The paper's own specification of this step is internally inconsistent. Eq. 14 defines C as the source joint distribution C = D_S(\\hat y_d, y_d), and Eq. 16 solves w_y from \\hat\\mu = \\hat C w_y. Algorithm 1, however, computes \\hat C = (c_{y_d}')^T (y_s_d'), the outer product of the normalized target-prediction histogram and the normalized source-label histogram. This matrix is rank one. Moreover, because the QP constraint is w_y^T D_S(y_d) = 1, for every feasible w_y we have \\hat C w_y = c_{y_d}' = \\hat\\mu, so the objective in Eq. 16 is identically zero and the returned w_y is arbitrary. The trivial feasible point w_y = 1 always satisfies the constraint, meaning Algorithm 1 as written cannot distinguish 'w/ wy' from 'w/o wy', yet Table IV attributes a gain of +0.067 PLCC / +0.045 SROCC to this weight. No source-domain predictions are even passed into Algorithm 1, so Eq. 14's joint distribution cannot be computed from the stated inputs. Together with the absence of code and the absence of any report of H's target-domain accuracy, the published text does not establish that the reported gains come from the distortion-guided weighting mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DWIT-PCQA, a no-reference point cloud quality assessment method that transfers image quality priors to point clouds without using point cloud MOS labels. Point clouds are projected into multi-view images, a shared image encoder extracts features, and an adversarial domain adaptation framework aligns source (image) and target (point cloud) feature distributions. To handle distortion distribution mismatch, the paper introduces a distortion-weighted cross-entropy loss whose weights are estimated from a source-trained distortion classifier, together with a contrastive quality-aware feature disentanglement loss. Experiments on TID2013/KADID-10k to SJTU-PCQA/WPC report large gains over the prior IT-PCQA method and competitive performance with supervised PCQA methods.","tokens_in":1769,"tokens_out":2248,"duration_ms":85679,"significance":"If the claims hold, this is a practically valuable result: it would show that point cloud quality can be predicted for new scenes using only image MOS labels and unlabeled target point clouds, addressing the scarcity of PCQA annotations. The paper includes a reasonable empirical protocol with multiple source and target datasets, comparisons against FR and NR PCQA methods, and an ablation study. I credit the authors for not using target MOS labels in training and for reporting gains over the most relevant prior art. However, the central mechanism for the reported gains — the distortion-based distribution weighting — is specified inconsistently in the text, and the claimed improvement from that mechanism is not isolated in the ablation. These issues need to be resolved before the contribution can be fully assessed.","major_comments":[{"comment":"The specification of the distribution-weight computation is internally inconsistent. Eq. (14) defines C as the source joint distribution D_S(ŷ_d, y_d), which requires the classifier's predictions on the source domain. Algorithm 1, however, receives only source distortion labels and target predictions, and computes C_hat = (cy_t_d')^T (ys_d') — the outer product of the normalized target-prediction histogram and the normalized source-label histogram. This matrix is rank one and is not an estimate of the joint distribution in Eq. (14). Moreover, with the constraint w_y^T D_S(y_d) = 1 and with mu_hat = cy_t_d', every feasible w_y satisfies C_hat w_y = cy_t_d' = mu_hat, so the QP objective in Eq. (16) is identically zero over the feasible set and the returned w_y is arbitrary. As a result, the +0.067 PLCC / +0.045 SROCC gain attributed to w_y in Table IV cannot be traced to the distortion-guided weighting mechanism as written. Please correct Algorithm 1, for example by passing H's source predictions as input and forming C_hat from the source joint histogram, and confirm the ablation result with the corrected weights; releasing code would also help resolve this issue.","section":"Section IV-B (Algorithm 1, Eqs. (14)–(16))"},{"comment":"The 'decomposition' of the DA objective is not a formal derivation. Eq. (2) aligns the quality-conditioned feature distributions D_S(z|y) and D_T(z|y). The paper replaces it with Eq. (3), alignment of distortion-conditioned distributions, plus Eq. (4), which only penalizes the distance between D_S(z|y) and D_S(z|y_d) within the source domain. No inequality is provided showing that minimizing these two sub-objectives controls the target quality-conditional term D_T(z|y) in Eq. (2). In particular, Eq. (4) does not constrain the target domain conditional distribution, which is the term most relevant to target quality prediction. Please provide a bound connecting the sub-objectives to the original objective, or state explicitly that this is a heuristic surrogate.","section":"Section III-A (Eqs. (2)–(4))"},{"comment":"The ablation does not isolate the effect of the distribution weight w_y. The row 'LReg + LDWCE (w/o wy)' is compared with the row 'LReg + LDWCE (w/ wy) + LCls'; the simultaneous addition of LCls is confounded with the effect of w_y. The claim that 'the distortion-based importance-weighted alignment is necessary and effective' requires an additional condition such as LReg + LDWCE (w/o wy) + LCls. Please also report this comparison on WPC, not only on SJTU-PCQA, to support the generality of the weighting mechanism.","section":"Section V-D (Table IV)"},{"comment":"The entire weight estimation relies on the source-trained classifier H producing meaningful pseudo-labels on the target point-cloud projections, but the paper never reports H's accuracy, confusion matrix, or prediction entropy on the target domain. If H's target predictions are noisy, then the estimated D_T(y_d), and hence w_y, is essentially noise, and the gains in Table IV may not reflect distortion-guided alignment. Please report quantitative evidence about H on the target projections (at minimum, confidence/entropy statistics and stability of w_y across H initializations) and a sensitivity analysis of the final SROCC/PLCC to the quality of H.","section":"Section IV-D (Distortion Distribution Prediction)"}],"minor_comments":[{"comment":"There is a dimension mismatch in the stated network architecture: G maps features to 256 dimensions, but H and R are described with first FC layers of 2048 inputs. Please clarify whether H and R operate on the 2048-d backbone features or on G's 256-d output.","section":"Section IV-A / IV-D / IV-E"},{"comment":"The word 'constractive' should be 'contrastive'.","section":"Section III-C"},{"comment":"The table caption should explicitly state that transfer methods use unlabeled target-domain data during training, whereas the general IQA baselines do not; this is the intended unsupervised domain adaptation protocol, but the asymmetry should be flagged to keep the comparison fair.","section":"Table II caption"},{"comment":"All weighting factors in Eq. (22) are set to 1 with no sensitivity study; a brief robustness check for lambda_1 to lambda_4 would help establish that the reported results do not hinge on a particular tuning.","section":"Section V-A"},{"comment":"The t-SNE visualization in Fig. 7 is qualitative; a quantitative measure of feature-space regularity, such as a domain alignment error or a nearest-classifier accuracy, would strengthen the claim.","section":"Section V-E (Fig. 7)"}],"recommendation":"major_revision","confidential_remarks":"The core empirical claim — that large gains over IT-PCQA can be obtained without target MOS labels — is attractive, and the paper's overall evaluation is thorough. The main barrier is that the published Algorithm 1 does not compute the quantity defined in Eq. (14), and the QP as written appears to return an arbitrary feasible weight. This is fixable by a corrected algorithm and additional ablations, so I recommend major revision rather than rejection. The absence of code makes it impossible to determine whether the reported Table IV numbers were produced by the algorithm as printed or by a different procedure; releasing code would be important in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The idea is the right response to a real bottleneck: PCQA datasets are tiny, and borrowing IQA priors through domain adaptation is the obvious workaround. The two new components—distortion-weighted alignment and quality-aware contrastive disentanglement—are natural extensions of your own IT-PCQA, and the reported gains are large and consistent across four transfer settings. The comparison against supervised PCQA methods is also honest: you land within a few percent of methods that train on point cloud MOS, without ever seeing a target label. On the empirical side, I believe the results as reported.\n\nBut the paper has a load-bearing flaw in the written specification of the key weighting step. Theorem 1 and Eq. 14 define C as the joint distribution D_S(c_yd, y_d), and Eq. 16 solves for w_y from C w_y = mu. Algorithm 1, however, computes C as the outer product of the target-prediction histogram and the source-label histogram. That matrix is rank one, and given the constraint w_y^T D_S(y_d) = 1, every feasible w_y makes the QP objective identically zero. In other words, Algorithm 1 as written cannot produce anything other than the trivial w_y = 1, so the claimed benefit of the weighting mechanism in Table IV is not supported by the paper's own pseudocode. The problem is compounded by the ablation: the \"w/ wy\" row also adds the LCls branch, so the +0.067 PLCC / +0.045 SROCC gain cannot be cleanly attributed to the weights. And no source-domain predictions are even fed into Algorithm 1, so the joint distribution in Eq. 14 is not computable from the stated inputs.\n\nThe theory in Section III is also informal: decomposing Eq. (2) into (3) and (4) without bounding the omitted term is not a derivation; it's a motivation. That part I can live with if the experiments are solid, but the algorithmic inconsistency is not a presentation issue. It is the mechanism that supposedly drives the headline results.