{"id":"f8a4a3ae-5f93-4c71-a5cd-9100b049118e","arxiv_id":"2504.16545","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"ToF-Splatting is the first 3D Gaussian Splatting SLAM system designed for sparse ToF depth, using a retrained multi-frame depth integration module with monocular cues and outlier filtering.","lead":"This paper builds a SLAM system that turns very sparse ToF depth points (64 per frame) into dense 3D maps by merging depth guesses from multiple camera views and a monocular depth prior. The result is a 3D Gaussian Splatting SLAM pipeline that tracks and maps better than prior sparse-ToF methods on a real indoor benchmark.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ZJUL5 gains may be explained by the multi-frame module; the outlier-filtering prior using the network's own no-ToF prediction is circular and unablated, so the headline SOTA claim needs a cleaner controlled comparison.","rationale":"The reader identified the same dependency on the multi-frame integration network and flagged inadequately-validated densified depth as the weakest assumption, with data snooping and error bars as supporting concerns. My stress-test focuses on the most load-bearing version: the system's SOTA margins on ZJUL5, the only real sparse-ToF dataset, are most plausibly driven by the retrained DoD network and its monocular and multi-view priors. The core loop -- tracking poses feed the module, module output supervises tracking and seeding, and the outlier filter compares ToF points to the module's own no-ToF prediction -- means the filter's effectiveness is not independent evidence of robustness. The paper also self-admits a training/eval mismatch: the module is retrained on ScanNetv2 with simulated sparsity [0%, 0.03%], then evaluated on real ZJUL5 noise; this makes cross-dataset transfer the linchpin, yet the only real-data evaluation is ZJUL5, with no error bars or per-seed variance. I do not claim the method is wrong; the claim is that the evidence as written cannot separate the module's contribution from the pipeline's, so the headline 'first 3DGS SLAM suited for sparse ToF' and the SOTA statement should be presented with that caveat or with the missing ablations. A CONDITIONAL verdict fits: accept only with the controlled comparison and error bars.","tokens_in":16314,"tokens_out":1752,"duration_ms":15096,"concrete_test":"Run ToF-Splatting on ZJUL5 with the same hyperparameters and training set, but with the outlier filter disabled (quantile q=1.0) and report tracking/mapping; also run with D_k replaced by DELTAR-densified depth and by raw sparse-depth nearest-neighbour upsampling. If ATE and F-score degrade sharply when the filter is disabled or the module is replaced, then D_k and its self-prior are indeed the driving factor and the ablation is needed; if the gains persist, the multi-frame module concern is not the dominant explanation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is state-of-the-art tracking and mapping with 64-point ToF input on ZJUL5 (Table 1: ATE 0.051 vs 0.107; Table 2: F-score 0.664 vs 0.604). Load-bearing for that claim is D_k, the dense depth from the retrained multi-frame module, which supervises both phases: it seeds Gaussians (Sec. 3.4) and drives L_depth in tracking (Eq. 5). If D_k is biased, the whole system inherits that bias. The paper's own description shows a loop: the network is conditioned on poses from tracking, and the outlier filter (Sec. 3.2) computes a no-ToF prediction from the same network and discards ToF points above quantile q=0.75 against it. If that prior is wrong, filtering removes valid measurements and enforces the network's hallucinated depth, so the filter can only inflate agreement with the prior, not correct it. The critical missing evidence is a controlled comparison isolating the multi-frame module: ablating it (e.g., using DELTAR densified depth or raw sparse depth for seeding/supervision) and ablating the outlier filter. With those ablations absent, the 2x ATE gain over ToF-SLAM and the mapping gains could be attributed to the pretrained-and-retrained DoD network plus DepthAnything priors rather than to multi-frame integration, keyframe handling, or the 3DGS backend. This is a correctness-risk concern, not a consistency contradiction: the paper's evidence does not yet identify which component is responsible for the central SOTA result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ToF-Splatting, a 3D Gaussian Splatting (3DGS) based dense SLAM system designed for extremely sparse Time-of-Flight depth input (e.g., 8x8 = 64 points per frame). The pipeline combines three modules: a tracking frontend that optimizes poses with photometric and geometric losses, a mapping backend that seeds and optimizes 3D Gaussians, and a multi-frame integration module that extends the authors' prior Depth on Demand (DoD) framework to fuse sparse ToF depth, multi-view geometry from a keyframe buffer, and monocular cues from Depth