{"id":"c618bee0-7cd8-46cd-8803-5194d0305d95","arxiv_id":"2412.19964","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A proposed Mamba-based two-branch fusion network for depth estimation reports competitive KITTI and DDAD numbers, but the method and its promised noisy-pose benchmark are not actually specified.","lead":"The paper describes a depth estimation network that fuses single-view and multi-view predictions with a Mamba-based backbone and an attention-based fusion module, reporting low error rates on KITTI and DDAD. It also announces a new robustness benchmark for noisy camera poses, but the benchmark, the fusion equations, and the training details are never actually given.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central noisy-pose robustness claim has no supporting experiment: the announced benchmark is never defined or run, and Tables 1-3 all report clean-pose results.","rationale":"The reader's REJECT verdict is justified, and my stress-test reaches the same conclusion through a slightly different route. The reader's weakest_assumption focuses on the unrecoverable architecture; that is real and serious. However, the more load-bearing failure is empirical: the paper's central contribution, stated in the abstract and introduction, is robustness under noisy camera poses and a new benchmark for it. That benchmark and all noisy-pose evaluations are absent. Tables 1-3 are clean-pose results; no noise model, perturbation range, or protocol is given. Therefore the main claim cannot be verified even in principle from the submitted manuscript. I did not find an internal contradiction in the clean-pose numbers, though Table 1 shows Ours is worse than AFNet on DDAD AbsRel (0.107 vs 0.088) while better on RMSE, which weakens the unqualified 'competitive performance' phrasing but is not the central issue. The prose-only fusion description compounds the problem: Section 3.2 describes group-wise correlation, variance volume, and multiplication by attention weights without equations, so no independent reimplementation can be attempted from the paper. The claimed 2.8% KITTI RMSE improvement and 15.3% DDAD RMSE improvement are likewise not accompanied by code or checkpoints. In sum, the argument does not support the central claim, so no adjustment to the reader's verdict is needed.","tokens_in":7736,"tokens_out":4374,"duration_ms":44074,"concrete_test":"Reconstruct and run the missing noisy-pose protocol: take the KITTI Eigen and DDAD validation splits, add controlled Gaussian noise to source camera poses (e.g. translation sigma in {0.01, 0.05, 0.1} m, rotation sigma in {0.1, 0.5, 1.0} deg), and evaluate Ours, AFNet, MVS2D, and a single-view baseline such as Monodepth2 with identical training. Report AbsRel and RMSE versus noise level. If the protocol cannot be recovered from the paper, or if at moderate noise Ours does not degrade less than the single-view baseline and more gracefully than AFNet, the robustness claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's stated novelty is robustness to imperfect camera poses: the abstract announces a new robustness benchmark and claims current methods fail under noisy poses, and the method is motivated as adaptively selecting the robust branch. Yet no noisy-pose experiment appears anywhere. The benchmark is not defined: no pose perturbation model, no noise magnitudes, no split, no protocol. Table 1, the main comparison, uses standard KITTI/DDAD with given clean poses; the ablations in Tables 2 and 3 likewise report only clean KITTI metrics. The sentence 'we find current methods will fail when given noisy pose settings' is asserted without a result, and the proposed two-branch fusion is never compared with any baseline under pose noise. This is the most load-bearing gap because it is the paper's motivating and differentiating claim: even a fully specified architecture and reproducible clean-pose numbers would not establish robustness. The prose-only description of the fusion module (Section 3.2) and the unrun benchmark together make the central claim untestable as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-branch depth estimation network for fusing single-view and multi-view cues, built around a Depth-Mamba feature extraction backbone and an attention-based adaptive fusion module. The authors claim that existing multi-view and fusion methods fail under noisy camera poses, introduce a new robustness benchmark for this setting, and report competitive results on KITTI and DDAD, including an AbsRel of 0.042 on