{"id":"aeadf7dc-1a59-490a-8855-26ca2448a792","arxiv_id":"2412.17769","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"ActiveGS combines Gaussian splatting with a voxel map and a viewpoint-distribution confidence score to guide an RGB-D robot toward poorly reconstructed areas.","lead":"This paper builds a robot mapping system that mixes a Gaussian splatting map with a coarse voxel map, and uses a confidence score per surface patch to decide where the robot should look next. The method improves active scene reconstruction for drones and ground robots compared with several recent alternatives.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The confidence score k_i (Eq. 6) is never validated against actual reconstruction error, and the low-confidence ROI threshold is unspecified; if k_i misranks under-reconstructed primitives, the exploitation term guides the planner to the wrong places.","rationale":"Read in good faith, ActiveGS is a clear, reproducible system: the hybrid map representation is well specified, the code is released with a verification tag, and the eight-scene simulation study provides genuine comparative evidence against FBE, FisherRF, and NARUTO. The central claim, however, is not just that the system works but that it works by 'actively collecting scene information in under-reconstructed and unexplored areas.' The under-reconstructed part is carried entirely by the confidence score in Eqs. (6)-(8). That score is a plausible heuristic, but the paper provides no evidence that it tracks actual per-primitive reconstruction quality, and the one ablation that could isolate it (Ours vs. Ours(w/o ROI)) removes frontier and confidence ROI jointly. The missing low-confidence threshold compounds the issue: without it, even the algorithm's exact behavior is under-specified. All three of the reader's concerns are legitimate; I focus on the confidence proxy because it is the causal heart of the proposed method. The proposed test is feasible with the released code and would settle whether the proxy is informative. This does not move the verdict: the paper should remain CONDITIONAL pending that validation, exactly as the reader concluded. No concern about author conduct or intent is raised; the issue is an unvalidated design assumption and a missing parameter.","tokens_in":11524,"tokens_out":8649,"duration_ms":88031,"concrete_test":"Compute a per-primitive error proxy for each final GS map: render all held-out test viewpoints, assign each pixel's absolute RGB and depth residual to the primitives contributing to it via the weights w_i in Eq. (3), and aggregate per primitive. Then measure the rank correlation between k_i and this error proxy scene-by-scene, and rerun the planner with the low-confidence ROI threshold set to, say, the 10th, 50th, and 90th percentiles of the k_i distribution. If the correlation is weak or the PSNR/completeness comparison against FBE and FisherRF changes materially with the threshold, the confidence-guided exploitation is not doing the claimed work; if the correlation is strong and results are threshold-insensitive, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ActiveGS's exploitation term in Eq. (9) rests on the assumption that a primitive's confidence k_i = gamma_i exp(beta_i) (Eqs. 6-8) identifies under-reconstructed areas. The score uses only viewpoint count, distance weighting, normal alignment, and direction dispersion; it contains no photometric error, depth residual, or training signal. This assumes sparse or directionally biased observations are the cause of poor reconstruction, which need not hold: a surface observed densely from varied angles can still be wrong due to specularity, sensor noise, or optimization failure, while a simple planar surface seen from three directions can be accurate but receive low k_i. The paper never reports any correlation between k_i and per-primitive reconstruction error, and the ablation 'Ours (w/o ROI)' removes both frontier-based and confidence-based ROI sampling, so the specific benefit of confidence is not isolated. Moreover, Sec. III-E says voxels containing 'low-confidence Gaussian primitives' are added to the ROI set, but no threshold is defined anywhere in Sec. IV; this missing parameter makes the ROI sampling step non-reproducible and leaves open the possibility that the reported gains depend on an unstated threshold. The limitation paragraph in Sec. V mentions only perfect localization, not this unvalidated proxy. If k_i misranks primitives, the planner wastes time inspecting phantom low-confidence regions or misses genuine errors; with the large exploration weight phi=1000 the effect can be masked early, but it dominates once exploration saturates. This is the weakest load-bearing step in the argument for the paper's second claim and for the causal reading of the abstract.