Anything V2. The integration module also includes an outlier filtering step based on the network's own no-ToF prediction. Experiments on the real-world ZJUL5 dataset report lower ATE than ToF-SLAM (0.051 m vs. 0.107 m) and higher mapping F-score (0.664 vs. 0.604), with additional simulated experiments on TUM RGB-D and Replica, runtime measurements, and an explicit limitations paragraph.","tokens_in":16681,"tokens_out":3651,"duration_ms":37237,"significance":"If the reported results hold up, ToF-Splatting is a meaningful advance: it is the first 3DGS-based SLAM system specifically adapted to sparse, noisy ToF sensors, and the ZJUL5 experiments demonstrate that a multi-frame integration strategy can support dense mapping from very few depth points. The paper does several things well: it evaluates on real sparse-ToF data rather than only simulations, it ablates the contribution of monocular versus multi-view cues within the integration module, it includes robustness studies for depth sparsity, noise, and temporal sparsity, and it reports runtime and limitations transparently. However, the central SOTA claim is not yet fully supported, because the experiments do not isolate the multi-frame integration module from the pretrained components (DoD, Depth Anything V2) that feed it, and the outlier filter is unablated and has a circular flavor. These are correctness-risk concerns rather than internal contradictions, and they are addressable with additional controlled experiments.","major_comments":[{"comment":"The outlier filter is load-bearing for the ZJUL5 results, since real ToF points are described as very noisy and outlier-prone, yet it is not ablated and its design raises a circularity concern. The filter computes a no-ToF depth prediction with the same DoD network, discards ToF measurements whose L1 error against that prediction exceeds quantile q=0.75, and then runs the same network again using the filtered points. If the no-ToF prediction is biased, this procedure removes valid measurements and enforces the network's own hallucinated depth instead of correcting it. The paper should report an ablation of the filter (filter on/off), the fraction of points discarded per sequence, and ideally compare the filter against an independent geometric consistency check or against filtering with a different reference depth.","section":"§3.2, Outlier Handling"},{"comment":"The central claim that 'multi-frame integration' is responsible for the large gains over ToF-SLAM is not directly supported by the experiments. Table 3 ablates monocular and multi-view cues inside the DoD-based module, but there is no ablation that replaces the module itself -- for example, using raw sparse depth, DELTAR-densified depth, or single-frame DoD depth for seeding and supervision while keeping the rest of the pipeline fixed. Without such a controlled comparison, the Table 1–2 improvements could be attributed to the retrained DoD network (the authors' own prior work) and the Depth Anything V2 priors rather than to the multi-view integration, keyframe handling, or the 3DGS backend. This is the key experiment needed to support the stated state-of-the-art claim.","section":"§4.2, Ablation Studies and Tables 1–3"},{"comment":"The quantitative evaluation on ZJUL5 is reported without error bars or multiple runs, and the dataset contains only seven real sequences. Given that several hyperparameters (q=0.75, νth=0.1, σ=0.98, λtrack, λmap, N=4, b=15 cm) appear manually selected and are not varied in a sensitivity analysis, it is difficult to assess whether the 2x ATE margin over ToF-SLAM is robust. The paper should state the hyperparameter selection protocol (e.g., validation split or sensitivity plots) and report variance across repeats or at least per-sequence stability. The claim of state-of-the-art on 'reference datasets' also rests on a single real dataset; the TUM and Replica results are simulated and are not compared against the same baselines used on ZJUL5.","section":"Tables 1 and 2"},{"comment":"The mapping comparison with ToF-SLAM is partially ambiguous because, as the paper states, 'for each scene, we collect the predicted pose of each frame and render depth and color from 3DGS' for the ToF-Splatting meshes, whereas the ToF-SLAM numbers are presumably produced by its own pipeline. It would strengthen the comparison to confirm that the same reconstruction protocol (TSDF voxel size, truncation, marching cubes parameters) is applied to both methods, since mesh-based F-score is sensitive to these choices. At minimum, the paper should state explicitly whether the protocol for all baselines is identical and whether any method's reported numbers are taken from the original paper rather than recomputed under the same protocol.","section":"§4.1, Table 2"}],"minor_comments":[{"comment":"The text says 'ZJUL5 dataset' has '8 sequences available' in the Table 1 caption, while Section 4 states seven indoor scene recordings and Table 2 