{"id":"05a2081c-f73d-4ae2-a6be-fab2337f0cf7","arxiv_id":"2506.23542","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GIGA-ToF denoises ToF depth video by fusing motion-invariant per-pixel correlation graphs across frames with an unrolled graph-regularized network, beating prior methods on the synthetic DVToF benchmark.","lead":"This paper introduces a multi-frame Time-of-Flight depth denoising network that fuses graph structures, which stay similar across frames even when depth values shift, instead of fusing depth features directly. The method reports large accuracy gains on synthetic data and qualitative robustness on real Kinect v2 data, and could improve temporal consistency for 3D cameras.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (6) makes the fused graph asymmetric for per-pixel confidence, so the graph Laplacian in Eq. (9) is not PSD and the claimed low-pass MAP unrolling is mathematically unsupported.","rationale":"The reader identified the reliability of cross-frame graph mapping as the weakest assumption. I disagree that this is the most load-bearing issue. Even if the mapping is perfectly reliable, the graph fusion formula as written breaks the symmetry required for the entire MAP and spectral-filter interpretation. The central claim in the abstract and Sec. 4.2 is that the network is obtained by unrolling the MAP solution of a graph-smoothness prior on the fused graph. That requires the graph Laplacian to be symmetric positive semidefinite; otherwise the 'prior' is not a proper prior, the low-pass filtering interpretation is false, and the unrolled update is not solving the stated optimization. This is an internal mathematical inconsistency, not a disagreement with consensus. The empirical state-of-the-art numbers could still be true, but the paper's theoretical contribution, motion-invariant graph fusion with an interpretable MAP unrolling, would be invalid as stated. A conditional acceptance with the requirement to symmetrize the fusion (e.g., Wtilde = Phi^{1/2} What Phi^{1/2} + W) and re-derive the unrolling, or to explicitly treat the network as a directed graph filter and drop the PSD and low-pass claims, is the appropriate verdict. This reinforces the reader's conditional verdict but for a different, more fundamental reason, so the verdict itself is unchanged.","tokens_in":15967,"tokens_out":23507,"duration_ms":224417,"concrete_test":"Analytically compute a two-pixel instance: set What^{t-1} = [[0,1],[1,0]], W^t = 0, Phi = diag(1,2). Eq. (6) gives Wtilde^t = [[0,1],[2,0]], so Wtilde_12 = 1 != Wtilde_21 = 2. The Laplacian L = diag(row sums) - Wtilde = [[1,-1],[0,0]] gives x^T L x = x_1^2 - x_1 x_2, which is negative for x = (1,2), disproving PSD. This check settles that the fused Laplacian is not the symmetric PSD matrix assumed in Sec. 5.2. If the authors intend a symmetric confidence application (e.g., Phi^{1/2} What Phi^{1/2} + W), the reported Eq. (6) and the network implementation must be revised accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing concern is an internal inconsistency in the core graph construction. Eq. (6) defines the fused graph as Wtilde^t = Phi^{t,t-1} What^{t-1} + W^t, where Phi is a non-negative diagonal 'mapping confidence' matrix. Since What^{t-1} and W^t are symmetric, the transpose of Wtilde^t is What^{t-1} Phi + W^t. Equality with Wtilde^t requires Phi What^{t-1} = What^{t-1} Phi, which fails for a spatially varying diagonal Phi (e.g., two pixels with phi_1 != phi_2). Thus the fused graph is asymmetric. The graph Laplacian Ltilde^t = Dtilde^t - Wtilde^t in Eq. (9) is then not symmetric, and the quadratic prior x^T Ltilde x is not non-negative in general; a two-node example suffices. Sec. 5.2 explicitly claims this Laplacian is symmetric and positive semi-definite and concludes the solution is a low-pass graph spectral filter. That claim is false under Eq. (6). Consequently the MAP objective in Eq. (11) is not a proper Bayesian prior, the unrolling in Eqs. (12)-(14) is not solving a PSD graph-regularized problem, and the paper's stated interpretability and generalization argument loses its foundation. The network may still perform well empirically, but the central theoretical claim of the title and abstract is unsupported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GIGA-ToF, a multi-frame Time-of-Flight depth denoising network that fuses graph structures across frames rather than depth features. The authors define intra-frame graphs on ToF raw data, map the reference frame's graph into the current frame via a cross-frame geometric attention mechanism, and form a