KITTI and an RMSE of 6.12 on DDAD. Ablation studies on the backbone and fusion module are also reported. The central noisy-pose robustness claim, however, is never experimentally supported, and the method itself is described only at a high level without mathematical or architectural specificity.","tokens_in":7963,"tokens_out":4649,"duration_ms":48802,"significance":"If the claims were substantiated, robustness to imperfect camera poses would be a practically important contribution for autonomous driving, and the use of a mamba-based backbone with adaptive single-view/multi-view fusion could be of interest to the multi-view depth estimation community. The paper also deserves credit for reporting evaluations on two public benchmarks and for including ablations of the backbone and fusion choices. As written, however, the announced robustness benchmark is not defined or executed, the architecture is described only in prose, and the main comparison table has unverifiable baseline citations. These issues currently prevent the results from being assessed or reproduced, so the paper's scientific contribution is not yet established.","major_comments":[{"comment":"The central claim of the paper is robustness to noisy camera poses, but no noisy-pose experiment appears anywhere. The abstract announces a new robustness benchmark and states that current methods fail under noisy poses, and Section 4.4 repeats that current methods fail, yet there is no definition of the benchmark, no noise model, no perturbation magnitudes, no evaluation protocol, and no table or figure with noisy-pose results. Tables 1-3 all report clean-pose results on the standard KITTI and DDAD splits. Because the proposed adaptive fusion is motivated specifically as selecting the more robust branch under pose noise, the paper's main differentiating claim is untestable as submitted.","section":"Abstract and Section 4.4"},{"comment":"The proposed method is described only in prose. Section 3.2 states that a group-wise correlation volume and a variance volume are constructed and that the variance volume is multiplied by attention weights, but no equations define these volumes, the attention module, the 3D hourglass network, or the depth regression. Section 3.1 does not give layer counts, channels, resolutions, or the structure of the Depth-Mamba and Pyramid Mamba modules, and Section 3.4 provides only the final MAE loss without specifying whether branch-level losses are applied to the single-view and multi-view outputs. Without these details, the link between the described architecture and the numbers in Table 1 cannot be verified, and the implementation is not reproducible.","section":"Section 3"},{"comment":"The main comparison is compromised by inconsistent sourcing. The text calls AFNet the current state-of-the-art and cites it as [2], but reference [2] is AdaBins, and no reference for AFNet appears in the bibliography. The asterisks attached to many entries in Table 1 are never explained, making it unclear which results were produced by the authors and which were taken from other papers. In addition, Section 4.5.1 says the backbone ablation compares ConvNeXt-T, VMamba, and Depth-Mamba, but Table 2 lists only Base-Mamba, Base-VMamba, and Base-DepthMamba, with no ConvNeXt-T row. These issues prevent the reader from validating either the state-of-the-art comparison or the claimed ablations.","section":"Table 1 and references"},{"comment":"The DDAD performance claim is stated selectively. The text reports an RMSE improvement of 15.3% over AFNet and MaGNet, but in Table 1 the proposed method has DDAD AbsRel 0.107, which is worse than AFNet's 0.088, and SqRel 1.49, which is also worse than AFNet's 1.41. The 15.3% figure applies only to RMSE relative to AFNet, not to MaGNet, where the relative RMSE improvement is about 33.7%. The overall claim of competitive performance on DDAD therefore depends on a single metric and omits the fact that the method is worse on two of the three reported error measures.","section":"Section 4.4 and Table 1"}],"minor_comments":[{"comment":"Reference [10] is cited for the DDAD dataset, but [10] is the Monodepth2 paper by Godard et al.; the DDAD dataset should be cited to its original source (Guizilini et al., 2020).","section":"Section 4.1"},{"comment":"The notation in the loss function is inconsistent with the metrics: Eq. (1) uses y_i and y_i^hat, while Eqs. (2)-(4) use y_i and y_i^pred; please unify the notation.","section":"Equation (1)"},{"comment":"The text 'Building on this, [MVSNet] pioneered...' uses