{"id":"47a127b5-dd17-4951-9ba8-541cb3d78b15","arxiv_id":"2501.01723","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"IGAF, an incremental attention fusion network, reports state-of-the-art RMSE on NYU v2 and zero-shot gains on Middlebury, Lu, and RGB-D-D depth super-resolution benchmarks.","lead":"This paper introduces IGAF, a neural network module that fuses low-resolution depth maps with high-resolution color images to produce sharper depth maps. The method reports slightly better accuracy than previous models on four standard depth super-resolution benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on an incomplete baseline set; the paper cites recent guided-DSR methods (DCTNet, SDFNet) but omits them from every comparison table.","rationale":"The reader's weakest_assumption was about bicubic downsampling vs real sensor degradation. That is a legitimate generalization concern, but the paper does include a real-world dataset experiment (Table 4) where IGAF leads, which partially addresses it. The more load-bearing issue for the stated central claim is the completeness of the comparison set. The paper itself lists DCTNet and SDFNet in the literature review as relevant techniques but excludes them from all quantitative comparisons; because the headline is explicitly 'outperforms all baselines,' a single omitted method with a better published number invalidates the claim. This is a concrete, checkable threat. The reader's rationale also noted missing newer baselines and lack of error bars, so we partially agree with the reader's diagnosis, though we weight the baseline omission more heavily than the degradation model. The verdict remains CONDITIONAL pending the concrete comparison test; if omitted methods are shown to beat IGAF, it should move to REJECT, but that is not yet established.","tokens_in":12184,"tokens_out":6002,"duration_ms":54101,"concrete_test":"Retrieve the official reported RMSE values for DCTNet (Zhao et al., CVPR 2022, ref. 28) and SDFNet (Zhao et al., 2023, ref. 29) on NYU v2 for x4/x8/x16 and on the zero-shot datasets (Middlebury, Lu, RGB-D-D) using the same evaluation protocol, and insert them into Tables 1, 2, 3, and 5. If either method is below IGAF on any row, the SOTA claim is refuted. Separately, retrain IGAF with 5 seeds on NYU v2 and report mean plus/minus standard deviation for x4, x8, x16 to check whether the 0.02 RMSE margin over SUFT is within noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical SOTA: IGAF 'outperforms all baselines' on NYU v2 and zero-shot datasets. That claim can only be evaluated against the methods actually compared. The comparison tables include DG, SVLRM, DKN, FDSR, SUFT, CTKT, JIIF, and a few others, but Section 2 cites DCTNet (ref. 28, CVPR 2022) and Spherical Space Feature Decomposition (ref. 29, 2023) as relevant prior work; neither appears in Tables 1-5, nor do other recent guided-DSR methods mentioned in the survey. Since these works report results on the same benchmarks (NYU v2, Middlebury, Lu) and were published before this preprint, their absence is not due to availability. If either achieves lower RMSE than IGAF (e.g., below 1.12 for NYU v2 x4), the abstract's 'all baseline models' statement is false. Secondary, the margins over SUFT are small (NYU x4: 1.12 vs 1.14; Lu x16: 4.14 vs 4.16) and no variance or multi-seed statistics are reported, so the improvements may not be significant even against the included baselines. The bicubic degradation assumption is real but partly mitigated by the real-world RGB-D-D experiment; the omitted-baselines problem directly threatens the headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes IGAF, a dual-stream convolutional architecture for guided depth super-resolution. An RGB stream and a bicubically upsampled LR depth stream are processed by three incremental guided attention fusion (IGAF) modules, each built from filtered wide-focus (FWF) feature extractors and two spatial attention fusion (SAF) blocks; the final output is a depth-refinement head added to a global residual connection. The model is trained on NYU v2 with an L1 loss and evaluated at x4, x8, and x16 on NYU v2, Middlebury, Lu, and RGB-D-D, including a real-world RGB-D-D setting with sensor-collected LR/HR pairs. The paper reports lower RMSE values than the included baselines in every configuration and provides ablations of the fusion module on NYU v2 x4.","tokens_in":12460,"tokens_out":5678,"duration_ms":56388,"significance":"If the reported improvements are statistically reliable, IGAF would be a competitive method for guided depth super-resolution, and the IGAF module