MAP problem with a graph Laplacian smoothness prior on the fused graph. The solution is unrolled into iterative graph filters, yielding an interpretable network. Experiments on a self-generated synthetic DVToF dataset report state-of-the-art MAE, AbsRel, δ1, and TEPE, with qualitative generalization results on real Kinect v2 data.","tokens_in":16328,"tokens_out":7807,"duration_ms":74811,"significance":"The core idea of exploiting temporal self-similarity of graph structures instead of depth features is novel and potentially impactful for temporal ToF denoising. The reported improvements are large (at least 37.9% MAE and 13.2% TEPE over the best compared methods), and the claimed cross-dataset generalization to Kinect v2 is practically important. The paper also ships source code and an algorithm unrolling formulation, which are positive for reproducibility. However, the theoretical claims about the MAP objective and graph spectral filtering currently rest on a symmetry/PSD property that the construction does not satisfy, and the empirical claims lack error bars, so the significance can only be assessed after these issues are resolved.","major_comments":[{"comment":"The fused graph Wtilde^t = Phi^{t,t-1} What^{t-1} + W^t is not symmetric when Phi^{t,t-1} is a spatially varying diagonal matrix, because the transpose is What^{t-1} Phi^{t,t-1} + W^t, which equals Wtilde^t only if Phi commutes with What^{t-1}. Consequently the graph Laplacian Ltilde^t = Dtilde^t - Wtilde^t in Eq. (9) is not symmetric, and the quadratic form x^T Ltilde x is not non-negative in general; for a two-node example with W_12=a and W_21=b, x=(1,2)^T gives x^T L x = -a+2b, which can be negative. The claim in Sec. 5.2 that the graph Laplacian is symmetric positive semi-definite and that the solution is a low-pass graph spectral filter is therefore false. The MAP prior in Eq. (11) and the unrolled iteration in Eq. (14) are not supported by the stated construction. The paper should modify the fusion to preserve symmetry and PSD, e.g., Phi^{1/2} What Phi^{1/2} + W^t, and re-derive the unrolling accordingly.","section":"Sec. 4.2, Eq. (6); Sec. 5.2"},{"comment":"The fidelity term in Eq. (12) is not the MAP objective from Eq. (11). Whereas Eq. (7) and Eq. (24) depend on the observed noisy measurements y_i^t and y_q^t, Eq. (12) uses ||(Xa^{r-1})^{-1} xq^{r-1} ⊙ (xi - xi^{r-1})||^2, which effectively replaces y_i^t with the previous estimate xi^{r-1} and y_q^t with xq^{r-1}. The alternating minimization described in the text therefore does not solve the MAP problem stated in Eq. (11). The authors should justify this replacement as a proximal or surrogate approximation, or re-derive the unrolled filter from the original data fidelity term.","section":"Sec. 4.3, Eq. (12)"},{"comment":"All quantitative claims are based on a single run of a self-generated synthetic dataset with no standard deviations, number of seeds, or statistical tests. The central performance claims ('outperforming other methods by at least 37.9% in MAE and 13.2% in TEPE') should be accompanied by multiple training runs with different seeds and a report of variance, especially because the dataset generation involves randomly generated camera paths.","section":"Sec. 6.1 and Sec. 6.2, Table 1"},{"comment":"The paper does not specify how the mapping confidence matrix Phi^{t,t-1} is computed or how its non-negative diagonal structure is enforced in the network, despite this matrix being load-bearing for the fusion in Eq. (6). The text only states that the GIGA module computes W^{t-1}, W^{t,t-1}, and Phi^{t,t-1} at 1/8 scale. Without this architectural detail, the method is not fully reproducible and the asymmetry issue cannot be assessed empirically.","section":"Sec. 5.1 and Sec. 5.2"}],"minor_comments":[{"comment":"The notation 'fix x_q^t = y_q^t = x_q^{t,r-1}' conflates the observed measurement with the previous iterate; please clarify the initialization and the update order so that the reader can track which quantity is fixed in each alternating step.","section":"Sec. 4.3, Eq. (12)"},{"comment":"There is a typo in the first paragraph of Sec. 6: 'syntheic' should be 'synthetic'.","section":"Sec. 6"},{"comment":"UDA is cited as [1] in the text but as [2] in Table 1; the authors should reconcile whether they are referring to the CVPR 2019 paper or the IEEE TPAMI 2022 journal version.","section":"References and Table 1"},{"comment":"The description of DVToF dataset generation is brief; for reproducibility, please provide details of the random camera paths, the range of motions, the scene content, and the exact noise model parameters beyond citing [34].","section":"Sec. 6.1"},{"comment":"The output dimensions and module boundaries in Figure 3 are difficult to read at print resolution; please increase the font size or separate the three components more clearly.