{"id":"580a8a3b-5256-46c8-b348-46846249cd0e","arxiv_id":"2412.19225","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"SigNet turns sparse depth completion into a depth enhancement problem and reports state-of-the-art accuracy on NYUv2, DIML, SUN RGBD, and TOFDC with a small 3.3M-parameter model.","lead":"This paper introduces SigNet, a depth completion network that first fills sparse depth maps with classical image processing tools and then treats the result as a blurry image to be sharpened with RGB guidance. It reports top results on four depth benchmarks while using far fewer parameters than prior state-of-the-art models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The degradation model in Eq. (3) is asserted but never validated; the coarse depth from hole-filling is unlikely to be a convolution of the target depth, leaving the degradation-aware mechanism as an untested narrative.","rationale":"The reader's weakest assumption correctly identifies Eq. (3) as the load-bearing premise. My stress test agrees and sharpens it: the premise is not merely unproven but likely false as a generative description of the densification process, because IP-Basic and colorization fill holes through nearest-neighbor or optimization-based propagation, not by convolving the true depth with a small kernel. This does not by itself refute the empirical SOTA claim, since a network could still learn useful RGB-D fusion while the degradation narrative is inaccurate. Therefore the conditional verdict remains appropriate: the paper should provide either a direct validation of the degradation assumption or an ablation that removes the degradation branch and shows a substantial performance drop attributable to the modeled degradation, not to extra parameters or attention. I agree with the reader's diagnosis, and my proposed check would settle whether the degradation model is a real mechanism or a post-hoc interpretation.","tokens_in":14510,"tokens_out":5285,"duration_ms":59774,"concrete_test":"On DIML training data, take the exact densified coarse depth Z used by SigNet and the ground-truth depth Y. For each local patch, fit the best linear convolution kernel K_emp minimizing ||K_emp * Y - Z||² with regularization, and record the residual distribution. Compare against a fixed Gaussian kernel and against a shuffled-alignment baseline. If the best-fit convolution residual is not substantially smaller near invalid-depth regions and is not localized near edges, Eq. (3) is rejected as a description of coarse-depth formation. Additionally, on the trained network, apply the learned K(D) to Y and compare with K_emp; a large mismatch shows the degradation bridge does not model the actual relation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim transforms depth completion into degradation-aware enhancement, and this transformation rests on Eq. (3): Z = K(D; θ) ⊗ Y + n. The paper never tests whether coarse depth Z actually arises from such a convolution. In practice, Z is produced by IP-Basic morphological hole-filling or colorization-based interpolation, both nonlinear inpainting operations. These cannot generally be represented as convolving the true dense depth Y with a single kernel, even a spatially varying one, especially on DIML's edge shadows and irregular holes. The auxiliary loss Ld (Eq. 13) trains K to make K ⊗ Y approximate Z; since K is a learned function of the input-derived D, this objective can be satisfied by a degenerate kernel and does not verify that the degradation is edge-concentrated or that the DCT decomposition of D corresponds to physically meaningful degradation. If Eq. (3) is not a faithful model, then DADF's adaptive RGB selection and conditional Mamba have no principled basis, and the reported ablation gains could come from an attention mechanism with extra parameters rather than from modeling degradation. No diagnostic is provided linking the estimated degradation to actual coarse-depth errors, so the central conceptual claim remains empirically unfalsified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"SigNet reframes depth completion as depth enhancement: sparse depth is first densified by non-CNN tools (IP-Basic/colorization) to obtain a coarse dense map, which is then enhanced by a CNN under an assumed degradation model Z = K(D;θ) ⊗ Y + n. A Degradation-Aware Decomposition and Fusion (DADF) module estimates an implicit degradation representation, decomposes it via DCT into components, uses those