lists seven named scenes. Please correct this inconsistency.","section":"Abstract and §4.1, Table 1"},{"comment":"The notation in the isotropy loss is confusing: the text says diag(·) extracts the diagonal values and then refers to the average of the resulting vector, but the equation writes ||diag(Sj) - diag(Sj)·1_{3x1}||_1, which suggests diag(Sj) is a scalar. Please make the distinction between the diagonal vector and its mean explicit.","section":"§3.4, Eq. (10)"},{"comment":"The sentence 'we prove this approach effective' overstates what is shown; the keyframe selection and seeding strategy are not ablated in the experiments. Please rephrase to 'we show empirically' or add an ablation, and in the meantime the claim should be softened.","section":"§3.4, Initialization"},{"comment":"In the TUM comparison, ToF-Splatting's 0.02%/0.04% simulated density results are compared with RGB-only methods and with a 100% density RGB-D MonoGS baseline. Please clarify that the RGB-D 100% row is an upper-bound reference and not a sparse-input competitor, and describe how the sparse points are sampled (e.g., random versus grid) since this affects the comparison.","section":"§4.2, Table 4"},{"comment":"The temporal sparsity experiment provides first 50 frames with ToF depth to establish scale, but the figure and text do not state how the scale is maintained afterward for frames without ToF measurements. A brief explanation of how the multi-frame module obtains metric scale in those frames would help the reader interpret the graceful degradation shown in Figure 10.","section":"Supplementary, §9 Temporal Sparsity"}],"recommendation":"major_revision","confidential_remarks":"The paper's main evidence is a single real dataset and a set of simulated experiments, and the strongest reported gains depend on a retrained version of the authors' own prior work (DoD) combined with a strong external monocular prior (Depth Anything V2). The missing ablations -- particularly the outlier filter and a module-replacement study -- are essential before the SOTA claim can be considered established. I would also encourage the editor to check whether the hyperparameter settings were selected on the ZJUL5 test sequences, since the paper does not describe a validation protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a genuine step forward for sparse-ToF SLAM. The authors build the first 3DGS-based pipeline that takes 8x8 (64-point) ToF depth and turns it into dense depth via a retrained multi-view Depth on Demand network with monocular cues, then feeds that into a MonoGS-style Gaussian SLAM. On the only real sparse-ToF benchmark (ZJUL5), it roughly halves the ATE of the previous approach (0.051 vs 0.107 m) and improves F-score from 0.604 to 0.664. That is a solid, credible result.\n\nWhat is new is the combination: extending DoD from two-view to multi-view by iterating over source views, injecting Depth Anything V2 monocular depth as a normalized cue, and an outlier rejection step that compares sparse ToF points against a no-ToF depth prior before re-running integration. The ablation in Table 3 shows both monocular and multi-view cues contribute, with multi-view doing the heavy lifting. The paper also includes robustness sweeps on Replica (sparsity, noise, temporal subsampling) and a TUM comparison that is fair in context.\n\nThe soft spots are real but mostly about attribution, not about whether the system works. There are no error bars or multiple seeds reported, so we do not know how sensitive the ZJUL5 numbers are. Hyperparameters like outlier quantile q=0.75 are chosen with knowledge of the target dataset. More importantly, the paper never isolates the multi-frame integration module against a strong depth-completion baseline inside the same 3DGS pipeline; we do not get to see whether a simple DELTAR-densified depth used with MonoGS would close much of the gap. The outlier filter is also under-ablated, and it has a mild circular flavor: it rejects ToF points based on the network's own no-ToF prediction, so if that prior is biased the filter enforces the hallucination rather than correcting it. That concern is worth testing, but it is not a load-bearing flaw: the filter could plausibly remove real sensor outliers.\n\nThe pose-depth loop (poses used to generate D_k, D_k used for pose tracking) is an iterative refinement, not a test-set fit, because the network is trained on ScanNet and evaluated on external datasets. I would not call that circularity fatal.