{"id":"277e37b7-bc82-4f3f-9eef-cd86b608fe48","arxiv_id":"2507.19948","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"UniCT Depth fuses event and image data with a CNN-Transformer hybrid to improve monocular depth estimation, achieving lower average errors on MVSEC and DENSE benchmarks.","lead":"This paper introduces UniCT Depth, a neural network that combines event camera data and regular images to estimate depth, especially in dark or high-contrast scenes. The method uses both convolutional and transformer components to fuse the two data types, and reports accuracy improvements over existing depth estimation methods on standard benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline SOTA claim rests on unverifiable baseline tables; Table 2's Transform-based row is internally inconsistent and no protocol/code is supplied.","rationale":"The Reader's weakest assumption — that the baseline numbers in Tables 1–3 are accurate and comparable — is also the single most load-bearing assumption for the paper's central claim. The claim is exclusively empirical, and the evidence consists of comparison tables with no code and no evaluation protocol. The anomalies the Reader flags are concrete and internal to the paper: the Transform-based[2024] row in Table 2 is implausibly low on δ metrics and missing RMSE log, and the SRFNet row in Table 3 is implausibly high relative to its MVSEC standing. These inconsistencies cannot be dismissed as mere style issues because the headline margin over SRFNet is computed from those very numbers. I do not see a separate, more fundamental flaw in the architecture or loss: the ablation study in Table 4 is internally coherent, and the reported improvements from adding CMSA+MFSA and DCC are directionally consistent with the design narrative. The weakness is external verification, not internal contradiction. The appropriate verdict is unchanged: CONDITIONAL, conditioned on releasing code and on re-running or explaining the baseline numbers. If the authors provide reproducible evaluation scripts and the corrected tables preserve the margins, the claim should be upgraded to ACCEPT; if the anomalous rows shift materially, the claim should be downgraded. I agree with the Reader rather than proposing a different concern, because this is the point where the strongest claim would actually break.","tokens_in":13554,"tokens_out":5112,"duration_ms":62213,"concrete_test":"Obtain the authors' code and evaluation configuration, or request it as a condition of release. Then independently run official implementations of SRFNet, IEBins, Transformer-based[2024], and the other Table 2/3 baselines using the same train/test split, same event voxelization (B=5, 224×224), same depth cutoffs, and same metric code as UniCT Depth. In particular, recompute Table 2's Transform-based row: if its δ<1.25 recovers to roughly 0.6 and RMSE log becomes non-missing, all claimed gains must be rederived; if it remains near 0.351, the comparison as printed is invalid. Also report mean ± std over at least 3 seeds for Ours and the top baselines on at least the day1/night1 sequences. The SOTA claim holds only if the baseline tables remain materially unchanged under this protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: UniCT Depth outperforms existing image-, event-, and fusion-based methods on key metrics. Every support for that claim is a comparison table (Tables 1–3), so the load-bearing condition is that the baseline numbers are produced under an equivalent, correct evaluation protocol. This condition is not currently secured. Table 2's 'Transform-based[2024]' row lists δ<1.25=0.351 and δ<1.25^3=0.480, far below every other row (e.g., IEBins δ<1.25=0.638), and omits RMSE log entirely, with no explanation for either anomaly. Table 3 gives SRFNet a 30 m Avg.Error of 6.11 and Abs.Rel of 0.51, more than four times the corresponding errors of Transformer-based on the same dataset; this is hard to reconcile with SRFNet being second-best on MVSEC unless different training or evaluation settings were used. The paper also reports no train/test split, no baseline hyperparameters, no error bars or seeds, and no code. Therefore the claimed margins — including the 5.56% Abs.Rel gain over SRFNet in Table 2 — may shrink, change rank, or vanish under a corrected protocol. This is an evidentiary gap, not an accusation of misconduct, but it is exactly the point at which the strongest claim needs independent support.