{"id":"692e6906-6ced-4941-82a1-64867563d6c7","arxiv_id":"2412.20082","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"MambaVO improves deep visual odometry by adding Mamba-based matching refinement and a smoothed training objective, achieving state-of-the-art absolute trajectory error on EuRoC, TUM-RGBD, KITTI, and TartanAir.","lead":"MambaVO is a visual odometry system that uses a Mamba-based sequence model to refine image matching between frames before optimizing camera poses. It reports accuracy gains over prior deep visual odometry methods on standard benchmarks while running in real time.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed KITTI SOTA may be inflated by Metric3D v2's pretrained depth prior, which was trained on KITTI; the paper neither discloses this nor ablates it, so the central monocular comparison is not yet verified.","rationale":"Reading the paper in good faith, the proposed system is coherent: GIM initializes a Point-Frame Graph, GMM refines matches with Mamba blocks, BA optimizes poses, and TAP stabilizes training. The experiments are extensive and the ablations in Table 7 show each module matters. The reader's weakest assumption identifies the initialization as load-bearing, and Table 7 supports that strongly. My stress test goes one step further: the initialization is not trained on TartanAir alone. Sec. 3.1.1 uses Metric3D v2, a pretrained metric-depth foundation model, and Sec. 5 only says 'our network' is trained on TartanAir. Because Metric3D v2's training mix includes KITTI, the KITTI comparison is potentially contaminated: MambaVO receives a KITTI-trained metric depth and scale prior that DPVO and DROID-VO, trained only on TartanAir, do not receive. The effect is plausibly large because monocular ATE in meters depends on scale, and Table 7 shows removing the geometric initialization destroys KITTI accuracy (192.75/201.16 vs 50.21). This does not mean the authors acted improperly; it means the paper lacks a critical control. If the check I propose shows the KITTI advantage disappears, the central claim of SOTA on KITTI must be reframed as a property of the pretrained depth prior, not the proposed modules. If the advantage persists, the conditional acceptance stands as reasonable. Since the reader's verdict was already CONDITIONAL and my concern adds a specific verification condition without changing the overall disposition, I recommend no change to the verdict.","tokens_in":15821,"tokens_out":8953,"duration_ms":99132,"concrete_test":"Inspect the training-set list in Metric3D v2 [22]; if KITTI is included, re-run MambaVO on KITTI under two conditions: (a) replace Metric3D with a depth network trained only on TartanAir, and (b) evaluate with Sim(3) scale alignment so no metric depth prior is needed. If the KITTI ATE advantage over DPVO (50.21 vs 53.03 w/o loop) disappears or reverses, the SOTA claim is attributable to the pretrained depth prior and must be reported as such.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of SOTA ATE on KITTI (Table 3: 50.21 m w/o loop, 21.84 m w/ loop) rests on a component whose training data are not disclosed. In Sec. 3.1.1, Eq. (2), the Geometric Initialization Module obtains metric depth from Metric3D v2 [22] before PnP. The paper states in Sec. 5 that the model is trained only on TartanAir, but Metric3D v2 is a pretrained foundation model whose documented multi-dataset training mix includes KITTI. Monocular VO has scale ambiguity, and Metric3D provides absolute metric scale and strong geometric priors specifically on KITTI, while DPVO and DROID-VO are trained only on TartanAir and receive no such KITTI-trained prior. Table 7 confirms the load-bearing role of this initialization: removing PnP raises KITTI ATE to 201.16 m and removing the geometric feature raises it to 192.75 m, while the full system is 50.21 m. Thus the outdoor accuracy gains could be almost entirely attributable to the pretrained geometric depth prior, not to the Mamba refinement or BA. The paper neither ablates replacing Metric3D with a TartanAir-only depth backbone nor reports which datasets the pretrained components were trained on. This is a missing disclosure and a missing ablation, not an accusation of misconduct; it blocks verification of the claimed KITTI SOTA and weakens the assertion that the comparison is a fair monocular setting.