{"id":"541062e8-d74d-41cf-9701-990bcd504425","arxiv_id":"2504.19104","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"MISO uses learned multiresolution submap initialization and feature-space submap alignment to make neural SDF SLAM substantially faster while keeping reconstruction accuracy competitive.","lead":"MISO is a method for neural-implicit SLAM that organizes maps into multiresolution submaps and aligns them directly in feature space, avoiding expensive decoding of the full scene geometry. A generalist reader might care because it claims large speedups in back-end optimization for large-scale 3D reconstruction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Feature-space submap alignment assumes cross-submap feature comparability that is neither guaranteed nor established; the FastCaMo-Large result shows feature alignment can fail.","rationale":"The reader identified the key load-bearing assumption: cross-submap feature comparability in Eq. (16). I agree that this is the weakest point in the central claim, because if it fails, MISO cannot deliver its advertised speed advantage. My stress-test adds two specific observations from the manuscript. First, the paper offers no theoretical mechanism for comparability; sharing the decoder does not make independently optimized feature grids canonical, and Proposition 1 only addresses feature initialization, not cross-submap feature geometry. Second, the paper itself reports in Sec. VI-C that the coarse feature alignment had to be skipped on FastCaMo-Large, which is direct evidence that the feature cost can mislead the optimizer on the large-scale target domain. I also considered the evaluation-fairness issues noted by the reader, such as per-scene decoder pretraining on Newer College and skipped alignment on FastCaMo-Large, but those affect the strength of the empirical claims rather than the core mechanism. The conditional verdict is appropriate: the central idea is plausible and the ablations are helpful, but the feature-comparability assumption needs either a direct validation or a stated condition under which it is safe.","tokens_in":21823,"tokens_out":4277,"duration_ms":47591,"concrete_test":"Pick two overlapping submaps from ScanNet scenes 0011 or 0207 optimized from ground-truth poses; freeze networks and features; evaluate the pairwise cost in Eq. (16) at both hierarchy levels over a grid of perturbed relative poses around the known ground-truth T_uv (rotations of 0, 3, 6, 10 deg and translations of 0, 0.05, 0.1, 0.2 m). If the global minimum of the cost surface is not at ground truth within the alignment tolerance, or if the cost is flat near the true pose (ratio of true-pose cost to median perturbation cost close to 1), the feature-comparability assumption is refuted for that setup. Run the same perturbation grid with the SDF cost in Eq. (18) as a positive control to confirm the test has power.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central speed claim rests on Eq. (16), which minimizes L2 distance between the level-l feature fields of two submaps evaluated at overlapping grid vertices. This cost is only a valid surrogate for geometric alignment if the feature codes of independently optimized submaps are comparable for the same surface. Nothing in the formulation guarantees this: each submap solves Problem 1 with its own feature grid F and only the decoder D_theta is shared; there is no coupling, canonicalization, or reconstruction loss that forces equal geometry to produce equal features. Since D_theta is nonlinear, small feature distance need not imply close SDF fields. The paper supplies no proof or mechanism, only the ablation in Table VI showing that adding feature stages helps on two ScanNet scenes. The manuscript itself contains a direct contrary data point: in Sec. VI-C, the coarse-level feature alignment is skipped on FastCaMo-Large 'as we observe it leads to degraded results on these datasets.' That is an admission that the feature cost in Eq. (16) at the coarse level is not reliably minimized at the correct relative pose on the large-scale target domain. If the feature cost has spurious minima, the hierarchical optimizer is pulled into the wrong basin and must be rescued by the SDF stage (Algorithm 2 line 6), which removes the claimed speed advantage. The load-bearing assumption is therefore empirically plausible on selected indoor scenes but not established, and there is positive evidence it can fail.