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniCT Depth, a monocular depth estimation method that fuses event camera data with intensity images. The architecture is a U-Net-style encoder-decoder whose encoder uses CcViT-DA blocks, each containing a ViT Dual Self-Attention module (with a spatial-window CMSA branch and a channel-wise MFSA branch) plus a Detail Compensation Convolution (DCC) block. The method is evaluated on the MVSEC and simulated DENSE datasets against image-based, event-based, and fusion-based baselines using standard depth metrics. The authors report state-of-the-art or competitive results, particularly lower Average Error across cut-off distances, and provide ablations of the attention branches, the DCC block, and the input modalities.","tokens_in":13852,"tokens_out":2543,"duration_ms":32152,"significance":"If the reported results are reproducible and the baseline comparisons are fair, the paper makes a useful architectural contribution to event-image fusion for monocular depth estimation. The combination of windowed spatial attention, channel-wise cross-modal attention, and a convolutional detail-compensation path is plausible and well motivated. The ablation study (Table 4) and modality study (Table 5) give some support for the design choices, and the reported inference speed (25 FPS on an RTX 3090) indicates practical potential. However, the significance is conditional on verification: no code, no evaluation protocol details beyond basic training settings, no error bars or multi-seed runs, and at least one internally inconsistent baseline row (Table 2) currently undermine the claim that the method 'outperforms existing methods' on MVSEC and DENSE.","major_comments":[{"comment":"The Transform-based[2024] row in Table 2 is internally inconsistent: RMSE log is missing for both scenes, and the δ<1.25, δ<1.25², and δ<1.25³ values (0.351/0.437/0.480 on Outdoor day1) are far lower than every other method in the same table, including methods with worse Abs.Rel. This is not explained anywhere in the text. Because the headline claim of 'best in 10 out of 12 scores' depends on the correctness of all baseline rows, the authors must either correct this row, explain the evaluation protocol that produced it, or remove it and re-evaluate the comparison, ideally with code or a detailed protocol that makes the numbers reproducible.","section":"Table 2"},{"comment":"The experimental setup is under-specified. Section 4.1 states learning rate, optimizer, epochs, image size, and loss weights, but it does not specify the train/test sequence split for MVSEC, the number of runs or seeds, the baselines' training or evaluation settings (e.g., whether each baseline was re-trained under the same protocol or numbers are taken from prior papers), or the exact procedure for computing Avg.Error at cutoff distances. Tables 1-3 report single runs with no error bars. Consequently, the claimed margins, including the 5.56% and 7.16% Abs.Rel improvements over SRFNet in Section 4.2, cannot be independently verified or meaningfully compared. Please provide complete reproducibility details, including code, or report the results as preliminary.","section":"Section 4.1 / Tables 1-3"},{"comment":"The SRFNet row on the DENSE dataset is difficult to reconcile with the method's strong performance on MVSEC. For example, SRFNet's Avg.Error at 30m is 6.11 and Abs.Rel is 0.51, while the same method is second-best on MVSEC in Table 2 (Abs.Rel 0.234/0.335). Unless the DENSE evaluation used a different protocol (e.g., different training set, input resolution, or ground-truth preprocessing), these values appear implausibly poor. The authors should describe how each baseline was evaluated on DENSE and explain or correct these numbers; otherwise the '13.3% / 30% / 10.9%' improvement claims in Section 4.2 are not trustworthy.","section":"Table 3"},{"comment":"The Detail Compensation Conv (DCC) block is not fully specified. The text says the block uses global max/average pooling, concatenation, a convolution to produce a single channel, a sigmoid, a dot product, and then 'two convolutional layers and activation functions', but it does not give kernel sizes, number of channels, layer counts, or the roles of the intermediate 'weighted output' with shape H×W×1. Without these details, the block cannot be reproduced from the paper. Please provide the full layer-by-layer specification, either in the main text, an appendix, or accompanying code.","section":"Section 3.2 / Figure 2(d)"}],"minor_comments":[{"comment":"The text says 'best results in 10 out of 12 scores', but Table 2 has 10 score columns (five metrics for two scenes); the count should be clarified or the statement corrected.","section":"Section 4.2 / Table 2"},{"comment":"The notation list in Figure 2 includes P_w, N_w, N_h, C_h, and C_g but some of these are not defined in the body text; please define all symbols in the text or caption for clarity.","section":"Figure 2"},{"comment":"The abbreviation for the proposed block is inconsistently capitalized: 'CcViT-DA' in the title/abstract and 'CCViT-DA' in Section 3.2; please standardize.","section":"Section 3.2"},{"comment":"Minor language issues: 'enhances edge representations' in the abstract should be 'enhance', and Section 3.2 contains 'It relies on standard self-attention... which is susceptibility to gradient' — these need copy-editing.","section":"Abstract / Section 5"}],"recommendation":"major_revision","confidential_remarks":"The