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MambaVO, a monocular learning-to-optimize visual odometry system, and MambaVO++, its loop-closure variant. The pipeline initializes matches with EfficientLoFTR and PnP using Metric3D depth, extracts DINOv2 context features, refines and reweights matches with a Mamba-based Geometric Mamba Module, and optimizes poses and map points with differentiable bundle adjustment. A Trending-Aware Penalty is introduced to stabilize the nested training loss. Experiments on TartanAir, EuRoC, KITTI, and TUM-RGBD report state-of-the-art ATE, with ablations, a matching-AUC diagnostic, runtime, and GPU-memory comparisons.","tokens_in":16143,"tokens_out":5712,"duration_ms":58117,"significance":"If the accuracy claims hold, this is a solid empirical contribution to deep visual odometry: it is a plausible first use of state-space models for VO, the semi-dense geometric initialization is well motivated, the matching-AUC evaluation usefully isolates the matching improvement, and the ablation study in Table 7 supports the role of each proposed module. The real-time operation and lower GPU memory relative to DROID-VO, DPVO, and V2V are practically relevant strengths. However, the central KITTI SOTA claim is not yet verified because the geometric initialization relies on a pretrained metric-depth model with possible KITTI training overlap, and the empirical claims lack uncertainty and reproducibility details.","major_comments":[{"comment":"The claimed KITTI SOTA is not yet verified because the Geometric Initialization Module obtains metric depth from Metric3D v2 before PnP, and Metric3D v2's documented multi-dataset training includes KITTI; the statement in Sec. 5 that the model is trained only on TartanAir does not cover this pretrained component. Table 7 shows that this initialization is load-bearing: removing PnP raises KITTI ATE from 50.21 m to 201.16 m, and removing the geometric feature raises it to 192.75 m. Since DROID-VO, DPVO, and V2V do not receive a KITTI-trained metric-depth prior, the monocular comparison in Table 3 may conflate the pretrained depth prior with the proposed modules. Please disclose the training data of all pretrained components and add an ablation that replaces Metric3D with a depth backbone not trained on KITTI, or trained only on TartanAir, while keeping the rest of the pipeline fixed.","section":"Sec. 3.1.1, Eq. (2); Sec. 5, Table 3"},{"comment":"The paper reports mean ATE over five runs but no standard deviation, per-seed values, or significance testing, and no code is released. Several SOTA margins are small (e.g., TUM-RGBD AVG 0.076 vs. 0.089 for DPVO; EuRoC AVG 0.094 vs. 0.117 for V2V), so without variance information the central accuracy claim is difficult to assess. Please report run-to-run variability (std or min-max), specify the five seeds, and release code or detailed hyperparameters, including the number of Mamba blocks B and the number of initial matches k, which are used throughout Sec. 3 but never given values.","section":"Sec. 5.1, Tables 1-4"},{"comment":"The matching-AUC comparison needs a more precise protocol. It is not stated how the matching outputs of DROID-VO, DPVO, and V2V are extracted, whether the same reference keyframes, the same number of matches, and the same inlier thresholds are used, and whether MambaVO's Poselib evaluation is applied identically to all methods. Without this control, the reported AUC@1 degree improvement (0.471 vs. 0.399 for V2V) may reflect evaluation choices rather than matching quality.","section":"Sec. 5.2, Table 5"}],"minor_comments":[{"comment":"There are several typos: 'Divo-v2' should be 'DINOv2', 'mathing features' should be 'matching features', and 'ARE' in the Fig. 5 caption should be 'ATE'.","section":"Sec. 3.1.2, Sec. 3.2.1, Fig. 5 caption"},{"comment":"The notation for the matching loss is ambiguous: Eq. (12) defines L_match, then Eq. (13) redefines L_match on the left while using L_match on the right. Please introduce a distinct symbol for the gradient-weighted matching loss, such as L_match_weighted.","section":"Eqs. (12)-(13)"},{"comment":"The row label 'w.o. History balance' is inconsistent with the text's 'trend-based balance parameter' in Sec. 3.3.2; the terminology should be aligned.","section":"Table 7"},{"comment":"The runtime and memory comparison is underspecified: please state the GPU model, input resolution, sequence length, and whether the reported GPU memory is peak or average usage, so that the comparison with DROID-VO, DPVO, and V2V is reproducible.","section":"Sec. 5.4, Fig. 6"},{"comment":"Hyperparameters B (number of Mamba blocks) and k (number of initial matches) are never assigned