could be a useful reusable cross-modal fusion component. The manuscript has concrete strengths: the code and trained models are promised on GitHub, the architecture is described at module level, zero-shot transfer is tested on three additional datasets, and a real-sensor RGB-D-D experiment partially addresses the synthetic-degradation concern. The main weakness is that the empirical evidence is not yet strong enough to support the headline state-of-the-art claim: recent methods cited in the paper itself are absent from all comparison tables, and no uncertainty or significance information is reported for any RMSE value.","major_comments":[{"comment":"The abstract and Section 5 claim state-of-the-art results compared to 'all baseline models', but the comparison tables include only a subset of the methods cited as relevant prior work. DCTNet (ref. 28) and the spherical-space feature decomposition method (ref. 29), both cited in Section 2, do not appear in any table, and other recent guided-DSR methods discussed in the survey [6] are likewise absent. Because these methods report results on the same benchmarks, their omission prevents the reader from verifying the central claim. Please add comparisons to these methods, or explicitly restate the claim as being with respect to the included baselines only.","section":"§2, §5 (Tables 1–5)"},{"comment":"Every quantitative result is a single RMSE value with no standard deviation, number of seeds, or significance test. Several reported margins are very small, for example NYU v2 x4 (1.12 vs. SUFT 1.14), Lu x16 (4.14 vs. JIIF 4.16), and Middlebury x16 (3.24 vs. JIIF 3.31). Without repeated runs or error bars, these improvements cannot be distinguished from run-to-run variation, so the state-of-the-art conclusion is not statistically supported. Reporting mean +/- std over at least three seeds, or otherwise providing variance information, is necessary for the headline empirical claim.","section":"§5, Tables 1–5"},{"comment":"The ablation study is conducted only on NYU v2 at x4, and several component-wise differences in Table 7 are as small as 0.02 RMSE (for example, 1.14 vs. 1.12 between the relocated-skip variant and the full model). With single runs, the conclusion that each component is necessary is not supported. In addition, design choices such as N=10 in the FWF module and the number of IGAF modules are selected on NYU v2, so the zero-shot results should be interpreted as obtained with NYU-v2-tuned hyperparameters; the paper should either extend the ablation or qualify this point.","section":"§6, Tables 6–7"},{"comment":"The evaluation protocol uses LR depth maps produced by bicubic downsampling of HR ground truth, as stated in Section 4. Although this is consistent with prior work and the real-world RGB-D-D experiment mitigates the concern, the abstract's generalization claim for Middlebury, Lu, and RGB-D-D is stated without this qualification. Please explicitly qualify the state-of-the-art claim as holding under the bicubic-degradation protocol, or add experiments with sensor-realistic noise and missing pixels.","section":"§4"}],"minor_comments":[{"comment":"The figures contain embedded 'Click to buy NOW!' and 'PDF-XChange Product' watermark text that should be removed from the published version.","section":"Figures 1–4"},{"comment":"The notation A_A1, b_A1, A_A2, and b_A2 is not defined; please specify the shapes and roles of these weight and bias tensors.","section":"§3.2.1, Eqs. (5)–(6)"},{"comment":"The paper states that RMSE is the evaluation metric but does not give its formula or unit; a one-line definition would improve reproducibility.","section":"§4"},{"comment":"The concluding paragraph repeats 'in a zero-shot setting' twice; the sentence should be rephrased.","section":"§7"},{"comment":"Some reference labels are shared between tables with different method names (for example DKN/FDKN and DJF/DJFR); please verify that each table entry points to the correct bibliography entry.","section":"Tables 2 and 4"}],"recommendation":"major_revision","confidential_remarks":"The central architectural idea is reasonable and the real-world RGB-D-D experiment is a useful addition, but the empirical comparison is the deciding issue. The missing recent baselines and the absence of any variance reporting directly undermine the state-of-the-art claim, so I cannot recommend acceptance in the current form. No circularity issue arises in the benchmark evaluation itself."