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"ActiveGS proposes an active scene reconstruction framework for a mobile robot with an RGB-D camera, based on a hybrid map that combines 2D Gaussian surfels for high-fidelity dense reconstruction with a coarse voxel map for exploration and path planning. The key novelty is a per-primitive confidence score k_i = gamma_i exp(beta_i), derived from the spatial distribution of observing viewpoints (Eqs. 6-8), which is used in two ways: to define an exploitation utility term in the candidate-viewpoint utility function (Eq. 9) and to select ROI voxels around 'low-confidence' Gaussian primitives for targeted viewpoint sampling (Sec. III-E). The framework iterates between incremental GS training and view planning, and is evaluated in Habitat/Replica on 8 scenes with 5 trials each, against frontier-based exploration (FBE), FisherRF, NARUTO, and two ablations, plus a real-world UAV experiment. The paper claims superior rendering (PSNR) and mesh (completeness) quality relative to these baselines.","tokens_in":11738,"tokens_out":4277,"duration_ms":45952,"significance":"If the claims hold, ActiveGS is a useful and timely contribution: it shows that a Gaussian-splatting representation can support online active reconstruction with fast map updates and explicit quality-oriented planning. The paper has clear strengths: a reproducible open-source implementation with a reproducibility certificate, a comparatively extensive simulation evaluation across 8 Replica scenes with multiple trials, a real-world UAV demonstration, and a simple confidence rendering mechanism that avoids expensive per-candidate gradient computations. The central comparative claim is supported by the reported experimental design, although the key confidence proxy is not directly validated against reconstruction error and one planning parameter is underspecified. These issues are addressable and do not appear to invalidate the main framework, but they are load-bearing for claim (ii) about confidence-guided inspection.","major_comments":[{"comment":"The confidence score k_i is a geometric heuristic based only on the set of viewpoints that observe primitive i: it captures distance, normal alignment, and direction dispersion, but contains no photometric error, depth residual, or training signal. The paper never validates k_i against per-primitive reconstruction error, despite the exploitation term in Eq. (9) and the low-confidence ROI selection in Sec. III-E both assuming that low-k_i primitives are exactly the under-reconstructed ones. I recommend reporting a quantitative relation between k_i (or rendered confidence K) and per-primitive error, for example depth residual, photometric error, or distance to a TSDF ground-truth surface, and/or adding an ablation in which the planner is driven by actual per-primitive error instead of k_i. Without such validation, the central mechanism for 'targeted inspection around under-reconstructed surfaces' rests on an untested assumption.","section":"Sec. III-C, Eqs. (6)-(8)"},{"comment":"The criterion for a 'low-confidence Gaussian primitive' is never specified. Sec. III-E states that voxels containing low-confidence Gaussian primitives are added to the ROI set, but neither Sec. IV-A nor any later section defines the threshold (e.g., an absolute k_i value, a percentile, or a fraction of primitives). This makes the ROI sampling step non-reproducible and leaves open the possibility that the reported gains depend sensitively on an unstated threshold. Please define the threshold explicitly and provide an ablation over threshold values.","section":"Sec. III-E and Sec. IV-A"},{"comment":"The ablation 'Ours (w/o ROI)' removes all ROI-based candidate generation, including both frontier-based and confidence-based ROI sampling. Consequently, it cannot isolate the specific benefit of confidence-based targeted inspection, which is the core of claim (ii). A more informative ablation would include a frontier-only ROI variant or a confidence-only ROI variant, so that the contribution of the confidence signal can be separated from the general benefit of adding non-local candidate viewpoints. The comparison with Ours† changes the confidence formula but does not resolve this confound, because both Ours and Ours† use the same combined ROI selection.","section":"Sec. IV-B, ablation study"},{"comment":"The headline claim that the approach 'achieves the best performance in both rendering and mesh quality across all test scenes' is supported only by plotted curves in Fig. 4; no numerical table of means and standard deviations is provided, and no statistical significance testing is reported. With five trials per scene, a table reporting mean plus standard deviation for each method and each of the 8 scenes, together with an indication of whether the differences with respect to FBE and FisherRF are significant, would substantially strengthen the comparative