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MISO, a hierarchical back-end for neural implicit SDF SLAM. Locally, each submap is a multiresolution feature grid, and features are initialized by learned hierarchical encoders before joint pose-feature optimization. Globally, submap base poses are aligned by minimizing feature-space costs at increasing resolutions (Eq. 16) followed by a final SDF-based refinement (Eq. 19), and submaps are fused by averaging features and decoding (Eqs. 20-21). The claims are that this avoids expensive geometric decoding for the bulk of the optimization, yielding substantial speedups and improved robustness over MIPS-Fusion, Vox-Fusion++, and ICP-style baselines, with evaluations on ScanNet, FastCaMo-Large, and Newer College.","tokens_in":22100,"tokens_out":4446,"duration_ms":47408,"significance":"If the central claims hold, MISO would be a practically valuable back-end for large-scale neural implicit SLAM: it directly exploits multiresolution features for alignment, reports large wall-clock reductions over geometry-based submap alignment, and includes a useful closed-form least-squares analysis (Proposition 1) that motivates the learned initialization. The manuscript is clearly written, uses public benchmarks, and provides ablations on the encoder/decoder contributions (Table V), the hierarchical alignment stages (Table VI), and the feature-space metric (Fig. 10). The main weaknesses are that the speed claims are confounded by pre-training and by per-scene decoder pre-training on Newer College, and that the key feature-comparability assumption for Eq. (16) is not established and is contradicted by the FastCaMo-Large coarse-level failure. These are evidential gaps rather than formal contradictions, so the claims remain plausible but not yet convincingly demonstrated.","major_comments":[{"comment":"The central speed advantage of global alignment rests on the feature-space cost in Eq. (16), but the manuscript states in Sec. VI-C that on FastCaMo-Large the coarse-level feature alignment is skipped 'as we observe it leads to degraded results on these datasets.' This is direct evidence that the feature cost can have spurious minima on the target domain, and Sec. V provides no mechanism or guarantee that independently optimized submap feature fields are comparable for the same surface. Please report feature-only alignment accuracy on FastCaMo-Large (without the SDF rescue stage in Algorithm 2 line 6), quantify the failure modes of Eq. (16), and either add an analysis of when feature comparability holds or restrict the robustness claim to settings where it is verified.","section":"VI-C and Sec. V (Eq. 16)"},{"comment":"The local-mapping speed comparison in Table I is confounded by pre-training: MISO uses pre-trained encoder and decoder weights and is run for 20 epochs, while iSDF and Neural Points are trained from scratch for 100 epochs. Table V shows that the from-scratch variant No-ED reaches F-score 59.3 on scene 0011 at 100 epochs, essentially matching Full MISO at 10 epochs (59.4), and No-E with only a pre-trained decoder reaches 52.8 at 10 epochs. This suggests that a large part of the reported speed advantage is attributable to pre-training rather than to the hierarchical optimization scheme itself. Please add an ablation that fixes the pre-trained decoder across all variants and reports wall-clock time to reach a fixed F-score or loss, separating the contribution of hierarchical initialization from the contribution of pre-training.","section":"Table I and Table V"},{"comment":"The paper claims in Sec. VI-A that pre-trained weights are used on ScanNet and FastCaMo-Large 'without additional fine-tuning,' but Sec. VI-D states that on Newer College 'we first pre-train a decoder for each scene.' This per-scene decoder pre-training on the test domain weakens the stated generalization claim and makes the comparison with PIN-SLAM and KISS-ICP less direct, since those baselines do not receive per-scene pre-trained decoders. Please either rephrase the claims to acknowledge per-scene pre-training as a limitation, or add an experiment on Newer College without per-scene pre-training to show how much of the reported accuracy depends on this adaptation.","section":"VI-D and VI-A"}],"minor_comments":[{"comment":"Reference [48] misspells the venue name as 'International Conference on Computer Cision'; this should be 'International Conference on Computer Vision' (or 'ICCV').","section":"References"},{"comment":"The sentence 'We do not report the F-score for No-ED at epoch 10 because it fails to produce a mesh due to insufficient training' appears after Table VI but refers to the ablation in Table V; please move it to the discussion of Table V.","section":"Sec. VI-E"},{"comment":"The definition of the vertex index set I^uv_l as 'level-l vertices in submap u that lie within the