values; please report them, together with the other training and inference settings, in the experimental section.","section":"Sec. 3.2 and Sec. 5"}],"recommendation":"major_revision","confidential_remarks":"The Metric3D training-data overlap is the main review risk. I do not interpret it as misconduct, but it must be resolved before the KITTI result can be accepted. If the ablation requested in Major Comment 1 shows that MambaVO still outperforms DPVO and V2V with a KITTI-free depth prior, I would support acceptance after the uncertainty and reproducibility items in Major Comment 2 are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a serious, well-engineered VO paper, and the first to use a Mamba SSM for inter-frame matching refinement in a learning-to-optimize pipeline. The core idea is sensible: semi-dense matches from EfficientLoFTR plus PnP give a strong initial pose, the GMM (history fusion + Mamba blocks) refines matching and weights, and differentiable BA does the final optimization. The TAP training smoothing is a small but reasonable trick. The experiments are extensive — four benchmarks, ablations for every module, and a matching AUC diagnostic that isolates the matching quality gain. On EuRoC, TUM-RGBD, and TartanAir, the reported ATE improvements over DPVO and V2V are plausible and the ablations back up the role of each piece.\n\nBut the KITTI claim has a problem. The geometric initialization uses Metric3D v2 for metric depth before PnP, and Metric3D v2 was trained on a large dataset mix that includes KITTI. The paper says the model is trained only on TartanAir, which is true for the VO network itself, but the depth backbone leaks KITTI-specific scale and geometry into the system. DPVO and V2V don't have that. Table 7 shows the depth/PnP initialization is load-bearing — removing it degrades KITTI ATE from 50.21 to 201.16 m. So the KITTI SOTA may be mostly the depth prior, not the Mamba refinement or BA. The paper neither states Metric3D's training data nor ablates a TartanAir-only depth backbone, so the comparison is not yet verified as a fair monocular test. This is a missing disclosure, not evidence of misconduct, but it blocks the headline claim.\n\nOther soft spots are smaller: no code, no error bars despite 'five runs,' and key hyperparameters (k, B, the matching feature dimension) are unstated. The paper is reproducible only in the loose sense. Self-citations are not load-bearing.\n\nWho is this for? Researchers in deep VO/SLAM will want to read the GMM and TAP ideas and the ablation methodology. It deserves peer review — the system is novel and the experimental design is mostly solid — but a serious referee should push for code, variance numbers, and the Metric3D ablation on KITTI before accepting. My recommendation: send it to review, with a clear request for that ablation and disclosure.","headline":"MambaVO is a well-built learning-to-optimize VO system with credible gains, but the KITTI SOTA claim rests on Metric3D v2's KITTI-trained depth prior, which the paper never discloses or ablates.","tokens_in":16750,"tokens_out":3882,"would_cite":true,"duration_ms":35977,"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":"MambaVO claims state-of-the-art monocular trajectory accuracy by refining semi-dense matches with a Mamba sequence model before bundle adjustment, while running in real time on four benchmarks.","keywords":["visual odometry","Mamba","state space model","bundle adjustment","learning to optimize","semi-dense matching","monocular SLAM","loop closure"],"falsifier":"Run MambaVO on a sequence where the semi-dense matcher produces a low inlier ratio and the monocular depth estimate is grossly wrong, and measure ATE. If the trajectory error jumps toward the no-PnP ablation values (KITTI ATE from 50.21 m to roughly 201 m), the system's robustness is inherited from the pretrained initialization, not from the Mamba refinement. Conversely, if GMM and BA still hold the pose when the initial matches are deliberately perturbed, the refinement is the load-bearing component.","tokens_in":15612,"feed_emoji":"📷","tokens_out":7886,"duration_ms":66786,"temperature":0.7,"pith_summary":"The paper tries to establish that the accuracy ceiling of learning-to-optimize visual odometry is set by ambiguous inter-frame matching, and that a Mamba-based sequential refinement of semi-dense matches removes enough of that ambiguity to set state-of-the-art absolute trajectory error on EuRoC, TUM-RGBD, KITTI, and TartanAir under a monocular setting. This