claim.","section":"Sec. IV-B, Fig. 4"}],"minor_comments":[{"comment":"The exploration weight phi=1000, travel weight delta=0.5, NROI=30, and Ntotal=100 are hand-set without any sensitivity analysis. Since phi scales the exploration term in Eq. (9) relative to a confidence-based term of unknown scale, a brief sensitivity study over phi would increase confidence that the results are not brittle.","section":"Sec. IV-A"},{"comment":"The color-consistency condition avg(|I(u)-I*(u)|) > 0.5 uses a threshold of 0.5 on RGB values that appear to be in [0,1]; please clarify the color range or the intended interpretation, since this threshold seems high for typical photometric differences.","section":"Sec. III-B, Eq. (4)"},{"comment":"The real-world experiment is described qualitatively and no quantitative reconstruction metrics are reported. Adding even a small set of numbers, such as mission time, PSNR, or completeness, would make the real-world validation more convincing.","section":"Sec. IV-C"},{"comment":"There are several minor typos and wording issues, for example 'combing explicit' in Sec. II-A should be 'combining explicit', and 'the applicability our framework' in Sec. IV-C should be 'the applicability of our framework'. These do not affect the technical content.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of RA-L and the reproducibility certificate plus open-source code are commendable. The main risk is that the confidence heuristic is never validated against reconstruction error, and the low-confidence threshold is unspecified; both issues are fixable with additional experiments and reporting. I would not reject on the current evidence, but I would like to see the confidence-validation analysis and the threshold definition before accepting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ActiveGS is a solid, practically useful paper for anyone doing active reconstruction with Gaussian splatting. The new piece is the viewpoint-distribution confidence score per Gaussian primitive, plugged into a hybrid GS/voxel planning loop. That combination is genuinely new relative to the cited literature, and the confidence rendering is fast, which matters for online use. Credit is due for the careful evaluation: eight Replica scenes, five trials, means and std, and a real UAV experiment, plus released code with a reproducibility certificate.\n\nThe soft spots are real but not disqualifying. The confidence score is never checked against actual per-primitive reconstruction error, so we don't know if low-confidence primitives are actually under-reconstructed. The ablation labeled 'w/o ROI' removes both frontier-based and confidence-based ROI sampling, so it doesn't isolate the confidence contribution; the comparison against Ours† does isolate the spatial-distribution part, but not the ROI part. The threshold defining 'low-confidence' voxels is never stated, which is a reproducibility gap. And with phi=1000 as a constant weight, the exploration term likely dominates, so it's unclear how much of the gain comes from exploration vs exploitation; a sensitivity analysis on phi would help. The missing concurrent baselines (ActiveSplat, GS-Planner, HGS-planner) are acknowledged in the related work but not evaluated, which weakens the 'state-of-the-art' phrasing but is understandable given timing.\n\nNone of these flaws undermines the central result: the method works and beats the baselines they did compare against. The confidence heuristic is plausible and the experiments support the qualitative claim that it helps. I would not block publication on the missing threshold or the phi sensitivity; those are easy fixes. The bigger gap is the lack of validation of k_i against error, but that's an improvement for future work, not a fatal flaw.\n\nI would recommend sending this to serious peer review; it deserves referee time. If I were reviewing, I'd ask for the threshold, a phi sweep, and ideally a variant with frontier-only ROI to separate the confidence effect. For a robotics audience, this is a worthwhile read.","headline":"A useful, well-engineered paper on active GS reconstruction; the confidence heuristic is plausible and the experiments support the main claim, but the confidence-specific ROI benefit is not cleanly isolated and the low-confidence threshold is unspecified.","tokens_in":12421,"tokens_out":2991,"would_cite":true,"duration_ms":28501,"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":"ActiveGS uses per-primitive confidence to steer an RGB-D camera toward under-reconstructed regions, beating active reconstruction baselines.","keywords":["active scene reconstruction","Gaussian splatting","confidence modelling","next-best-view planning","hybrid map representation","RGB-D mapping","exploration and exploitation","under-reconstructed surface inspection"],"falsifier":"A direct test would be to run a full ActiveGS mission