overlapping region' is vague; please specify how the overlapping region is computed from the current base poses, since this affects both the cost and its gradient.","section":"Sec. V, Eq. (16)"},{"comment":"Table IV does not report optimization time, although the paper's headline claim is computational efficiency; please include runtime or convergence-time data for the Newer College experiments.","section":"Table IV"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid systems contribution and the issues are addressable with additional experiments and more careful claims. I recommend major revision rather than reject: the headline speed/robustness claims need to be decoupled from pre-training effects, and the FastCaMo-Large coarse-alignment failure must be explained or the claim about feature-space robustness narrowed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Jack,\n\nMISO is worth a look. The genuinely new piece is the idea of aligning neural implicit submaps directly in the learned feature space, hierarchically from coarse to fine, instead of decoding geometry first. The local hierarchical initialization from residuals, motivated by a clean linear least-squares derivation, is also a real trick and seems to deliver large speedups: 0.7–1.5 s vs tens of seconds on ScanNet, and the ablations support the individual components. That is a meaningful engineering contribution for neural implicit SLAM.\n\nThe soft spots are real, though. The load-bearing assumption in Eq. (16) is that independently optimized submaps have comparable feature fields for the same surface. Nothing in the formulation enforces that—submaps solve separate local problems with only the shared decoder, and a nonlinear decoder means small feature distance does not imply close SDF values. The paper gives only an ablation, and the ablation contains an admission: on FastCaMo-Large they skip the coarse feature alignment because it degrades results. That is a direct data point that the feature cost can fail in the target domain. It is fine to have a method that needs the SDF stage as a safety net, but then the speed advantage is smaller, and the paper should say so.\n\nThe evaluation also gives MISO an advantage that is not fully acknowledged. In Table I, MISO gets pre-trained encoder/decoder from Replica while iSDF and Neural Points train from scratch, and get 100 epochs vs 20. On Newer College, the decoder is pre-trained on each test scene, which undermines the generalization claim. These choices are disclosed, but they make the headline speedups look better than a fair comparison would.\n\nNone of this is fatal. The method is coherent, the paper is transparent about limitations, and Proposition 1 is a legitimate theoretical motivation rather than a circular step. The feature-comparability concern is the one thing I'd push on: give an experiment that varies the relative pose and plots the feature cost surface, or show that the learned features are consistent across overlapping submaps after local optimization. If the answer is 'not always, but the SDF stage rescues it', say that explicitly.\n\nI'd send this to peer review. It is a serious, potentially useful contribution that needs a fairer experimental setup and a more careful statement of when feature alignment is safe. For a reading group, it's a decent pick if people are tracking neural implicit SLAM.","headline":"MISO is a genuinely new neural implicit SLAM back-end with real speedups, but the feature-space alignment assumption is only partially tested and the baselines are not run under equal pre-training conditions.","tokens_in":22641,"tokens_out":3375,"would_cite":true,"duration_ms":34042,"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":"MISO claims that fast global consistency in neural implicit SLAM comes from aligning and fusing multiresolution submaps directly in feature space, skipping geometry decoding for most of the optimization.","keywords":["neural implicit SLAM","multiresolution feature grid","submap alignment","signed distance function","hierarchical optimization","feature-space registration","loop closure","pose graph optimization"],"falsifier":"Take two overlapping submaps of the same scene built from disjoint depth or LiDAR observations with a known ground-truth relative pose, then optimize only the feature-based cost in (16) over that relative pose; if its global minimum does not sit at the ground-truth pose, or within its uncertainty, the comparability premise that carries MISO fails.","tokens_in":21590,"feed_emoji":"🗺️","tokens_out":5884,"duration_ms":58173,"temperature":0.7,"pith_summary":"MISO claims that neural implicit SLAM can be made fast