an unnumbered citation placeholder; MVSNet should be given a numbered reference consistent with the bibliography.","section":"Related Works"},{"comment":"The implementation details omit several settings needed for reproducibility, including the depth range, the source-view selection and baseline distances, the input resolution for training and testing, and the number of training samples used from each camera in DDAD.","section":"Section 4.3"},{"comment":"The phrase 'Ablation studies prove the effectiveness' overstates what a single-dataset ablation can establish; 'support' or 'demonstrate' would be more appropriate.","section":"Section 4.5"},{"comment":"There is a typo in the Conclusions section: 'The ablation studies certificate the effectiveness' should be 'certify the effectiveness' or 'demonstrate the effectiveness.'","section":"Conclusions"}],"recommendation":"reject","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper's headline claim—robustness to noisy poses—is never tested. The abstract announces a new robustness benchmark, but all reported experiments use clean-pose KITTI and DDAD. That mismatch makes the central contribution unverifiable.\n\nWhat's genuinely there: the two-branch architecture (single-view + multi-view) with adaptive attention fusion is a reasonable idea, and the clean-pose numbers are competitive: AbsRel 0.042 on KITTI and RMSE 6.12 on DDAD. The DDAD RMSE is a real improvement over AFNet if the numbers are correct. Using a Mamba-based backbone is a plausible extension of VMamba to depth estimation.\n\nThe soft spots are serious. The method description has no equations—the attention volume, variance volume, and depth regression are all prose. That prevents any form of independent implementation. The announced noisy-pose benchmark never appears: no perturbation model, no noise levels, no protocol, no results. The comparison table has citation errors: AFNet is cited as [2], but [2] is AdaBins; MaGNet is misspelled and there is a stray reference. The ablation tables also only report clean-pose metrics, so the claimed robustness is entirely unsupported.\n\nThere is also no code or data release, which—combined with the missing equations—means the reported numbers cannot be checked. The writing is sloppy throughout (e.g., 'Muti-scale,' 'certificate' as a verb, inconsistent citing of the same works).\n\nThe core idea might have legs, but this manuscript is far too incomplete to act as a basis for evaluation. The authors need to define and run the robustness benchmark, add the architectural equations, and fix the references. If they do that, the clean-pose numbers could make a decent workshop or conference paper.\n\nRecommendation: desk reject. Not ready for peer review until the benchmark and equations are added. I would not cite this in its current form, and I would not bring it to a reading group.","headline":"Announces a noisy-pose robustness benchmark that never appears in the experiments; the clean-pose numbers are competitive but unverifiable.","tokens_in":8475,"tokens_out":4040,"would_cite":false,"duration_ms":39016,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that fusing single-view and multi-view depth estimates with an attention-based adaptive mechanism and a shared mamba backbone produces depth maps that stay accurate even when camera poses are noisy, reporting AbsRel 0.042…","keywords":["multi-view depth estimation","monocular depth estimation","single-view and multi-view fusion","adaptive attention fusion","Mamba backbone","noisy camera poses","robustness benchmark","autonomous driving"],"falsifier":"Train the exact described architecture from scratch on the KITTI Eigen split with the stated recipe (40 epochs, AdamW, one-cycle schedule, 128 depth hypotheses, three input frames) and check whether it reaches AbsRel 0.042 and RMSE 1.695 on KITTI and RMSE 6.12 on DDAD; if the published recipe cannot reproduce these numbers, or if swapping the attention-weighted fusion for concatenation does not produce the claimed 33.3% AbsRel gap, the central claim fails.","tokens_in":7541,"feed_emoji":"🚗","tokens_out":6586,"duration_ms":55957,"temperature":0.7,"pith_summary":"This paper claims that multi-view depth estimation can be made robust to inaccurate camera poses by fusing a single-view depth branch with a multi-view depth branch instead of trusting either one alone. The proposed two-branch network shares a mamba-based feature extraction backbone across both branches, builds a group-wise correlation volume and a variance volume