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong empirical component and an interesting idea, but the theoretical unrolling narrative needs substantial repair: the fused graph asymmetry breaks the PSD/low-pass interpretation, and the fidelity term in Eq. (12) differs from the stated MAP objective. Both issues are likely fixable within the manuscript's scope, but they currently affect the central claims. I would also encourage the editor to ask the authors to release the DVToF dataset generation code, not only the network code, to allow independent verification of the reported margins."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's core idea—fusing graph structures across frames rather than depth features—is genuinely new, and the ablations support it. The reported gains on DVToF are large, and the real-data generalization results are plausible. But there is a real math problem in the interpretability claim. Eq. (6) defines the fused graph as Wtilde^t = Phi^{t,t-1} What^{t-1} + W^t, with a spatially varying diagonal Phi. Since What and W^t are symmetric but Phi does not generally commute with What, Wtilde is asymmetric. Then the Laplacian Ltilde = Dtilde - Wtilde is not symmetric, let alone PSD. Section 5.2 explicitly claims it is symmetric PSD and that the solution is a low-pass graph spectral filter. That claim is false. The MAP objective in (11) is then not a proper Bayesian prior, and the unrolling in (14) is not solving the stated PSD-regularized problem. The network may still work empirically, but the central theoretical foundation is unsupported. There is also a silent approximation in Eq. (12), where y_i and y_q are replaced by previous estimates in the fidelity term without comment. That is a smaller issue, but combined with the lack of released code/data and error bars, it means the empirical claims need a careful check. The graph-fusion mechanism itself is well motivated and the ablation study is convincing: graph fusion beats feature fusion, and attention helps. So the paper deserves a serious referee, but it needs a major revision to fix the mathematical claims or to reframe the contribution as purely empirical.","headline":"Genuinely new graph-fusion idea with strong reported results, but the paper's key interpretability claim rests on a symmetric-PSD Laplacian that Eq. (6) does not yield.","tokens_in":16868,"tokens_out":2162,"would_cite":true,"duration_ms":21323,"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":"Graph structures around a pixel stay similar across ToF frames even when depth values shift; fusing those graphs rather than depth features removes noise while preserving edges and temporal stability.","keywords":["time-of-flight depth denoising","graph signal processing","graph Laplacian regularization","algorithm unrolling","multi-frame fusion","temporal consistency","geometric attention"],"falsifier":"An annotated dataset with occlusion masks and known inter-frame correspondences would settle the claim: run GIGA-ToF on occluded and fast-moving regions and compare against a variant that sets the reference graph contribution to zero ($W^{t,t-1}=0$). If fusing the mapped reference graph does not reduce error in exactly those regions, or if the learned confidence $\\Phi^{t,t-1}$ does not down-weight mismapped pixels, then graph structures are not self-similar where temporal fusion matters most.","tokens_in":15771,"feed_emoji":"📷","tokens_out":13300,"duration_ms":123681,"temperature":0.7,"pith_summary":"Time-of-flight depth video is noisy, and existing denoisers either process frames independently, which causes temporal jitter, or fuse depth features across frames, which blurs details because the depth value at a given pixel shifts with camera and object motion. The paper argues that the neighborhood-correlation graph around each pixel is nearly motion-invariant even when depth values shift, so the right cross-frame information to fuse is the graph, not the depth. It builds an eight-connected intra-frame graph for each frame, maps the previous frame's graph into the current frame through learned cross-frame attention, and fuses the two with a per-pixel confidence weight. Denoising is then posed as a maximum-a-posteriori problem combining a ToF-noise data-fidelity term and a graph-smoothness prior on the fused graph, and the solution is unrolled into iterative