and globally consistent by optimizing the multiresolution feature grids themselves, rather than the geometry decoded from them. Locally, pre-trained hierarchical encoders initialize each level of the grid from residual inputs, so very few optimization epochs are needed. Globally, submaps are aligned by matching interpolated feature vectors in overlapping regions, with a short SDF-based refinement at the end. If the claim holds, large-scale neural signed-distance-function mapping becomes dramatically cheaper and stays accurate on real indoor and outdoor benchmarks.","feed_headline":"MISO aligns neural submaps without decoding geometry","feed_subtitle":"Matching learned feature grids directly, instead of surfaces, makes large-scale neural SLAM faster and more reliable.","key_machinery":"The central object is the multiresolution feature grid: a hierarchy of regular voxel grids whose trilinearly interpolated feature vectors are concatenated and decoded by a fixed MLP into signed distance values. Two mechanisms carry the argument. For local SLAM, a closed-form least-squares analysis shows the optimal initialization of a level's features is a linear function of coarser levels' residuals; this motivates learned 3D-CNN encoders that predict each level's grid from voxelized residuals. For global alignment, a correspondence-free cost compares the interpolated features of two submaps at overlapping grid vertices, and hierarchical coarse-to-fine optimization over submap poses minimizes it; only the final stage decodes SDF values, followed by feature-space averaging to fuse submaps in the world frame.","core_discovery":"The central claim is that both local submap optimization and global submap fusion can be carried out hierarchically in implicit feature space instead of in decoded geometry. For local SLAM, the paper proves that in a linear-least-squares special case, the optimal initialization of a level's features is a linear function of the residuals from coarser levels, and it converts this insight into learned hierarchical encoders that predict feature grids from voxelized residual inputs. For global alignment, two submaps are aligned by minimizing the squared Euclidean distance between trilinearly interpolated feature vectors at overlapping grid vertices, and only the final refinement stage uses predicted SDF values. The paper argues that this combination yields accuracy comparable to or better than SDF-based, correspondence-based, and ICP baselines while being substantially faster.","pith_inferences":["The paper does not say this, but if cross-submap feature comparability holds, MISO's feature-based alignment cost could let multiple agents fuse maps without ever exchanging raw geometry or decoded surfaces.","A direct extension of the linear-least-squares initialization result is that any grid-based implicit representation trained with a squared residual loss could be warmed up by a learned residual encoder, not just SDF grids.","The paper's own ablation notes mild domain mismatch after synthetic-only pretraining, which suggests alignment and initialization quality on a new deployment should track how well the pretraining distribution matches the scene.","Because alignment consumes only grid vertices, replacing dense grids with sparse or factored structures would likely preserve the speed gains while easing the memory cost the paper identifies as a limitation."],"forward_implications":["Local submap optimization needs only about 20 epochs with learned hierarchical initialization, compared with 100 for the SDF and neural-point baselines, cutting back-end optimization time from tens of seconds to roughly one to eight seconds on ScanNet scenes.","Global submap alignment avoids decoding SDF for most iterations, bringing alignment time down from roughly 59–138 seconds to about 10–19 seconds while matching or improving final rotation and translation errors.","Feature-space fusion by weighted averaging of submap features yields a globally consistent SDF and mesh after loop closures, including on large indoor and outdoor LiDAR sequences.","The same hierarchical back-end is positioned to extend beyond SDF reconstruction to other scene properties such as radiance, semantics, or uncertainty.","In incremental mode, the same submap and alignment machinery improves trajectory accuracy substantially over the odometry initialization on outdoor Newer College sequences."],"supporting_citations":[{"why":"Supplies the SDF and free-space measurement cost formulations used in local submap optimization.","marker":"[11]"},{"why":"Provides the neural-point baseline and the incremental tracking-and-mapping loop that MISO