components to select high-frequency RGB features, and fuses the resulting RGB-D features with a conditional Mamba. The paper reports state-of-the-art RMSE on NYUv2, DIML, SUN RGBD, and TOFDC with a 3.3M-parameter model, along with ablations, density experiments, complexity comparisons, and a KITTI failure-case analysis.","tokens_in":14805,"tokens_out":5584,"duration_ms":54366,"significance":"If the reported numbers are reproducible, the paper demonstrates a strong empirical result: replacing direct sparse-depth convolution with densification-plus-enhancement yields competitive or better depth completion with substantially fewer parameters. The DCT-based decomposition of an implicit degradation and the conditional Mamba fusion are interesting mechanisms, and the evaluation across four datasets is broad. The paper is also honest about a KITTI failure case, which is a positive sign. However, the central conceptual claim—that the method is degradation-aware in a meaningful sense—rests on an untested assumption about how coarse depth relates to target depth, and the reported gains could plausibly be attributed to the extra attention/global-modeling components rather than to the degradation mechanism.","major_comments":[{"comment":"The degradation model Z = K(D;θ) ⊗ Y + n is the conceptual foundation of the paper, but it is asserted rather than validated. The coarse depth Z is obtained by IP-Basic morphological hole-filling or colorization-based interpolation, which are nonlinear operations that need not be representable as a convolution of the target depth with any kernel. The auxiliary loss Ld in Eq. (13) only enforces that the learned K applied to the ground-truth Y reproduces Z, so a degenerate K can satisfy this objective without the representation D capturing the actual degradation mechanism. As a result, the claim that DADF is 'degradation-aware' is not currently supported by any direct diagnostic. I request at least one of: (i) quantitative evaluation of how well K(D;θ)⊗Y predicts the actual coarse-depth errors, (ii) an ablation in which D is replaced by a fixed or random condition, or (iii) visual/statistical evidence that the DCT-decomposed degradation components correlate with depth boundaries and errors. Without such evidence, the reported improvements could come from the attention or Mamba components rather than from modeling degradation.","section":"Sec. 3.1, Eq. (3)"},{"comment":"The degradation loss Ld uses the ground-truth dense depth Y to supervise the kernel K, so the degradation bridge is not 'self-supervised' as stated in the abstract and introduction; it is supervised by the same ground truth that trains the reconstruction loss Lr. If the intended claim is that no annotations beyond the existing GT depth are required, that should be stated precisely. The distinction matters because the paper presents the degradation bridge as a conceptual novelty, and the current terminology is misleading.","section":"Abstract and Sec. 3.3, Eq. (13)"},{"comment":"The state-of-the-art claim rests on single-run metrics with no error bars or repeated runs, and most baseline numbers are borrowed from AGG-Net and TPVD rather than re-run under the authors' protocol. Several improvements are small relative to typical run-to-run variation (e.g., NYUv2 RMSE 83 vs 88 mm and REL 0.012 vs 0.011; TOFDC RMSE 87 vs 92 mm). Please report mean ± std over at least three runs for the main comparisons and, if possible, re-run the strongest baselines under the same training settings. In addition, Table 3 mislabels the final row as 'v' when the text describes 'SigNet-vi'; as printed, the table does not unambiguously support the stated 13 mm gain from conditional Mamba.","section":"Sec. 4, Tables 1-4"}],"minor_comments":[{"comment":"Table 2 caption says the density comparison is on TOFDC, but Sec. 4.1.3 states the comparison is on DIML; please correct the dataset label.","section":"Table 2 and Sec. 4.1.3"},{"comment":"The implementation details call the loss weight γ, whereas Eq. (14) defines λ; unify the notation.","section":"Sec. 6.3, Eq. (14)"},{"comment":"The Figure 8 caption labels (a) as 'Iterations of DADF' and (b) as 'Mamba Fusion', while the text refers to Fig. 8(a) as Mamba fusion and Fig. 8(b) as iterations; the caption and text are inconsistent.","section":"Fig. 8 and Sec. 4.2"},{"comment":"Equation (1) uses h for the degradation while Eq. (3) uses a kernel K(D;θ); the relationship between h and K, and the origin of the implicit degradation D (which network, with what input), should be defined more precisely.","section":"Sec. 3.1, Eqs. (1) and (3)"},{"comment":"The paper does not state whether code or trained models will be released; please provide code or a clear reproducibility statement.