from the extracted features, and uses an attention-based adaptive fusion module to weight the two branches per region. On the KITTI Eigen split the method reports an AbsRel of 0.042 and RMSE of 1.695, and on DDAD an RMSE of 6.12, improving RMSE by 2.8% over AFNet on KITTI and by 15.3% on DDAD. The authors also introduce a robustness benchmark with noisy pose settings and report that existing single-view, multi-view, and fusion methods degrade sharply under it.","feed_headline":"Fused depth network cuts DDAD RMSE by 15.3% over AFNet","feed_subtitle":"Two-branch design with Mamba backbone and attention weighting keeps depth accurate when camera poses are noisy.","key_machinery":"The load-bearing mechanism is the adaptive fusion of two cost volumes: the multi-view branch produces a group-wise correlation volume from reference and source features, while the variance volume is constructed from the same features; a multi-scale attention module followed by a 3D hourglass network turns the correlation volume into attention weights, and the element-wise multiplication of variance volume with those weights yields the attention volume used for depth regression. The feature extraction backbone is a \"depth-mamba\" variant of the Mamba selective state-space sequence model adapted to image feature extraction, augmented with local feature blocks to compensate for mamba's weaker spatial localization. This mechanism is what lets the network choose the more reliable branch per location.","core_discovery":"The central claim is that the reliability of depth estimates can be improved by adaptively selecting between single-view and multi-view evidence rather than merging them blindly. The paper proposes a two-branch architecture in which a shared depth-mamba backbone extracts multi-scale features from the reference and source images; a multi-view branch builds a group-wise correlation volume, a single-view branch builds a variance volume, and a multi-scale attention module plus a 3D hourglass network produces attention weights that are multiplied onto the variance volume to form the final attention volume. Depth is then obtained by residual regression from that volume. On this design the paper reports AbsRel 0.042 on KITTI and RMSE 6.12 on DDAD, and it attributes the gain to the mamba backbone and to the attention-based fusion, with ablations showing the fusion module lowers AbsRel by 33.3% versus simple concatenation and by 23.6% versus cross-attention. The paper further claims these design choices keep the system accurate under noisy camera poses, a setting where current methods fail.","pith_inferences":["A natural next test is to visualize the learned attention weights directly: if the fusion is truly adaptive, the weights should be low in occluded or textureless regions where multi-view matching is unreliable.","The robustness benchmark could be extended to perturb camera intrinsics as well as extrinsics, probing sensitivity to calibration errors beyond pose noise.","Because the architecture description lacks implementation specifics, independent replication would currently require reverse-engineering the described modules; releasing code and exact layer configurations would let the community verify and build on the claimed improvements.","The adaptive-fusion idea may transfer to fusing learned depth with LiDAR or radar depth, where the same tension exists between dense monocular cues and sparse but accurate geometric measurements."],"forward_implications":["Multi-view depth systems could be deployed in autonomous driving without relying on highly accurate pose estimates, since the adaptive fusion keeps the single-view branch as a fallback where geometry is unreliable.","The same two-branch fusion principle can be applied to other multi-modal estimation tasks where one modality is robust and another is precise but brittle.","The reported mamba backbone with local feature blocks offers a practical alternative to CNN and transformer backbones for stereo and multi-view depth networks.","The noisy-pose robustness benchmark introduced in the paper gives a standard way to compare depth estimators under realistic pose errors rather than only ideal poses.","If the reported numbers hold, the method sets a new accuracy bar on the KITTI Eigen split with AbsRel 0.042 and on DDAD with RMSE 6.12."],"supporting_citations":[{"why":"MaGNet, the single-view-plus-multi-view fusion baseline that the proposed method compares against and improves on DDAD by 15.3% RMSE.","marker":"[1]"},{"why":"AdaBins, the adaptive-bin