paper presents an interesting architecture, but the empirical claim is currently unverifiable due to missing protocol details, no code release, and at least one clearly anomalous baseline row. The issues in Tables 2 and 3 are not minor typographical problems; they concern the comparability of the central comparison. I believe the authors can address them within the scope of a revision by providing exact evaluation settings, re-running or correcting baseline evaluations, and ideally releasing code. If the anomalies turn out to be unreproducible or the baselines were misconfigured, the stated improvements may shrink, so the revision should be checked carefully against the corrected numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a straightforward application of DaViT's dual attention and a CBAM-style spatial attention to event-image fusion for monocular depth. The components are not new, and the authors say so for the attention blocks. What is new is the specific combination plus the DCC detail block, and the ablations show the combination buys a few points on MVSEC. That is a legitimate, modest contribution.\n\nThe paper does a few things well. It cites DaViT explicitly for CMSA and MFSA, which is honest. The ablation design is reasonable: it isolates the two attention branches, the DCC block, and the input modalities. The modality ablation (Table 5) is a nice sanity check. Qualitative examples suggest the method actually handles occlusion cases where one modality is missing. The FPS numbers are useful.\n\nNow the soft spots, in proportion. The central empirical claim is \"SOTA across key metrics,\" and every support is a comparison table. That is fine only if the baseline rows are produced under the same protocol. There is at least one internal inconsistency: Table 2's \"Transform-based[2024]\" row shows delta<1.25=0.351 and delta<1.25^3=0.480, far below every other row, with RMSE log missing and no note. Table 3 gives SRFNet an Abs.Rel of 0.51 and 30m Avg.Error of 6.11, hard to square with SRFNet being second-best on MVSEC; if that number is right, the method was run out of its training distribution or under different settings. The paper reports no train/test split, no baseline hyperparameters, no code, and no error bars or seeds. So the claimed margins, including the 5.56% Abs.Rel gain, could shift under a corrected protocol. This is an evidentiary gap, not misconduct.\n\nMinor issues: the DCC block description omits kernel sizes and layer counts, so it is not reproducible from the text. Table 4 suggests DCC alone slightly hurts (row 7 vs row 5) but helps when combined with the dual SA block; that interaction is not discussed, which is a missed chance to explain the mechanism. The gains over SRFNet on MVSEC mean Avg.Error are small (1.84 vs 1.99 at 20m), so the practical significance is modest.\n\nWho is this for? People working on event-image fusion and real-time depth for robotics. A serious referee should engage with it, but mostly to verify the baseline numbers. If the authors release code and correct the Table 2/Table 3 anomalies, the paper is publishable as an incremental but sound application. As it stands, the SOTA claim is not secured. My recommendation: send to peer review, but require protocol transparency, corrected baseline rows, and multi-seed statistics before acceptance.","headline":"Incremental but plausible fusion architecture; the SOTA claim rests on baseline tables with at least one internal inconsistency, so the verdict hinges on protocol transparency.","tokens_in":14393,"tokens_out":2087,"would_cite":false,"duration_ms":23520,"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":"UniCT Depth claims that fusing event streams and intensity images through a dual-branch self-attention encoder yields the lowest average depth error yet reported on the MVSEC and DENSE benchmarks.","keywords":["monocular depth estimation","event camera","image-event fusion","vision transformer","dual self-attention","detail compensation convolution","MVSEC dataset","DENSE dataset"],"falsifier":"Recompute the MVSEC and DENSE comparisons using the official releases of SRFNet and the 2024 transformer-based fusion method under the same training schedule; if the claimed 5.56% and 7.16% Abs.Rel improvements over SRFNet or the margin over the transformer baseline do not reproduce, the central claim fails. A simpler check is to correct the 'Transform-based[2024]' row in Table 2, whose $\\delta<1.25$ values of about 0.35 are far below every other method's roughly 0.6 and whose RMSE log is absent.","tokens_in":13378,"feed_emoji":"📏","tokens_out":9010,"duration_ms":89646,"temperature":0.7,"pith_summary":"The paper claims that monocular depth estimation in difficult conditions improves when event-camera streams and ordinary intensity images are fused inside a single CNN-Transformer encoder that models both spatial context and cross-modal channel dependencies. It presents UniCT Depth, whose