on a static scene, then compare each Gaussian primitive's confidence value with its actual reconstruction error (for example, the residual between the primitive's rendered colour and depth and ground truth at its centre); if low-confidence primitives do not systematically coincide with high-error primitives, the exploitation term loses its justification.","tokens_in":11226,"feed_emoji":"🤖","tokens_out":7953,"duration_ms":71208,"temperature":0.7,"pith_summary":"ActiveGS is a framework for active scene reconstruction: a mobile robot with an RGB-D camera decides where to look next in an unknown scene, under a fixed mission time, so that the resulting map is as accurate as possible. The paper's central claim is that a hybrid map—2D Gaussian splatting for high-fidelity rendering plus a coarse voxel map for occupancy and path planning—lets the robot both explore unknown space and inspect under-reconstructed surfaces. The key idea is a per-Gaussian confidence score computed from the spatial distribution of viewpoints that have observed each surface patch; rendering this confidence into a map gives a fast way to score candidate views. The authors report that ActiveGS outperforms frontier-based exploration, a Fisher-information GS baseline, and a NeRF-based active reconstruction baseline in both rendering PSNR and mesh completeness across the tested indoor scenes, and they demonstrate the system on a real UAV. If this holds, autonomous robots could build denser, higher-fidelity maps with the same exploration budget, without needing ground truth during planning.","feed_headline":"ActiveGS: planner that beats active reconstruction baselines","feed_subtitle":"Confidence per Gaussian primitive guides the camera to under-reconstructed surfaces, lifting PSNR and mesh completeness.","key_machinery":"The machinery is the per-primitive confidence score $k_i = \\gamma_i \\exp(\\beta_i)$, where $\\gamma_i = \\sum_j (1 - d_{ij}/d_{\\mathrm{far}})\\, n_i \\cdot v_{ij}$ accumulates distance-weighted alignment between the primitive's normal and the observed view directions, and $\\beta_i = 1 - \\|\\mu_i\\|$ with $\\mu_i$ the mean of the normalized view directions measures how widely the primitive has been seen. Because each primitive carries $k_i$, the differentiable rasterizer renders a confidence map at any candidate viewpoint in one forward pass, making viewpoint evaluation cheap and ROI sampling possible. The hybrid voxel map contributes the exploration utility $U_V$ (visible unexplored voxels) and the frontier-based ROI set, and the final view utility is $U_{\\mathrm{view}} = \\phi U_V + U_G$ with $U_G = -\\mathrm{mean}(K_i)$.","core_discovery":"On its own terms, the paper discovers and validates a way to make Gaussian-splatting maps actively steer their own refinement. It defines a confidence value for every Gaussian primitive from the directions and distances of the viewpoints that saw it—higher when the primitive has been seen up close, along its normal, and from widely spread angles—and treats rendered confidence as the signal for which parts of the scene are under-reconstructed. The view planner weights exploration (unexplored voxels visible from a candidate pose) against exploitation (negative mean confidence) and generates targeted candidate viewpoints around frontier voxels and voxels containing low-confidence primitives. The paper claims this yields the best rendering and mesh quality across all test scenes compared with frontier-based exploration, FisherRF, and NARUTO, with ablations showing that ROI-based sampling and the viewpoint-distribution confidence formulation each contribute.","pith_inferences":["The paper does not verify that a low-confidence primitive actually corresponds to a high per-primitive reconstruction error; testing that correlation directly would tell practitioners whether the exploitation term is pointing at the right surfaces.","Since the confidence proxy uses only viewpoint geometry and ignores colour or depth residual, fusing a photometric error term into $k_i$ could make the planner robust to textureless or specular surfaces, where view coverage alone is misleading.","The framework assumes perfect localization; injecting pose uncertainty into the viewpoint distribution would make confidence more honest in real deployments, which the paper itself names as future work.","The low-confidence threshold used to select ROI voxels is not specified; making it explicit and measuring sensitivity would let users trade exploration against inspection in a principled way."],"forward_implications":["Under the reported settings, a robot using ActiveGS produces higher PSNR and mesh completeness than frontier-only exploration, FisherRF, and NARUTO for the same mission time.","Because confidence maps are rendered in a feed-forward pass, candidate-view evaluation avoids the expensive per-candidate gradient computations