\n\nOtherwise the protocol is reasonable—no target MOS used, standard 5-fold splits, multiple databases. The citation pattern is fine; self-citing IT-PCQA is appropriate. I would send this to review, but with a major-revision request: either correct Algorithm 1, report the actual computed weights, release code, or change the claims to match what the implementation really does. Without that, the specific contribution of the distortion-guided weighting remains unverified. A reader who cares about unlabeled PCQA will want to see this fixed.","headline":"Sensible extension of IT-PCQA with large reported gains, but the paper's central weighting algorithm is internally inconsistent and the ablation is confounded.","tokens_in":20838,"tokens_out":3649,"would_cite":false,"duration_ms":701275,"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 network trained only on distorted images and their quality scores can predict the perceptual quality of unlabeled point clouds from new scenes.","keywords":["cross-media transfer","blind quality assessment","point cloud quality assessment","domain adaptation","distortion distribution weighting","contrastive learning","no-reference metric","unlabeled training"],"falsifier":"Replace the computed distortion weights with uniform weights in the distortion-weighted cross-entropy loss on the TID2013-to-SJTU-PCQA setup; if SROCC does not drop materially, the distortion-distribution weighting is not doing the work attributed to it.","tokens_in":19773,"feed_emoji":"🧊","tokens_out":6316,"duration_ms":56516,"temperature":0.7,"pith_summary":"The paper claims that a network trained only on distorted images and their mean opinion scores can predict the perceptual quality of point clouds from unseen scenes without ever seeing point cloud quality labels. Its method, DWIT-PCQA, projects point clouds into multi-view images, aligns image and point-cloud features in a shared space, and uses estimated distortion distributions to weight the alignment toward common distortions. A quality-aware contrastive branch keeps the aligned features sensitive to quality. If this holds, point cloud quality assessment no longer needs expensive subjective labeling for every new dataset.","feed_headline":"Point cloud quality scored with zero point cloud labels","feed_subtitle":"Image-trained network scores unlabeled point clouds, beating older transfer methods by up to 40 percent.","key_machinery":"The central object is a decomposition of the domain adaptation objective with distortion as an intermediate variable, turning direct alignment of quality-conditioned features into two subobjectives: one that aligns features conditioned on distortion type and one that maintains consistency between distortion-aware and quality-aware feature representations. The working parts are a distortion classifier that labels target projections, a distribution weight computed from estimated distortion distributions through a quadratic program, a conditional discriminator trained with a distortion-weighted cross-entropy loss, and a contrastive loss over carefully chosen positive and negative samples. Together they re-weight the source feature distribution toward distortions common to both media while preserving the features that predict perceptual quality.","core_discovery":"The central claim is that cross-media blind quality assessment can be achieved by decomposing the standard domain adaptation objective into two subproblems: aligning feature distributions conditioned on distortion, and preserving the feature-to-quality mapping under biased distortion distributions. The method re-weights source features by the ratio of target to source distortion probabilities, computed from distortion pseudo-labels produced by a classifier, and feeds the re-weighted features into a conditional discriminator. Simultaneously, contrastive learning pulls together features with the same content and distortion while pushing apart features with different distortions, which keeps the aligned representation both distortion-aware and quality-aware. On image datasets as the source and point cloud datasets as the target, the paper reports SROCC gains over the prior transfer method of roughly 20 to 40 percent, landing close to fully supervised point cloud quality metrics.","pith_inferences":["Beyond the paper: the classifier's distortion accuracy on target point-cloud projections is never reported, so a decisive check is to measure it; near-chance accuracy would mean the computed weights are pseudo-label noise.","Beyond the paper: the transfer assumes the source distortion taxonomy can express point-cloud distortions, and genuinely novel target distortion types would likely be under-aligned by the importance weighting.","Beyond the paper: the same distortion-as-transition decomposition could apply to other cross-media quality tasks, such as video-to-mesh or rendered-to-real scenes, wherever a shared feature extractor and a distortion-type predictor are available."],"forward_implications":["Point cloud quality assessment for new scenes no longer requires point cloud quality labels; source image quality labels and unlabeled point cloud