core CcViT-DA blocks use two parallel attention branches — a windowed spatial self-attention (CMSA) and a channel-grouped modal-fusion self-attention (MFSA) — plus a Detail Compensation Convolution (DCC) block that restores local texture and edges. On the MVSEC and DENSE datasets, the method reports the lowest average absolute depth error among image-based, event-based, and fusion-based baselines, including a 5.56% and 7.16% relative improvement over SRFNet in Abs.Rel for day and night scenes. Ablations show that the dual attention branches together beat pure convolution, standard self-attention, and either branch alone, and that fusing both modalities outperforms using either one alone.","feed_headline":"Two-branch attention fusion cuts depth error on event-image data","feed_subtitle":"The model beats image-, event-, and fusion-based baselines on MVSEC and DENSE.","key_machinery":"The load-bearing component is the CcViT-DA Block, a CNN-Transformer module placed in the encoder's four downsampling stages. It contains Context Modeling Self-Attention (CMSA), which runs multi-head self-attention within non-overlapping spatial windows to capture local context for occluded regions, and Modal Fusion Self-Attention (MFSA), which applies single-head self-attention to a transposed token so that attention runs along the channel dimension in groups, modeling global dependencies and cross-modal correlation. The outputs of the two branches are dot-multiplied with the output of the Detail Compensation Convolution (DCC) block, a spatial-attention map built from channel-wise max and average pooling followed by sigmoid and convolutions, then concatenated and merged. These choices cut the attention complexity from $\\mathcal{O}(P^2 d)$ to $\\mathcal{O}(P P_w d) + \\mathcal{O}(P C C_g)$, which is what lets the dual attention run at interactive speeds.","core_discovery":"On its own terms, the paper discovers that a unified encoder with two complementary self-attention branches — one computing attention inside spatial windows and one computing attention across channel groups — fuses event and image features more effectively than prior single-branch or concatenation-based fusion, while a convolution-based detail compensation block sharpens the edges and textures that depth maps need in occluded or low-light scenes. The evidence is the set of benchmark numbers: the lowest mean Avg.Error at every cutoff distance on both MVSEC and DENSE, the best score in 10 of 12 metrics in Table 2, and the ablations where the full CcViT-DA block beats every partial configuration. The authors would summarize this as a demonstration that complementary spatial-window and channel-group attention, together with local convolution compensation, is a better way to combine event and image modalities than the alternatives compared.","pith_inferences":["Because the dual attention block is defined on generic feature maps rather than on any event-specific representation, the same encoder design could be dropped into other cross-modal dense prediction tasks (for example, depth from image plus LiDAR, or low-light semantic segmentation); the paper does not test this.","The windowed and channel-grouped attention reduces the cost of global self-attention, so the architecture should scale more gracefully to higher input resolutions than a standard transformer fusion encoder; this could be checked by evaluating at 512x512 or 1024x1024.","The implausibly low $\\delta<1.25$ values and the missing RMSE log for the 'Transform-based[2024]' row in Table 2 suggest the comparison against that baseline should be re-audited with official code before the claimed margin is taken at face value.","The single-modality ablations in Table 5 hint that the network could operate in a graceful-degradation mode if one modality is missing at test time, but the paper does not evaluate that scenario."],"forward_implications":["On both MVSEC and DENSE, the fused dual-modality model produces lower average depth error at 10m, 20m, and 30m cutoffs than any image-only, event-only, or fusion baseline reported.","The 5.56% (day) and 7.16% (night) relative Abs.Rel gains over SRFNet indicate that channel-wise attention adds information that spatial-prior-only fusion misses.","The ablation comparing row (8) with (6) and (7) shows the DCC block lowers mean error at all cutoffs, meaning local detail compensation is not redundant with either attention branch.","Using both modalities as input beats either modality alone in day and night scenes, confirming the fusion premise the architecture is built on.","The full model runs at about 25 FPS on an RTX 3090 while achieving these results, so the accuracy improvement does not sacrifice real-time operation."],"supporting_citations":[{"why":"Supplies the dual self-attention design (spatial windowed and channel-grouped attention) that CMSA and MFSA adapt for cross-modal fusion.","marker":"[Ding et al., 2022]"},{"why":"The SRFNet baseline that uses