\n\nBottom line: the paper earns a serious refereeing. It deserves revision, not rejection: add error bars, ablate the outlier filter and a DELTAR-based baseline, and discuss the quantile choice. If those hold, this is a useful citation for anyone combining learned depth completion with Gaussian SLAM.","headline":"First 3DGS-SLAM that handles 64-point ToF depth; strong ZJUL5 gains but missing ablations leave which component earns them unclear.","tokens_in":17277,"tokens_out":3351,"would_cite":true,"duration_ms":29420,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ToF-Splatting claims that 64-point 8×8 ToF measurements, integrated across keyframes with monocular cues, are enough to run accurate 3D Gaussian Splatting SLAM.","keywords":["time-of-flight","sparse depth","SLAM","3D Gaussian Splatting","dense mapping","multi-frame integration","depth completion","camera tracking"],"falsifier":"Run ToF-Splatting on a real sparse-ToF sequence with ground-truth depth and poses, disable the outlier filter, and check whether the filter removes points that lie within sensor noise of the ground truth; if such valid points are dropped, the filter is enforcing the network's prior rather than removing sensor outliers, and tracking error should track the network's depth error rather than the ToF noise level.","tokens_in":16104,"feed_emoji":"📡","tokens_out":6748,"duration_ms":55444,"temperature":0.7,"pith_summary":"This paper claims that very sparse time-of-flight (ToF) depth, as little as the 64 points from an $8\\times 8$ sensor, is enough to run an accurate dense SLAM system if the sparse measurements are integrated across multiple frames and combined with monocular cues. The authors propose ToF-Splatting, the first 3D Gaussian Splatting-based SLAM pipeline built for such sparse ToF input. A multi-frame integration module produces dense metric depth maps that seed new Gaussians and supervise both tracking and mapping. On the real-world ZJUL5 dataset, the system reports an absolute trajectory error of 0.051 m versus 0.107 m for the prior ToF-SLAM baseline, and a higher mapping F-score. If correct, this shows that the extreme sparsity and noise of low-power ToF sensors need not block their use in mobile and AR/VR SLAM.","feed_headline":"64-point ToF depth now drives full dense SLAM","feed_subtitle":"ToF-Splatting fuses sparse depth, motion, and monocular cues to halve trajectory error on ZJUL5.","key_machinery":"The load-bearing component is the multi-frame integration module, an extension of the Depth on Demand framework. Given a target frame, it iteratively refines a dense depth prediction using a set of source keyframes selected for sufficient parallax, their relative poses, the current sparse ToF measurements, and the RGB image paired with a normalized monocular depth prior from Depth Anything v2. The output dense depth map $D_k$ serves three purposes: it seeds new Gaussians during mapping, it supervises the geometric term of the tracking loss, and it supervises the mapping optimization. Its outlier handling step produces a prior depth prediction without the ToF points, discards measurements above a quantile of the $\\ell^1$ error against that prior, and repredicts with the cleaned sparse depth.","core_discovery":"The central discovery is that replacing single-frame depth completion with multi-frame integration makes a 3DGS-based SLAM system usable with sensors that return only 64 depth points per frame. The multi-frame integration module, an extension of Depth on Demand, combines the sparse ToF measurements with multiple keyframe views, the relative poses from tracking, and monocular depth priors from Depth Anything v2 to predict dense metric depth maps. These maps provide the geometric supervision that 3DGS-based SLAM needs; the system also filters ToF outliers by comparing each measurement against the network's prediction without ToF input. ToF-Splatting reports state-of-the-art tracking (ATE 0.051 m on ZJUL5) and mapping (F-score 0.664) among methods using sparse ToF, and on TUM RGB-D at 0.04% depth density it approaches the accuracy of a baseline using 100% dense depth.","pith_inferences":["Because the outlier filter compares each ToF measurement against the network's own no-ToF depth prediction, a systematically biased depth prior would silently discard valid measurements; the method's robustness ceiling is therefore set by how well the ScanNet-trained network transfers to the deployment scene.","The reported ZJUL5 and Replica results measure generalization to real and synthetic data, but the training sparsities stop at 0.03% and the test densities start at 0.02%, so a wider sweep of densities and sensor noise profiles would clarify where the gains saturate.","The multi-frame integration module could be evaluated in isolation with ground-truth poses and ToF measurements while varying the monocular prior, separating the contribution of multi-view geometry from that of the monocular cue.","The same keyframe-buffer integration idea transfers to NeRF-based SLAM or to depth completion over posed video, since the module only needs posed views and sparse depth as input."],"forward_implications":["Low-power ToF sensors with tens of depth points become viable for dense SLAM in mobile and AR/VR devices, removing the need for higher-resolution, more power-hungry depth sensors.","3DGS-based SLAM systems, previously thought to require dense and accurate depth, can