monocular baseline the paper calls AFNet, whose RMSE the proposed method improves by 2.8% on KITTI.","marker":"[2]"},{"why":"Provides the KITTI Eigen split used for training and evaluation.","marker":"[6]"},{"why":"Provides the DDAD dataset and its train/validation split used for training and evaluation.","marker":"[10]"},{"why":"MVS2D, the attention-driven 2D-convolution fusion method whose residual weakness under noisy poses motivates the proposed adaptive fusion.","marker":"[21]"},{"why":"Defines the Mamba selective state-space model that the proposed depth-mamba backbone is built from.","marker":"[25]"},{"why":"Defines VMamba and the multi-scale visual state-space feature extraction that the backbone adapts with local feature blocks.","marker":"[26]"}],"fun_headline_variants":["Mamba backbone plus adaptive fusion beat noisy-pose depth estimation","Adaptive fusion keeps depth accurate when camera poses are off","Two-branch depth net with attention fusion tackles noisy poses","DepthMamba fuses views adaptively for robust depth under pose noise","Attention-based fusion boosts depth net's noise resilience"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire experimental result rests on the assumption that the network described only in prose—shared depth-mamba backbone, group-wise correlation volume, variance volume, 3D hourglass, and attention-weighted multiplication—was implemented exactly as described and trained with the stated recipe, because no equations, layer dimensions, or branch-specific losses are provided to pin down the architecture.","fun_headline_variants_meta":{"raw":{"variants":["Mamba backbone plus adaptive fusion beat noisy-pose depth estimation","Adaptive fusion keeps depth accurate when camera poses are off","Two-branch depth net with attention fusion tackles noisy poses","DepthMamba fuses views adaptively for robust depth under pose noise","Attention-based fusion boosts depth net's noise resilience"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00044,"raw_usage":{"total_tokens":2230,"prompt_tokens":944,"completion_tokens":1286,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":1203}},"tokens_in":560,"tokens_out":1286,"duration_ms":9936,"temperature":1.0,"reasoning_tokens":1203,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:44:08.541267+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the exact described architecture from scratch on the KITTI Eigen split with the stated recipe (40 epochs, AdamW, one-cycle schedule, 128 depth hypotheses, three input frames) and check whether it reaches AbsRel 0.042 and RMSE 1.695 on KITTI and RMSE 6.12 on DDAD; if the published recipe cannot reproduce these numbers, or if swapping the attention-weighted fusion for concatenation does not produce the claimed 33.3% AbsRel gap, the central claim fails.","supporting_citations":[{"cited_title":"Multiview depth estimation by fusing single-view depth probability with multi-view geometry","cited_arxiv_id":null,"evidence_quote":"MaGNet, the single-view-plus-multi-view fusion baseline that the proposed method compares against and improves on DDAD by 15.3% RMSE."},{"cited_title":"Adabins: Depth estimation using adaptive bins","cited_arxiv_id":null,"evidence_quote":"AdaBins, the adaptive-bin monocular baseline the paper calls AFNet, whose RMSE the proposed method improves by 2.8% on KITTI."},{"cited_title":"Depth anything: Unleashing the power of large-scale unlabeled data","cited_arxiv_id":null,"evidence_quote":"Provides the KITTI Eigen split used for training and evaluation."},{"cited_title":"Deep ordinal regression network for monocular depth estimation","cited_arxiv_id":null,"evidence_quote":"Provides the DDAD dataset and its train/validation split used for training and evaluation."},{"cited_title":"Patchmatchnet: Learned multi-view patchmatch stereo","cited_arxiv_id":null,"evidence_quote":"MVS2D, the attention-driven 2D-convolution fusion method whose residual weakness under noisy poses motivates the proposed adaptive fusion."},{"cited_title":"Fast -mvsnet: Sparse-todense multi-view stereo with learned propagation and gaussnewton refinement","cited_arxiv_id":null,"evidence_quote":"Defines the Mamba selective state-space model that the proposed depth-mamba backbone is built from."},{"cited_title":"Computing the stereo matching cost with a convolutional neural network","cited_arxiv_id":null,"evidence_quote":"Defines VMamba and the multi-scale visual state-space feature extraction that the backbone adapts with local feature blocks."}],"review_version":1}