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The degradation-model validation is the key gate for this paper. If the authors can provide a direct link between the estimated degradation and actual coarse-depth errors, or an ablation that isolates the degradation mechanism, the paper may become acceptable. The duplicate 'v' in Table 3 and the DIML/TOFDC mismatch in Sec. 4.1.3 suggest the final manuscript needs careful proofreading. The borrowed baseline numbers are common practice in this field, but the lack of error bars is worth pushing on given the small margins on some datasets."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper deserves a real referee. SigNet is a depth completion network that first densifies sparse depth with classical hole-filling, then treats the task as depth enhancement using a learned degradation kernel and a DCT-based selection of high-frequency RGB features, fused via a conditional Mamba. It reports the best RMSE on DIML, SUN RGBD, and TOFDC and near-parity on NYUv2 with only 3.3M parameters, an order of magnitude smaller than BPNet or AGG-Net. The ablations show each added component helps, and the efficiency story is credible.\n\nWhat is truly new is the combination: non-CNN densification, degradation representation, DCT decomposition for adaptive RGB feature selection, and conditional Mamba fusion. The 'completion as enhancement' framing is not new in itself, and the paper's own footnote concedes the equivalence with depth super-resolution. That weakens the 'for the first time' claim, but the specific architecture is original and clearly engineered.\n\nThe soft spots are real but not fatal. The load-bearing premise is Eq. (3): coarse depth Z equals a learned degradation kernel convolved with target depth Y plus noise. That is an assumption, not a derivation. Coarse depth from morphological hole-filling or colorization is not generally a pure convolution of the dense target, especially at edges and irregular holes. The auxiliary loss Ld trains the kernel to reproduce Z from Y, which is supervised by ground truth, not self-supervised as claimed. The stress-test note is right: no diagnostic links the estimated degradation to actual coarse-depth errors, so the degradation-aware mechanism might be benefiting from extra parameters and attention rather than from a physically meaningful model. That said, the reported numbers stand independently of the narrative; the architecture works even if the story is over-fitted.\n\nMinor issues: no code, no error bars, and baseline numbers are mostly borrowed from prior papers. The KITTI failure case is honestly acknowledged and suggests the method is better suited to indoor or short-range outdoor scenes.\n\nBottom line: useful and practical advance for embedded and real-time RGB-D systems. If the authors release code and add a validation of the degradation model, I would be comfortable with acceptance. As it stands, it is a clear conditional.\n\nRecommendation: send to peer review. The empirical contribution and efficiency are strong enough to warrant careful referee time.","headline":"SigNet is a well-engineered, unusually compact depth completion model with consistent SOTA numbers, but its degradation-bridge story is a fitted assumption rather than a verified mechanism.","tokens_in":15313,"tokens_out":1879,"would_cite":true,"duration_ms":17645,"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":"SigNet claims that depth completion is better solved as depth enhancement: densify first, then let a learned degradation kernel select RGB edges to correct the coarse depth, achieving state-of-the-art RMSE on four benchmarks with a…","keywords":["depth completion","depth enhancement","degradation-aware network","RGB-D fusion","conditional Mamba","discrete cosine transform","sparse depth densification","state space model"],"falsifier":"Re-run the DIML or NYUv2 experiments with the degradation loss $L_d$ removed while keeping the DADF module; if RMSE barely changes and the learned kernel $K$ converges to a near-identity kernel, the degradation bridge is not what carries the performance. Alternatively, construct a synthetic test where coarse depth is corrupted by non-edge-centered