that slow Fisher information-based planning.","The hybrid representation solves the free-space gap of Gaussian maps: the voxel layer distinguishes unknown from free space, enabling collision-free paths and frontier-based exploration.","Targeted sampling around low-confidence primitives, rather than only local random sampling, raises reconstruction quality and reduces variance across trials.","Viewpoint-dispersion-aware confidence beats a naive count-based confidence, indicating that angular spread, not just number of views, is the informative signal."],"supporting_citations":[{"why":"Supplies the Gaussian surfel representation that the GS map is built on and the differentiable rasterizer used to render colour, depth, normals, and confidence.","marker":"[4]"},{"why":"Defines 3D Gaussian splatting, the rendering paradigm from which the online GS map is adapted.","marker":"[16]"},{"why":"Provides the probabilistic voxel map update (OctoMap) used for occupancy, unexplored-region tracking, and collision-free path planning.","marker":"[9]"},{"why":"Is the NeRF-based active reconstruction baseline (NARUTO) that ActiveGS compares against for rendering and mesh quality.","marker":"[5]"},{"why":"Is the GS-based active view selection baseline (FisherRF) that uses Fisher information for viewpoint evaluation, the main GS-based comparison point.","marker":"[12]"},{"why":"Defines frontier-based exploration, which supplies the frontier voxel logic for ROI generation and serves as the FBE baseline.","marker":"[36]"}],"fun_headline_variants":["ActiveGS: Gaussian splats guide their own rebuild","Confidence-driven exploration sharpens Gaussian maps","Let low-confidence Gaussians steer the camera to fix spots","Robots build better scenes by asking their Gaussians what's missing","ActiveGS: map gaps revealed by per-Gaussian confidence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a primitive's confidence score—computed only from how many viewpoints saw it, how far away they were, and how aligned they were with its normal—actually marks the regions where the Gaussian map is poorly reconstructed, and the paper never checks this against real reconstruction error.","fun_headline_variants_meta":{"raw":{"variants":["ActiveGS: Gaussian splats guide their own rebuild","Confidence-driven exploration sharpens Gaussian maps","Let low-confidence Gaussians steer the camera to fix spots","Robots build better scenes by asking their Gaussians what's missing","ActiveGS: map gaps revealed by per-Gaussian confidence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001069,"raw_usage":{"total_tokens":4443,"prompt_tokens":874,"completion_tokens":3569,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":3490}},"tokens_in":490,"tokens_out":3569,"duration_ms":23950,"temperature":1.0,"reasoning_tokens":3490,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T05:10:04.716952+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to run a full ActiveGS mission on a static scene, then compare each Gaussian primitive's confidence value with its actual reconstruction error (for example, the residual between the primitive's rendered colour and depth and ground truth at its centre); if low-confidence primitives do not systematically coincide with high-error primitives, the exploitation term loses its justification.","supporting_citations":[{"cited_title":"High-quality surface reconstruction using gaussian surfels,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian surfel representation that the GS map is built on and the differentiable rasterizer used to render colour, depth, normals, and confidence."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering,","cited_arxiv_id":null,"evidence_quote":"Defines 3D Gaussian splatting, the rendering paradigm from which the online GS map is adapted."},{"cited_title":"Octomap: An efficient probabilistic 3d mapping framework based on octrees,","cited_arxiv_id":null,"evidence_quote":"Provides the probabilistic voxel map update (OctoMap) used for occupancy, unexplored-region tracking, and collision-free path planning."},{"cited_title":"Naruto: Neural active reconstruction from uncertain target observations,","cited_arxiv_id":null,"evidence_quote":"Is the NeRF-based active reconstruction baseline (NARUTO) that ActiveGS compares against for rendering and mesh quality."},{"cited_title":"Fisherrf: Active view selection and mapping with radiance fields using fisher information,","cited_arxiv_id":null,"evidence_quote":"Is the GS-based active view selection baseline (FisherRF) that uses Fisher information for viewpoint evaluation, the main GS-based comparison point."},{"cited_title":"A frontier-based approach for autonomous exploration,","cited_arxiv_id":null,"evidence_quote":"Defines frontier-based exploration, which supplies the frontier voxel logic for ROI generation and serves as the FBE baseline."}],"review_version":1}