filters, giving an interpretable network. On the synthetic DVToF test set the method reports mean absolute error of 0.0193 m and temporal end-point error of 0.0637 m, with qualitative transfer to real Kinectv2 data.","feed_headline":"Fusing graphs, not depth, sharpens ToF video","feed_subtitle":"New network fuses pixel-correlation graphs across frames; MAE 0.0193 m and TEPE 0.0637 m on DVToF.","key_machinery":"The load-bearing mechanism is cross-frame graph fusion: intra-frame graphs $W^t$ encode eight-neighbor pixel correlations within a frame, the inter-frame attention graph $W^{t,t-1}$ connects each current-frame pixel to a $7\\times7$ neighborhood in the previous frame, and the mapped graph $\\hat{W}^{t-1}=W^{t,t-1}(W^{t-1}+I)(W^{t,t-1})^{\\top}$ moves the reference graph along two-hop and three-hop paths. The fused graph $\\tilde{W}^t=\\Phi^{t,t-1}\\hat{W}^{t-1}+W^t$ is weighted by a learned per-pixel confidence $\\Phi^{t,t-1}$ so that unreliable mappings, for instance at occlusions, are down-weighted. This fused graph defines the graph Laplacian in the MAP smoothness prior, and unrolling the solution produces iterative low-pass graph filters whose weights come from attention-based graph learning, which is named the graph-informed geometric attention (GIGA) module.","core_discovery":"The central discovery is temporal self-similarity of graph structure: depth values at corresponding pixels vary across frames, but the graph encoding which neighboring pixels are correlated stays similar, so a reference frame's graph can be transported to the current frame and fused with it. The transport is a mapped graph $\\hat{W}^{t-1}=W^{t,t-1}(W^{t-1}+I)(W^{t,t-1})^{\\top}$, combining two-hop and three-hop paths through the inter-frame attention graph, and the fusion is $\\tilde{W}^t=\\Phi^{t,t-1}\\hat{W}^{t-1}+W^t$, where $\\Phi^{t,t-1}$ is a learned diagonal confidence that down-weights unreliable mapping. The fused graph supplies the Laplacian for a graph-smoothness prior, and the likelihood comes from the ToF depth-noise distribution; the resulting MAP problem is unrolled into iterative filters whose kernels are the learned graph edge weights. The unrolled solution acts as a low-pass graph spectral filter, which the paper credits for its robustness to unseen real noise.","pith_inferences":["The paper leaves the confidence matrix $\\Phi^{t,t-1}$ unvalidated against occlusion or large-displacement ground truth; an occlusion-annotated benchmark is the natural next test of the motion-invariant graph assumption.","The same principle could extend beyond ToF: any representation invariant to per-pixel value shifts but sensitive to local geometry, such as surface normals or local shape indices, could be fused across frames instead of depth features.","Since the mapped-graph formula is a linear transport of edge weights, composing it over multiple reference frames would yield a recurrent graph-state update, potentially handling variable-length video without redesign.","Because the unrolled step is a graph Laplacian low-pass filter, one could compare the spectral response of fused and single-frame graphs to quantify how much cross-frame information contributes; the paper does not report such an analysis."],"forward_implications":["If graph structures are as motion-invariant as claimed, multi-frame ToF denoising can drop explicit scene-flow alignment: the attention-based graph mapping absorbs the correspondence problem.","Accuracy and temporal consistency improve in the same fused graph, so the usual tradeoff between per-frame sharpness and temporal stability need not apply to this class of methods.","The unrolled MAP solution gives the network a graph-spectral interpretation as a low-pass filter, which is the paper's stated reason it transfers to real Kinectv2 noise despite synthetic-only training.","Only the previous frame is used as reference with a forward-only update, so the same graph-fusion step can be chained over longer sequences; the paper reports the two-frame version and lists extension to more frames as future work."],"supporting_citations":[{"why":"It supplies the unrolled graph-Laplacian-regularization solver and prior-weight parametrization that this paper extends from single-frame to cross-frame fused graphs.","marker":"[17]"},{"why":"It establishes the multi-frame ToF denoising baseline that fuses depth features and defines the forward-only previous-frame reference choice.","marker":"[9]"},{"why":"It provides the raw-ToF data generation protocol and phase-unwrapping used to build the DVToF synthetic training and test data.","marker":"[34]"},{"why":"It