noise, such as large smooth-region offsets; if SigNet's edge-selection mechanism cannot correct it, the edge-concentration premise in the degradation model is falsified.","tokens_in":14329,"feed_emoji":"🧩","tokens_out":5435,"duration_ms":51543,"temperature":0.7,"pith_summary":"The paper tries to establish that depth completion can be recast as depth enhancement: first fill the sparse depth map with classical densification to get a coarse dense map, then treat the remaining error as a learned degradation that RGB edge information can compensate. SigNet makes this concrete by estimating a per-image degradation kernel that is decomposed in the DCT domain and used both to select high-frequency RGB features and to condition a Mamba sequence model that fuses RGB and depth. The payoff claimed is state-of-the-art RMSE on NYUv2, DIML, SUN RGBD, and TOFDC with only 3.3M parameters, e.g., 83 mm on NYUv2 and 55 mm on DIML. The redefinition matters because it removes the irregular-sparse-data mismatch that direct convolutions face, and it borrows a mature degradation-learning toolkit from image restoration.","feed_headline":"Depth filling as enhancement: 3.3M-param model tops four benchmarks","feed_subtitle":"SigNet densifies sparse depth first, then uses learned degradation to select RGB edges—55 mm RMSE on DIML.","key_machinery":"The load-bearing object is the degradation bridge in Eq. (3), $Z = K(D;\\vartheta) \\otimes Y + n$, which asserts that coarse depth $Z$ is a convolution of target depth $Y$ with a kernel generated from an implicit degradation map $D$, plus noise. DADF then decomposes $D$ with a 2D DCT into three components, builds an adaptive spectrum mask that multiplies RGB features to select high-frequency content, and uses the decomposed degradation as a condition inside a Mamba state-space block, computing the SSM matrices $B$ and $C$ from gated recurrent convolutions over the RGB-D features and degradation. This machinery converts the completion task into an enhancement task and supplies the mechanism for RGB-guided correction.","core_discovery":"SigNet's central claim is that the long-standing sparse-depth difficulty in depth completion is avoidable: dense coarse depth produced by non-CNN methods, together with a self-supervised degradation assumption $Z = K(D;\\vartheta) \\otimes Y + n$, turns completion into a depth super-resolution/enhancement problem. The degradation kernel $K$ is learned from an implicit degradation representation $D$ and is constrained by the loss $L_d = \\|K \\otimes Y - Z\\|_1$ so that the network explicitly models how coarse depth arises from target depth. Degradation typically sits at edges, so DADF decomposes $D$ via a 2D DCT into components, uses them to mask and select high-frequency RGB features, and feeds RGB-D features plus degradation into a conditional Mamba whose state parameters are generated from both the input and the degradation. The authors report that this design produces state-of-the-art numbers on four benchmarks and that each component, densification plus degradation, denoising, attention selection, DCT decomposition, and conditional Mamba, contributes in ablations.","pith_inferences":["Editorial inference: because the authors note depth completion and depth super-resolution share the same end goal, the degradation-aware DADF module could be applied directly to RGB-guided depth super-resolution with minimal changes.","Editorial inference: the reported failure on KITTI suggests the edge-localized degradation assumption is range- and sensor-dependent; a testable fix hinted by the paper is to supervise the degradation branch with dense edge maps from monocular depth priors.","Editorial inference: since the densification step is fixed and non-CNN, the method's accuracy ceiling may be tied to the quality of that initial fill; replacing it with a learned densifier could improve cases with large holes or very sparse input."],"forward_implications":["If SigNet's reported numbers hold, depth completion can be made both more accurate and far lighter: 3.3M parameters versus 89.9M for BPNet on NYUv2, with lower RMSE.","The densify-then-enhance recipe gives a general route around sparse-input convolution, so the same degradation bridge could apply to other dense prediction problems where input and output share resolution.","The self-supervised degradation loss $L_d$ links coarse and target depth without paired degradation examples, reducing the need for externally supplied