data suffice.","The transfer gap to fully supervised point cloud quality metrics is small, around 7 percent SROCC on one dataset and 4.8 percent on another, so the unlabeled approach is practically usable.","The reported gains over the previous image-to-point-cloud transfer method, roughly 20 percent SROCC on SJTU-PCQA and 40 percent on WPC, indicate that distortion-distribution weighting addresses a real failure mode of direct feature alignment.","Ablation results show each component contributes: plain image-trained regression reaches 0.667 SROCC, adding distortion-weighted alignment raises it to 0.749, and the full model reaches 0.828 on the TID2013-to-SJTU-PCQA setup."],"supporting_citations":[{"why":"The prior image-to-point-cloud transfer method that DWIT-PCQA compares against and improves on.","marker":"[15]"},{"why":"The source image dataset with MOS and distortion-type labels used to train feature extraction, classification, and regression.","marker":"[19]"},{"why":"The target point cloud dataset with subjective MOS used to evaluate transfer performance.","marker":"[13]"},{"why":"The second target point cloud dataset used to evaluate transfer performance.","marker":"[14]"},{"why":"Supplies the adversarial domain adaptation framework that the conditional discriminator builds on.","marker":"[26]"},{"why":"Supplies the error-decomposition theorem bounding the source-target error gap, used to motivate distortion-conditioned alignment.","marker":"[64]"},{"why":"Supplies the clustering-structure and generalized label-shift results used to justify computing weights from distortion pseudo-labels.","marker":"[66]"},{"why":"Supplies the quadratic-program formulation used to estimate the distortion distribution weights.","marker":"[67]"},{"why":"Supplies the contrastive objective used in the quality-aware feature disentanglement module.","marker":"[68]"}],"fun_headline_variants":["Image-trained model scores point clouds with no labels","Zero-shot point cloud quality via image transfer","Blind PCQA: Transfer image knowledge, skip annotations","Distortion-aware domain adaption scores unlabeled point clouds","Cross-media quality: From images to point clouds without labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the distortion classifier, trained only on source images, estimates the target point-cloud distortion distribution well enough that the computed importance weights improve alignment rather than adding noise.","fun_headline_variants_meta":{"raw":{"variants":["Image-trained model scores point clouds with no labels","Zero-shot point cloud quality via image transfer","Blind PCQA: Transfer image knowledge, skip annotations","Distortion-aware domain adaption scores unlabeled point clouds","Cross-media quality: From images to point clouds without labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000157,"raw_usage":{"total_tokens":1229,"prompt_tokens":959,"completion_tokens":270,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":194}},"tokens_in":575,"tokens_out":270,"duration_ms":7779,"temperature":1.0,"reasoning_tokens":194,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:59:57.578976+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the computed distortion weights with uniform weights in the distortion-weighted cross-entropy loss on the TID2013-to-SJTU-PCQA setup; if SROCC does not drop materially, the distortion-distribution weighting is not doing the work attributed to it.","supporting_citations":[{"cited_title":"Perceptual quality assessment of 3d point clouds,","cited_arxiv_id":null,"evidence_quote":"The second target point cloud dataset used to evaluate transfer performance."},{"cited_title":"No-reference point cloud quality as- sessment via domain adaptation,","cited_arxiv_id":null,"evidence_quote":"The prior image-to-point-cloud transfer method that DWIT-PCQA compares against and improves on."},{"cited_title":"Image database TID2013: Peculiarities, results and perspectives,","cited_arxiv_id":null,"evidence_quote":"The source image dataset with MOS and distortion-type labels used to train feature extraction, classification, and regression."},{"cited_title":"Adversarial discrimi- native domain adaptation,","cited_arxiv_id":null,"evidence_quote":"Supplies the adversarial domain adaptation framework that the conditional discriminator builds on."},{"cited_title":"On learning invariant representations for domain adaptation,","cited_arxiv_id":null,"evidence_quote":"Supplies the error-decomposition theorem bounding the source-target error gap, used to motivate distortion-conditioned alignment."},{"cited_title":"Do- main adaptation with conditional distribution matching and generalized label shift,","cited_arxiv_id":null,"evidence_quote":"Supplies the clustering-structure and generalized label-shift results used to justify computing weights from distortion pseudo-labels."},{"cited_title":"Detecting and correcting for label shift with black box predictors,","cited_arxiv_id":null,"evidence_quote":"Supplies the quadratic-program formulation used to estimate the distortion distribution weights."}],"review_version":1}