matters because current deep VO systems either rely on dense flow or patch matching that degrades in low-texture or repetitive scenes, or they sacrifice real-time operation for accuracy. MambaVO reports real-time frame rates (22–30 Hz on a single GPU) and lower GPU memory than comparable learning-to-optimize systems while improving pose accuracy. A second version, MambaVO++, adds loop closure and global optimization and reports larger gains on long sequences. If correct, the work shifts the design question for deep VO from better optimization geometry toward better sequential matching.","feed_headline":"MambaVO refines matches to beat deep VO benchmarks in real time","feed_subtitle":"A Mamba module sharpens pixel matches before bundle adjustment and cuts trajectory error on four benchmarks.","key_machinery":"The load-bearing object is the Point-Frame Graph (PFG), a sliding-window graph whose vertices are camera poses and map points and whose edges encode which points each frame observes. The geometric argument runs through the Geometric Initialization Module (GIM) and the Geometric Mamba Module (GMM): GIM uses semi-dense matching plus monocular metric depth to seed pixel correspondences and a PnP initial pose, and GMM runs Mamba blocks—selective state-space sequence models—over the matching tokens derived from current features fused with a history token via cross-attention and a GRU, outputting per-match pixel refinements and weights that feed a differentiable bundle adjustment layer. The Mamba scan is the mechanism that carries sequential structure: it lets matching in the current frame be refined using the history of previous matches inside the PFG.","core_discovery":"The central claim is that replacing the matching front end of a learning-to-optimize odometer with a sequence-aware refiner—semi-dense geometric initialization followed by Mamba blocks that refine pixel coordinates and matching weights over a Point-Frame Graph—yields the best monocular absolute trajectory error among learning-based VO systems on four public benchmarks, in both odometry and loop-closure variants. The paper further claims that the gains come specifically from matching quality: on EuRoC, MambaVO raises the matching precision metric AUC@1° from 0.399 to 0.471, and ablation shows that removing the PnP initialization degrades KITTI ATE from 50.21 m to 201.16 m, while removing the Mamba blocks makes KITTI fail. The Trending-Aware Penalty is claimed to smooth the nested pose/matching optimization, improving convergence speed and final error.","pith_inferences":["The paper leaves implicit that the sequence-refinement module is likely backbone-agnostic: because geometric and context features enter only as concatenated inputs, replacing the semi-dense matcher or the monocular depth source with a stronger foundation model should compound the accuracy gains without redesigning the Mamba refinement.","A direct testable extension is a second PnP pass on the refined matches before bundle adjustment; the current pipeline feeds refined matches only into BA, so the extra loop would isolate how much of the final accuracy comes from refinement versus optimization.","The trend-balance training rule is described for visual odometry, but the mechanism—reweighting two losses by their recent decreasing trends—is a generic stabilizer for nested differentiable optimization and could be evaluated on other learning-to-optimize problems."],"forward_implications":["If the claims hold, deep VO accuracy can be improved at the matching front end: refining a few hundred semi-dense correspondences sequentially beats dense-flow and patch-matching systems without needing stronger bundle adjustment or denser flow.","The loop-closure variant shows that classic place recognition and pose-graph optimization on top of the learned odometer cut trajectory error by more than half on long sequences, so learned matching and geometric global optimization complement each other.","The reported matching-precision gains (AUC@1° rising from 0.399 to 0.471 on EuRoC) give a per-frame proxy for downstream pose accuracy, suggesting matching quality is the bottleneck the system targets.","Real-time monocular operation at 22–30 Hz with under 5 GB GPU memory means the accuracy gains are available to platforms that cannot run offline or large-batch inference."],"supporting_citations":[{"why":"Supplies the selective state-space sequence blocks that GMM uses to refine matching tokens.","marker":"[18]"},{"why":"Provides the semi-dense