be bootstrapped from very sparse measurements when multi-view geometry is integrated across a keyframe buffer.","At 0.04% depth density on TUM RGB-D, the method approaches the tracking accuracy of a baseline that receives 100% dense depth, suggesting sparse active depth plus geometry can nearly substitute for dense RGB-D input indoors.","The reported runtime of about 1.5 frames per second shows the pipeline is not real-time as published, so deployment depends on faster 3DGS-based SLAM backends."],"supporting_citations":[{"why":"Supplies the Depth on Demand framework that the multi-frame integration module extends from two-view to multi-view processing.","marker":"[3]"},{"why":"Provides the ZJUL5 real sparse-ToF dataset, the sensor specifications, and the ToF-SLAM baseline that the paper must beat.","marker":"[19]"},{"why":"DELTAR is the depth completion network used by the ToF-SLAM baseline, representing the single-frame completion approach this work argues is insufficient.","marker":"[16]"},{"why":"Depth Anything v2 provides the monocular depth prior that is fed into the multi-frame integration module as a normalized cue.","marker":"[40]"},{"why":"ScanNet is the dataset and protocol used to retrain the multi-frame integration module and to build the ZJUL5 ground truth.","marker":"[5]"},{"why":"Shows how to integrate multiple views in the depth prediction by using a different source view at each iteration, which the paper adapts for the keyframe buffer.","marker":"[4]"},{"why":"MonoGS is the 3DGS-based SLAM baseline whose tracking and mapping losses, and its requirement for dense depth, frame the comparison.","marker":"[22]"}],"fun_headline_variants":["Multi-frame integration turns sparse ToF into dense SLAM","Sparse ToF dots become dense SLAM via multi-frame fusion","ToF-Splatting: SLAM from just 64 depth points per frame","Dense SLAM from 64-point ToF via multi-frame integration","Multi-frame cues rescue sparse ToF for 3DGS SLAM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system's tracking and mapping supervision rests on dense depth maps produced by a network retrained on ScanNet; if that network's depth prior is wrong for an unseen scene, errors in those maps dominate the optimization, and the outlier filter, which checks ToF points against the same prior, can throw away valid measurements.","fun_headline_variants_meta":{"raw":{"variants":["Multi-frame integration turns sparse ToF into dense SLAM","Sparse ToF dots become dense SLAM via multi-frame fusion","ToF-Splatting: SLAM from just 64 depth points per frame","Dense SLAM from 64-point ToF via multi-frame integration","Multi-frame cues rescue sparse ToF for 3DGS SLAM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1362,"prompt_tokens":895,"completion_tokens":467,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":373}},"tokens_in":511,"tokens_out":467,"duration_ms":3800,"temperature":1.0,"reasoning_tokens":373,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:00:56.869096+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ToF-Splatting on a real sparse-ToF sequence with ground-truth depth and poses, disable the outlier filter, and check whether the filter removes points that lie within sensor noise of the ground truth; if such valid points are dropped, the filter is enforcing the network's prior rather than removing sensor outliers, and tracking error should track the network's depth error rather than the ToF noise level.","supporting_citations":[{"cited_title":"Depth on demand: Streaming dense depth from a low frame-rate active sensor","cited_arxiv_id":null,"evidence_quote":"Supplies the Depth on Demand framework that the multi-frame integration module extends from two-view to multi-view processing."},{"cited_title":"Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor","cited_arxiv_id":null,"evidence_quote":"Provides the ZJUL5 real sparse-ToF dataset, the sensor specifications, and the ToF-SLAM baseline that the paper must beat."},{"cited_title":"Deltar: Depth estimation from a light-weight tof sensor and rgb image","cited_arxiv_id":null,"evidence_quote":"DELTAR is the depth completion network used by the ToF-SLAM baseline, representing the single-frame completion approach this work argues is insufficient."},{"cited_title":"ScanNet: Richly-annotated 3d reconstructions of indoor scenes","cited_arxiv_id":null,"evidence_quote":"ScanNet is the dataset and protocol used to retrain the multi-frame integration module and to build the ZJUL5 ground truth."},{"cited_title":"Mat- toccia","cited_arxiv_id":null,"evidence_quote":"Shows how to integrate multiple views in the depth prediction by using a different source view at each iteration, which the paper adapts for the keyframe buffer."},{"cited_title":"Kelly, and An- drew J","cited_arxiv_id":null,"evidence_quote":"MonoGS is the 3DGS-based SLAM baseline whose tracking and mapping losses, and its requirement for dense depth, frame the comparison."}],"review_version":1}