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a coherent, incremental architecture paper with a real but modest fusion idea, and an empirical evaluation that does not support the headline SOTA claim as written. The core IGAF module is clearly specified; Equations (5)-(7) describe a two-step cross-wise attention weighting that is a reasonable variation on WAFP-Net's weighted fusion. The FWF block is reused from the authors' earlier medical imaging work, which is fine as long as it is cited, and it is. The ablation on NYU v2 x4 shows every removed component hurts, with differences of 0.02-0.05 RMSE, and the real-world RGB-D-D experiment is a legitimate attempt to address the bicubic-downsampling worry. Zero-shot results on four datasets are useful if you trust the numbers.\n\nThe soft spots are not hidden. The baseline set is missing exactly the recent methods cited in Section 2: DCTNet and Spherical Space Feature Decomposition are mentioned but never appear in Tables 1-5. Both report on the same benchmarks before this preprint, so the absence is not availability. With those absent, the abstract's 'all baseline models' is not established. The margins over SUFT are around 0.02-0.05 RMSE, and with single runs, no seeds, and no significance testing, I can't tell whether the differences are real. The ablation is limited to one dataset and one scale, and the hyperparameter text ('empirically estimated') doesn't specify the search. The dangling phrase 'as can be verified through our ablations in' points to nothing; minor production issue. The self-cited WF block is component reuse, not circularity, so I don't hold that against them.\n\nWho is this for? Practitioners in guided depth SR who want a clear description of a plausible incremental fusion module. I wouldn't cite it as SOTA until the baselines and variance are added, but the architecture is concrete, the code is promised, and the claims are checkable. A serious referee could fix this with a major revision.","headline":"Incremental attention-fusion architecture for depth SR with a plausible module but an under-supported SOTA claim due to missing recent baselines and single-run metrics.","tokens_in":12989,"tokens_out":2703,"would_cite":false,"duration_ms":26913,"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":"This paper claims that incrementally fusing RGB and depth features with learned cross-modal attention outperforms all compared guided depth super-resolution models on four benchmarks.","keywords":["guided depth super-resolution","multimodal sensor fusion","attention fusion","cross-modal attention","convolutional neural networks","zero-shot generalization","depth map upsampling","RGB-D"],"falsifier":"Run the released model on low-resolution depth maps degraded with realistic sensor artifacts, such as randomly missing depth pixels, sensor noise, and slight RGB-depth misalignment, and compare RMSE against SUFT and JIIF; if IGAF no longer improves on those baselines under this degradation, the paper's claim of generalizable state-of-the-art performance fails. A simpler check is to retrain the model several times on NYU v2 with different random seeds and see whether the reported x4 RMSE of 1.12 is reproduced within a small margin.","tokens_in":11992,"feed_emoji":"🖼️","tokens_out":7452,"duration_ms":67227,"temperature":0.7,"pith_summary":"Guided depth super-resolution takes a low-resolution depth map and a sharp RGB image and tries to reconstruct a high-resolution depth map; the paper argues that the fusion step between the two modalities is where prior methods fail. Its claim is that fusing RGB structure into depth in one shot lets irrelevant textures leak through, causing blur, depth bleeding, and misaligned edges, and that this can be avoided by fusing incrementally: a naive joint representation is refined through two stages of attention-weighted cross-fusion before the depth stream moves on. The proposed model, built from three repeating IGAF modules, reports the lowest RMSE on NYU v2 at x4, x8, and x16 upsampling, and, trained only on NYU v2, also reports the lowest RMSE when transferred without fine-tuning to the Middlebury, Lu, and RGB-D-D datasets. It further reports the best result on RGB-D-D's real-sensor pairs, where low- and high-resolution depth come from different devices rather than from downsampling. If the results hold, the work makes a concrete case that iterative attention fusion is the component that lets depth super-resolution generalize across sensors and scenes.","feed_headline":"Incremental attention fusion beats all depth upsampling baselines","feed_subtitle":"One network, trained only on NYU v2, tops three unseen datasets and real-sensor depth pairs.","key_machinery":"The load-bearing object is the incremental guided attention fusion (IGAF) module, a fusion cell that consumes RGB and depth feature maps and returns refined versions of both. Inside, a filtered wide-focus block (FWF) first extracts multi-scale features: a channel-attention feature extractor is followed by three dilated-convolution branches with linearly increasing dilation rates, so the network sees both fine texture and broad structure. The actual fusion is done by two spatial attention fusion (SAF) blocks, each of which generates attention weights with a two-layer MLP from one input and applies those weights to the other input in a crosswise manner; the first SAF fuses RGB features with the element-wise product of the two streams, and the second fuses that joint representation with the depth stream. This two-step, weight-gated cross-multiplication is the mechanism that the paper credits for transferring only relevant RGB structure and suppressing depth bleeding.","core_discovery":"On the paper's own terms, the central discovery is that a cross-modal attention mechanism applied in two sequential steps, not one, is what makes guided depth super-resolution work. The first spatial attention fusion block combines the element-wise product of RGB and depth features with the RGB features; the second takes that output, passes it through a convolutional layer, and fuses it with the depth features. Because each block creates two-layer MLP attention weights and applies them crosswise, the depth stream is updated by RGB structure only where the naive fusion indicates the two modalities agree. The authors show this design outperforms addition and concatenation fusion (x4 RMSE 1.12 versus 1.23 and 1.22 on NYU v2) and that each component, including wide-focus dilation, two-layer MLP weights, and skip placement, contributes to the final accuracy. Their conclusion is that incremental guided fusion gives sharper edges, less bleeding, and better zero-shot transfer than the single-stage fusion used by previous state-of-the-art methods.","pith_inferences":["Independently re-running the released code on NYU v2 with the same train/test split would settle whether the reported margins, such as 1.12 versus 1.14 for SUFT at x4, are reproducible under different random seeds; the paper gives mean RMSE without error bars.","A natural next experiment is to degrade low-resolution depth with sensor-realistic artifacts, including missing pixels, infrared speckle noise, and RGB-depth misalignment, and test whether the zero-shot advantage over SUFT and JIIF survives; if it does, the method is a stronger candidate for real robotics and augmented-reality pipelines.","Because the IGAF module is a generic fusion cell, it could be lifted into other guided restoration tasks such as disparity upsampling, normal-map refinement, or low-light image enhancement, where a high-resolution guide channel carries structure that must not be over-transferred.","The paper does not report latency, parameter count, or FLOPs against baselines, so a practical follow-up would measure whether the reported accuracy gain is worth the compute in real-time systems."],"forward_implications":["On NYU v2 the model reports RMSE of 1.12 at x4, 2.48 at x8, and 5.00 at x16, each better than every baseline listed in the paper's comparison.","Because it is trained only on NYU v2, the Middlebury, Lu, and RGB-D-D results are zero-shot transfers; the paper's claim is that attention-based incremental fusion generalizes without retuning.","On RGB-D-D's real-world pairs, where the low- and high-resolution depth images come from different sensors, the model reports RMSE 7.01, lower than all compared baselines, suggesting the benefit is not an artifact of bicubic degradation alone.","The ablation replacing IGAF with addition or concatenation raises NYU v2 x4 RMSE to 1.23 or 1.22, and removing the wide-focus block, the two-layer MLP weights, or the chosen skip placement each worsens accuracy, so the paper attributes the gain to the fusion design rather than model scale.","The three repeated IGAF modules feed only the depth stream into the final refinement block after the last module, so the architecture also shows that the RGB stream can be dropped at the output stage once guidance has been applied."],"supporting_citations":[{"why":"The strongest prior baseline, compared on all four datasets; Figure 5 shows IGAF against SUFT qualitatively, so the central state-of-the-art claim is defined relative to it.","marker":"[24]"},{"why":"Introduces the RGB-D-D benchmark with real-sensor low- and high-resolution pairs and the FDSR baseline; the real-world generalization claim depends on this data.","marker":"[23]"},{"why":"Joint