initial correspondences and pixel-wise geometric features that seed the Point-Frame Graph.","marker":"[59]"},{"why":"Provides monocular metric depth used to lift 2D matches to 3D points for pose initialization.","marker":"[22]"},{"why":"Supplies the PnP solver that derives the initial pose and map points from the seeded correspondences.","marker":"[28]"},{"why":"Supplies the differentiable bundle adjustment layer, with tangent-space backpropagation, that optimizes poses and map jointly.","marker":"[50]"},{"why":"Dense-flow learning-to-optimize baseline that motivates the matching-refinement design and is compared on all benchmarks.","marker":"[49]"},{"why":"Patch-matching deep VO baseline whose matching and efficiency characteristics GMM is designed to improve.","marker":"[51]"},{"why":"Identifies gradient variance in differentiable BA layers, the training problem that the Trending-Aware Penalty targets.","marker":"[21]"},{"why":"Patch-graph global-optimization SLAM baseline compared in the loop-closure setting.","marker":"[34]"},{"why":"Supplies the context features queried at matched pixels and fused with geometric features.","marker":"[39]"}],"fun_headline_variants":["MambaVO sharpens matching to slash trajectory error","Sequence-aware matching refiner improves deep visual odometry","MambaVO lifts matching precision to cut pose drift","Deep VO gets a Mamba boost for sharper matches and poses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The end-to-end accuracy depends on the pretrained semi-dense matcher and monocular depth estimator being good enough that PnP produces a usable initial pose; if those fail in an unseen scene, the Mamba refinement and bundle adjustment cannot recover the trajectory.","fun_headline_variants_meta":{"raw":{"variants":["MambaVO sharpens matching to slash trajectory error","Sequence-aware matching refiner improves deep visual odometry","MambaVO lifts matching precision to cut pose drift","Deep VO gets a Mamba boost for sharper matches and poses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1452,"prompt_tokens":947,"completion_tokens":505,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":440}},"tokens_in":563,"tokens_out":505,"duration_ms":5140,"temperature":1.0,"reasoning_tokens":440,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:35:02.740495+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MambaVO on a sequence where the semi-dense matcher produces a low inlier ratio and the monocular depth estimate is grossly wrong, and measure ATE. If the trajectory error jumps toward the no-PnP ablation values (KITTI ATE from 50.21 m to roughly 201 m), the system's robustness is inherited from the pretrained initialization, not from the Mamba refinement. Conversely, if GMM and BA still hold the pose when the initial matches are deliberately perturbed, the refinement is the load-bearing component.","supporting_citations":[{"cited_title":"Efficient loftr: Semi-dense local feature matching with sparse-like speed","cited_arxiv_id":null,"evidence_quote":"Provides the semi-dense initial correspondences and pixel-wise geometric features that seed the Point-Frame Graph."},{"cited_title":"PoseLib - Minimal Solvers for Camera Pose Estimation, 2020","cited_arxiv_id":null,"evidence_quote":"Supplies the PnP solver that derives the initial pose and map points from the seeded correspondences."},{"cited_title":"Tangent space backpropa- gation for 3d transformation groups","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable bundle adjustment layer, with tangent-space backpropagation, that optimizes poses and map jointly."},{"cited_title":"Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras","cited_arxiv_id":null,"evidence_quote":"Dense-flow learning-to-optimize baseline that motivates the matching-refinement design and is compared on all benchmarks."},{"cited_title":"Deep patch vi- sual odometry","cited_arxiv_id":null,"evidence_quote":"Patch-matching deep VO baseline whose matching and efficiency characteristics GMM is designed to improve."},{"cited_title":"From variance to veracity: Unbundling and mitigating gradient variance in differentiable bundle adjustment layers","cited_arxiv_id":null,"evidence_quote":"Identifies gradient variance in differentiable BA layers, the training problem that the Trending-Aware Penalty targets."},{"cited_title":"Deep Patch Vi- sual SLAM","cited_arxiv_id":null,"evidence_quote":"Patch-graph global-optimization SLAM baseline compared in the loop-closure setting."}],"review_version":1}