spatial-prior interactive fusion; the paper's main comparison target, which UniCT Depth claims to beat by 5.56% and 7.16% Abs.Rel.","marker":"[Pan et al., 2024]"},{"why":"The 'Transform-based[2024]' baseline using a single unified Transformer encoder; the closest prior fusion approach the paper argues is computationally heavy and coarsely fused.","marker":"[Devulapally et al., 2024]"},{"why":"Provides the spatio-temporal voxel grid event representation the network uses to encode the asynchronous event stream.","marker":"[Zhu et al., 2019]"},{"why":"The U-Net encoder-decoder with skip connections that forms the backbone of UniCT Depth.","marker":"[Ronneberger et al., 2015]"},{"why":"The MVSEC dataset used for the main real-world day/night evaluation.","marker":"[Zhu et al., 2018]"},{"why":"The DENSE simulated dataset used to evaluate generalization.","marker":"[Hidalgo-Carrió et al., 2020]"},{"why":"DPT baseline representing image-only transformer depth estimation; also the comparison that shows image-only methods lose objects in low light.","marker":"[Ranftl et al., 2021]"}],"fun_headline_variants":["UniCT Depth: dual attention fusion sharpens event-image depth maps","Two attention branches beat baselines on event-image depth","Convolution-compensated transformer wins on event-image depth","Dual SA block with detail convolution lifts event-image depth","Attention plus convolution fusion tops event-image depth benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported advantage depends on the baseline numbers in Tables 1-3 being accurate and comparably configured, since the paper provides no code or evaluation protocol and one baseline row in Table 2 shows missing and implausible values.","fun_headline_variants_meta":{"raw":{"variants":["UniCT Depth: dual attention fusion sharpens event-image depth maps","Two attention branches beat baselines on event-image depth","Convolution-compensated transformer wins on event-image depth","Dual SA block with detail convolution lifts event-image depth","Attention plus convolution fusion tops event-image depth benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1417,"prompt_tokens":932,"completion_tokens":485,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":404}},"tokens_in":548,"tokens_out":485,"duration_ms":5785,"temperature":1.0,"reasoning_tokens":404,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:50:16.622236+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the MVSEC and DENSE comparisons using the official releases of SRFNet and the 2024 transformer-based fusion method under the same training schedule; if the claimed 5.56% and 7.16% Abs.Rel improvements over SRFNet or the margin over the transformer baseline do not reproduce, the central claim fails. A simpler check is to correct the 'Transform-based[2024]' row in Table 2, whose $\\delta<1.25$ values of about 0.35 are far below every other method's roughly 0.6 and whose RMSE log is absent.","supporting_citations":[{"cited_title":"Davit: Dual attention vision transformers","cited_arxiv_id":null,"evidence_quote":"Supplies the dual self-attention design (spatial windowed and channel-grouped attention) that CMSA and MFSA adapt for cross-modal fusion."},{"cited_title":"Srfnet: Monocular depth estimation with fine-grained structure via spatial reliability-oriented fusion of frames and events","cited_arxiv_id":null,"evidence_quote":"The SRFNet baseline that uses spatial-prior interactive fusion; the paper's main comparison target, which UniCT Depth claims to beat by 5.56% and 7.16% Abs.Rel."},{"cited_title":"Multi-modal fusion of event and rgb for monocular depth estimation using a unified transformer-based architecture","cited_arxiv_id":null,"evidence_quote":"The 'Transform-based[2024]' baseline using a single unified Transformer encoder; the closest prior fusion approach the paper argues is computationally heavy and coarsely fused."},{"cited_title":"Unsupervised event-based learning of optical flow, depth, and egomotion","cited_arxiv_id":null,"evidence_quote":"Provides the spatio-temporal voxel grid event representation the network uses to encode the asynchronous event stream."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation","cited_arxiv_id":null,"evidence_quote":"The U-Net encoder-decoder with skip connections that forms the backbone of UniCT Depth."},{"cited_title":"The multivehicle stereo event camera dataset: An event camera dataset for 3d perception","cited_arxiv_id":null,"evidence_quote":"The MVSEC dataset used for the main real-world day/night evaluation."},{"cited_title":"Learning monocular dense depth from events","cited_arxiv_id":null,"evidence_quote":"The DENSE simulated dataset used to evaluate generalization."},{"cited_title":"Vision transformers for dense prediction","cited_arxiv_id":null,"evidence_quote":"DPT baseline representing image-only transformer depth estimation; also the comparison that shows image-only methods lose objects in low light."}],"review_version":1}