extends.","marker":"[29]"},{"why":"Supplies the MIPS-Fusion submap alignment baseline and the FastCaMo-Large dataset used for large-scale evaluation.","marker":"[1]"},{"why":"Provides the correspondence-free submap alignment baseline from Vox-Fusion++.","marker":"[10]"},{"why":"Contributes the multiresolution grid encoding idea that underlies MISO's feature grids.","marker":"[9]"},{"why":"Motivates the 3D CNN encoder architecture used for voxelized residual inputs.","marker":"[45]"},{"why":"Motivates hierarchical encoders that predict lower-level feature grids from higher-level information.","marker":"[46]"},{"why":"Provides the ScanNet real-world RGB-D benchmark used for local mapping and submap alignment experiments.","marker":"[47]"},{"why":"Provides the Replica scenes used for offline pretraining of the encoders and decoder.","marker":"[49]"},{"why":"Serves as the odometry baseline and comparison for outdoor Newer College evaluation.","marker":"[53]"}],"fun_headline_variants":["Skip geometry decode: MISO aligns submaps in feature space","MISO: Fast neural SLAM by fusing feature grids","MISO: Hierarchical optimization in feature space for SLAM","MISO: Align neural submaps without decoding surfaces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the premise that feature vectors learned independently in different submaps are comparable, so aligning them in feature space is equivalent to aligning the underlying geometry; if that premise fails, the feature-based alignment cost can mislead the optimizer.","fun_headline_variants_meta":{"raw":{"variants":["Skip geometry decode: MISO aligns submaps in feature space","MISO: Fast neural SLAM by fusing feature grids","MISO: Hierarchical optimization in feature space for SLAM","MISO: Align neural submaps without decoding surfaces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000571,"raw_usage":{"total_tokens":2663,"prompt_tokens":869,"completion_tokens":1794,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":1724}},"tokens_in":485,"tokens_out":1794,"duration_ms":12492,"temperature":1.0,"reasoning_tokens":1724,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T06:01:33.501042+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two overlapping submaps of the same scene built from disjoint depth or LiDAR observations with a known ground-truth relative pose, then optimize only the feature-based cost in (16) over that relative pose; if its global minimum does not sit at the ground-truth pose, or within its uncertainty, the comparability premise that carries MISO fails.","supporting_citations":[{"cited_title":"iSDF: Real-time neural signed distance fields for robot perception,","cited_arxiv_id":null,"evidence_quote":"Supplies the SDF and free-space measurement cost formulations used in local submap optimization."},{"cited_title":"PIN-SLAM: LiDAR SLAM using a point-based implicit neural representation for achieving global map consistency,","cited_arxiv_id":null,"evidence_quote":"Provides the neural-point baseline and the incremental tracking-and-mapping loop that MISO extends."},{"cited_title":"MIPS-Fusion: Multi-implicit-submaps for scalable and robust online neural RGB-D reconstruction,","cited_arxiv_id":null,"evidence_quote":"Supplies the MIPS-Fusion submap alignment baseline and the FastCaMo-Large dataset used for large-scale evaluation."},{"cited_title":"Vox-Fusion++: Voxel-based Neural Implicit Dense Tracking and Mapping with Multi-maps","cited_arxiv_id":"2403.12536","evidence_quote":"Provides the correspondence-free submap alignment baseline from Vox-Fusion++."},{"cited_title":"Instant neural graphics primitives with a multiresolution hash encoding,","cited_arxiv_id":null,"evidence_quote":"Contributes the multiresolution grid encoding idea that underlies MISO's feature grids."},{"cited_title":"Con- volutional occupancy networks,","cited_arxiv_id":null,"evidence_quote":"Motivates the 3D CNN encoder architecture used for voxelized residual inputs."},{"cited_title":"NV AE: A deep hierarchical variational autoencoder,","cited_arxiv_id":null,"evidence_quote":"Motivates hierarchical encoders that predict lower-level feature grids from higher-level information."},{"cited_title":"ScanNet: Richly-annotated 3D reconstructions of indoor scenes,","cited_arxiv_id":null,"evidence_quote":"Provides the ScanNet real-world RGB-D benchmark used for local mapping and submap alignment experiments."},{"cited_title":"KISS-ICP: In defense of point-to-point ICP–simple, accurate, and robust registration if done the right way,","cited_arxiv_id":null,"evidence_quote":"Serves as the odometry baseline and comparison for outdoor Newer College evaluation."}],"review_version":1}