implicit image function baseline that appears in every comparison table and gives the closest prior results on several settings.","marker":"[25]"},{"why":"Cross-task knowledge transfer baseline (CTKT) in the NYU v2 and RGB-D-D tables; the claim of beating all baselines requires it.","marker":"[26]"},{"why":"NYU v2, the only training set; all results and the zero-shot claim start from this dataset.","marker":"[36]"},{"why":"Lu dataset used as one of the zero-shot evaluation sets; the paper's generalization claim covers it.","marker":"[39]"},{"why":"Deformable kernel network baseline used in the NYU v2, RGB-D-D, Lu, and Middlebury tables.","marker":"[42]"},{"why":"Middlebury dataset, another zero-shot evaluation set the paper claims to outperform on.","marker":"[37,38]"}],"fun_headline_variants":["Two-step attention fusion sharpens depth super-resolution","Incremental attention fusion tops depth SR benchmarks","Cross-modal attention in two steps beats single-stage depth SR","Zero-shot depth super-resolution via incremental attention fusion","IGAF: two-step attention fusion for state-of-the-art depth SR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The experiments assume that a low-resolution depth map produced by bicubic downsampling of a high-resolution ground truth behaves like the output of a real depth sensor, so the reported state-of-the-art numbers may not transfer to the missing pixels, noise, and misalignment found in actual sensor data.","fun_headline_variants_meta":{"raw":{"variants":["Two-step attention fusion sharpens depth super-resolution","Incremental attention fusion tops depth SR benchmarks","Cross-modal attention in two steps beats single-stage depth SR","Zero-shot depth super-resolution via incremental attention fusion","IGAF: two-step attention fusion for state-of-the-art depth SR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000781,"raw_usage":{"total_tokens":3456,"prompt_tokens":959,"completion_tokens":2497,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":2419}},"tokens_in":575,"tokens_out":2497,"duration_ms":19020,"temperature":1.0,"reasoning_tokens":2419,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:21:09.045104+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released model on low-resolution depth maps degraded with realistic sensor artifacts, such as randomly missing depth pixels, sensor noise, and slight RGB-depth misalignment, and compare RMSE against SUFT and JIIF; if IGAF no longer improves on those baselines under this degradation, the paper's claim of generalizable state-of-the-art performance fails. A simpler check is to retrain the model several times on NYU v2 with different random seeds and see whether the reported x4 RMSE of 1.12 is reproduced within a small margin.","supporting_citations":[{"cited_title":"Symmetric Uncertainty-Aware Feature Transmission for Depth Super-Resolution","cited_arxiv_id":null,"evidence_quote":"The strongest prior baseline, compared on all four datasets; Figure 5 shows IGAF against SUFT qualitatively, so the central state-of-the-art claim is defined relative to it."},{"cited_title":"Towards fast and accurate real-world depth super-resolution: Benchmark dataset and baseline","cited_arxiv_id":null,"evidence_quote":"Introduces the RGB-D-D benchmark with real-sensor low- and high-resolution pairs and the FDSR baseline; the real-world generalization claim depends on this data."},{"cited_title":"Joint implicit image function for guided depth super-resolution","cited_arxiv_id":null,"evidence_quote":"Joint implicit image function baseline that appears in every comparison table and gives the closest prior results on several settings."},{"cited_title":"Learning scene structure guidance via cross-task knowledge transfer for single depth super-resolution","cited_arxiv_id":null,"evidence_quote":"Cross-task knowledge transfer baseline (CTKT) in the NYU v2 and RGB-D-D tables; the claim of beating all baselines requires it."},{"cited_title":"Indoor segmentation and support inference from rgbd images","cited_arxiv_id":null,"evidence_quote":"NYU v2, the only training set; all results and the zero-shot claim start from this dataset."},{"cited_title":"Depth enhancement via low-rank matrix completion","cited_arxiv_id":null,"evidence_quote":"Lu dataset used as one of the zero-shot evaluation sets; the paper's generalization claim covers it."},{"cited_title":"Deformable kernel networks for guided depth map upsampling","cited_arxiv_id":"1903.11286","evidence_quote":"Deformable kernel network baseline used in the NYU v2, RGB-D-D, Lu, and Middlebury tables."}],"review_version":1}