degradation supervision.","Conditional Mamba improves over vanilla Vision Mamba by roughly 8 mm RMSE in ablations, showing that degradation-conditioned state parameters help global RGB-D fusion.","The method is stable across input densities: on TOFDC it beats TPVD at every tested density from 1% to 95%, with an average improvement of 9.7%."],"supporting_citations":[{"why":"Supplies the non-CNN densification tool $f_{ds}$ that turns sparse depth into coarse dense depth.","marker":"[14]"},{"why":"Colorization-based optimization used as an alternative densification tool for the coarse depth map.","marker":"[15]"},{"why":"Provides the degradation-model assumption $Z=hY+n$ that SigNet adapts into Eq. (3).","marker":"[59]"},{"why":"Vanilla Vision Mamba baseline that the conditional Mamba modifies by generating state parameters from degradation conditions.","marker":"[25]"},{"why":"State-space model backbone used for the conditional Mamba fusion block.","marker":"[10]"},{"why":"Strongest prior CNN baseline whose NYUv2, DIML, and SUN RGBD results SigNet is compared against and surpasses in RMSE.","marker":"[38]"},{"why":"Provides the TOFDC evaluation protocol and the prior state-of-the-art that SigNet outperforms by 5 mm RMSE.","marker":"[51]"}],"fun_headline_variants":["SigNet: depth completion as enhancement, SOTA on four benchmarks","Degradation-aware SigNet turns sparse depth filling into enhancement","SigNet uses learned degradation to pick RGB edges for depth","From completion to enhancement: SigNet's conditional Mamba tops four benchmarks","Densify first, then enhance: SigNet redefines depth completion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach stands on Eq. (3): that the coarse depth map is the target depth blurred by a learned, edge-concentrated kernel plus noise, so that RGB high-frequency information is the right signal to subtract the error; if real coarse-depth error is not mostly edge-localized blur, the adaptive selection and conditional Mamba lose their principled basis.","fun_headline_variants_meta":{"raw":{"variants":["SigNet: depth completion as enhancement, SOTA on four benchmarks","Degradation-aware SigNet turns sparse depth filling into enhancement","SigNet uses learned degradation to pick RGB edges for depth","From completion to enhancement: SigNet's conditional Mamba tops four benchmarks","Densify first, then enhance: SigNet redefines depth completion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00036,"raw_usage":{"total_tokens":1955,"prompt_tokens":965,"completion_tokens":990,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":900}},"tokens_in":581,"tokens_out":990,"duration_ms":9604,"temperature":1.0,"reasoning_tokens":900,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:48:32.070640+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the DIML or NYUv2 experiments with the degradation loss $L_d$ removed while keeping the DADF module; if RMSE barely changes and the learned kernel $K$ converges to a near-identity kernel, the degradation bridge is not what carries the performance. Alternatively, construct a synthetic test where coarse depth is corrupted by non-edge-centered noise, such as large smooth-region offsets; if SigNet's edge-selection mechanism cannot correct it, the edge-concentration premise in the degradation model is falsified.","supporting_citations":[{"cited_title":"In defense of classical image processing: Fast depth completion on the cpu","cited_arxiv_id":null,"evidence_quote":"Supplies the non-CNN densification tool $f_{ds}$ that turns sparse depth into coarse dense depth."},{"cited_title":"Colorization using optimization","cited_arxiv_id":null,"evidence_quote":"Colorization-based optimization used as an alternative densification tool for the coarse depth map."},{"cited_title":"Designing a practical degradation model for deep blind image super-resolution","cited_arxiv_id":null,"evidence_quote":"Provides the degradation-model assumption $Z=hY+n$ that SigNet adapts into Eq. (3)."},{"cited_title":"Bi- lateral propagation network for depth completion","cited_arxiv_id":null,"evidence_quote":"Strongest prior CNN baseline whose NYUv2, DIML, and SUN RGBD results SigNet is compared against and surpasses in RMSE."},{"cited_title":"Tri- perspective view decomposition for geometry-aware depth completion","cited_arxiv_id":null,"evidence_quote":"Provides the TOFDC evaluation protocol and the prior state-of-the-art that SigNet outperforms by 5 mm RMSE."}],"review_version":1}