derives the ToF depth-noise distribution from which the data-fidelity term in the MAP objective is obtained.","marker":"[13]"},{"why":"It formulates the graph-Laplacian-regularization prior that is applied to the fused graph.","marker":"[26]"},{"why":"It provides the attention operation used to compute inter-frame graph weights from geometric features.","marker":"[8]"},{"why":"It supplies the Kinectv2 noise statistics used to generate realistic training noise and to assess cross-dataset generalization.","marker":"[14]"},{"why":"It supplies the weighted-mode-filter multi-frame baseline against which temporal consistency and detail preservation are compared.","marker":"[23]"}],"fun_headline_variants":["Graphs stay still as depth moves: ToF denoise","Motion-invariant graph fusion clears ToF noise","Unrolled MAP denoiser keeps ToF depth crisp","Graph self-similarity stabilizes ToF depth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method must reliably carry the previous frame's pixel-correlation graph into the current frame, and the only protection against wrong mapping is a learned per-pixel confidence that was never trained or tested on occlusion labels; if that mapping misfires under occlusion or large motion, the fused graph injects wrong edges and the claimed sharpness and temporal stability would degrade.","fun_headline_variants_meta":{"raw":{"variants":["Graphs stay still as depth moves: ToF denoise","Motion-invariant graph fusion clears ToF noise","Unrolled MAP denoiser keeps ToF depth crisp","Graph self-similarity stabilizes ToF depth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000559,"raw_usage":{"total_tokens":2693,"prompt_tokens":1018,"completion_tokens":1675,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":1610}},"tokens_in":634,"tokens_out":1675,"duration_ms":14692,"temperature":1.0,"reasoning_tokens":1610,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:38:49.025646+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An annotated dataset with occlusion masks and known inter-frame correspondences would settle the claim: run GIGA-ToF on occluded and fast-moving regions and compare against a variant that sets the reference graph contribution to zero ($W^{t,t-1}=0$). If fusing the mapped reference graph does not reduce error in exactly those regions, or if the learned confidence $\\Phi^{t,t-1}$ does not down-weight mismapped pixels, then graph structures are not self-similar where temporal fusion matters most.","supporting_citations":[{"cited_title":"Deep unrolled graph laplacian regularization for robust time-of-flight depth denoising","cited_arxiv_id":null,"evidence_quote":"It supplies the unrolled graph-Laplacian-regularization solver and prior-weight parametrization that this paper extends from single-frame to cross-frame fused graphs."},{"cited_title":"Exploiting dual-correlation for multi-frame time-of- flight denoising","cited_arxiv_id":null,"evidence_quote":"It establishes the multi-frame ToF denoising baseline that fuses depth features and defines the forward-only previous-frame reference choice."},{"cited_title":"Deep end-to-end time-of-flight imaging","cited_arxiv_id":null,"evidence_quote":"It provides the raw-ToF data generation protocol and phase-unwrapping used to build the DVToF synthetic training and test data."},{"cited_title":"Time-of-flight range measurement in low-sensing environ- ment: Noise analysis and complex-domain non-local denois- ing","cited_arxiv_id":null,"evidence_quote":"It derives the ToF depth-noise distribution from which the data-fidelity term in the MAP objective is obtained."},{"cited_title":"Graph laplacian regulariza- tion for image denoising: Analysis in the continuous domain","cited_arxiv_id":null,"evidence_quote":"It formulates the graph-Laplacian-regularization prior that is applied to the fused graph."},{"cited_title":"Interpretable lightweight transformer via unrolling of learned graph smoothness pri- ors","cited_arxiv_id":null,"evidence_quote":"It provides the attention operation used to compute inter-frame graph weights from geometric features."},{"cited_title":"Tackling 3d tof artifacts through learning and the flat dataset","cited_arxiv_id":null,"evidence_quote":"It supplies the Kinectv2 noise statistics used to generate realistic training noise and to assess cross-dataset generalization."},{"cited_title":"Depth video en- hancement based on weighted mode filtering","cited_arxiv_id":null,"evidence_quote":"It supplies the weighted-mode-filter multi-frame baseline against which temporal consistency and detail preservation are compared."}],"review_version":1}