{"id":"9afcddf6-5b33-483c-9f7b-352d5aa2fffe","arxiv_id":"2504.21067","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"GauSS-MI estimates the expected visual information gain of a viewpoint from per-Gaussian reliability probabilities, and an active reconstruction system using it achieves high visual fidelity with fewer frames.","lead":"This paper introduces GauSS-MI, a metric for scoring how much visual information a next camera viewpoint would add to a 3D Gaussian Splatting reconstruction, and uses it in an active reconstruction system. The goal is to let a robot choose viewpoints that improve visual fidelity rather than only geometric coverage.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (20) is not Shannon mutual information: the §IV-C/Appendix A derivation drops the r=0 (unreliable) binary state and adds a T[i] weighting without an information-theoretic justification, so the paper's central theoretical claim is unsupported.","rationale":"The paper's stated contribution is an information-theoretic criterion for next-best-view selection. If Eq. (20) is not Shannon mutual information, the central theoretical claim collapses, even though the heuristic may still work in practice. The reader's rationale already flags the missing second binary state, but the reader's stated weakest assumption focuses on the hand-crafted inverse sensor model in Eq. (7). I find the derivation error in §IV-C to be the more direct and load-bearing problem: it is internal to the mathematics and does not depend on the sensor model being realistic. Even with a perfect inverse sensor model, the derivation would not produce Shannon MI because the s=0 term is omitted and the T[i] weighting is unjustified. The hand-crafted λL and λT are a secondary concern about calibration and sensitivity; they affect the numerical value of the heuristic but do not determine whether the quantity is MI. The empirical comparisons with FisherRF and random baselines are useful evidence for the heuristic's utility, and the paper is transparent about some limitations, but the title, abstract, and contribution list claim Shannon mutual information. Therefore the reader's REJECT verdict is appropriate, and my analysis does not change it.","tokens_in":18795,"tokens_out":11402,"duration_ms":117684,"concrete_test":"Re-derive Appendix A with the binary-state sum included. For a single Gaussian, prior o=P(r)/(1-P(r)), inverse-sensor odds δ from Eq. (8), and any measurement prior P(Z), compute the exact Shannon MI I_exact = Σ_Z [P(Z) Σ_{s∈{0,1}} P(s|Z) log(P(s|Z)/P(s))] and compare it with Eq. (20), which for one Gaussian and one pixel reduces to P(Z) [oδ/(1+oδ)] log(δ(1+o)/(1+oδ)). If I_exact differs from Eq. (20) for generic o and δ (e.g., o=1, δ=3), the derivation omits a nonzero term and the quantity is not Shannon MI. A second check: repeat with two Gaussians along a ray and compare whether the T[i] weighting in Eq. (13) reproduces the exact MI of the joint reliability vector.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Eq. (14), the paper derives I(r[i];z[j]) = P(z[j]=Z)P(r[i]|z[j]=Z) log(P(r[i]|z[j]=Z)/P(r[i])). Because r[i] is defined in §IV-B as binary (reliable/unreliable), this is only the s=1 term of the Shannon mutual information. The full conditional contribution for outcome Z is P(Z) [P(1|Z) log(P(1|Z)/P(1)) + P(0|Z) log(P(0|Z)/P(0))]. The s=0 term is omitted in Eqs. (14), (15), and Appendix A; it is generally nonzero and negative when the observation increases reliability, so the computed quantity is not MI. For example, with o=1 and δ=3, the included term equals 0.152 for P(Z)=0.5, while the exact conditional MI contribution is 0.066. The later substitution L=0 leading to Eq. (19) uses only this one-sided term. Additionally, Eq. (13) inserts T[i], the cumulative transmittance, as a multiplicative weight inside the MI sum; no step in the derivation justifies treating a rendering weight as a probability weight in an information-theoretic sum. Thus GauSS-MI is an expected negative-log-probability heuristic, not Shannon mutual information. The empirical section can still evaluate its usefulness, but contribution 2 and the paper's central claim are not established by the supplied mathematics.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents GauSS-MI, a metric intended to quantify the Shannon mutual information between a 3D Gaussian Splatting (3DGS) map and a candidate viewpoint, for use in active 3D reconstruction. The authors introduce a probabilistic reliability model for each Gaussian, update it with an inverse sensor model based on rendering loss, and derive a mutual information expression that is then used to select next best views in an active reconstruction system. The system is evaluated in simulation (three Flightmare scenes) and in real-world experiments with a robotic arm, comparing against FisherRF, FUEL, NARUTO, and random selection. The reported results show improved PSNR and efficiency, and the implementation is open-sourced.","tokens_in":19156,"tokens_out":4523,"duration_ms":46948,"significance":"If the central theoretical claim were correct, GauSS-MI would be a valuable real-time uncertainty metric for active 3DGS reconstruction, bridging information-theoretic view planning and high-fidelity visual reconstruction. The paper has notable strengths: it ships open-source code, presents extensive experiments including real-world validation, reports fast runtime (5.55 ms per evaluation), and includes uncertainty quantification comparisons (sparsification plots and AUSE) against FisherRF. However, the derivation of the mutual information contains a serious mathematical gap: Eq. (14) omits the unreliable-state term of the binary reliability variable, and Appendix A inverts the posterior odds. Consequently, the quantity computed in Eq. (20) is not Shannon mutual information as claimed, but a weighted heuristic. The core contribution of the paper is therefore not established by the supplied mathematics, despite the suggestive empirical results.","major_comments":[{"comment":"The mutual information between a binary reliability variable r[i] and an observation z[j] is defined as a sum over both outcomes s=0 and s=1. Equation (14) computes only the s=1 term, P(z[j]=Z)P(r[i]|z[j]=Z) log(P(r[i]|z[j]=Z)/P(r[i])), omitting the s=0 term P(z[j]=Z)P(¬r[i]|z[j]=Z) log(P(¬r[i]|z[j]=Z)/P(¬r[i])). This omitted term is generically nonzero; for instance, with prior odds o=1, inverse sensor odds δ=3, and P(Z)=0.5, the included term is 0.152 while the exact conditional MI contribution is 0.066. Therefore the quantity in Eqs. (14)-(15) and ultimately Eq. (20) is not the Shannon mutual information of the described probabilistic model. This invalidates the paper's central claim that GauSS-MI measures Shannon mutual information.","section":"§IV-C, Eq. (14)"},{"comment":"The derivation of the information gain function contains an algebraic error. With the definition o = P/(1-P), the correct inversion is P = o/(1+o). Equation (26) states P(r|Z1:k-1) = (1+o)/o, which is the reciprocal of the correct probability. Equation (27) similarly inverts the posterior odds. Although the final expression f(δ,o) in Eq. (28) may be consistent with the one-sided term if one starts from this inverted definition, the derivation as written is incorrect and the error is not acknowledged. A correct derivation starting from the proper posterior probability would produce a different expression unless the omitted binary outcome is also included.","section":"Appendix A, Eqs. (26)-(27)"},{"comment":"The cumulative transmittance T[i] is inserted as a multiplicative weight inside the summation for mutual information. In information theory, mutual information is a sum of terms each involving a probability mass or density; T[i] is a rendering weight, not a probability of a map state. No step in the derivation justifies replacing the contribution of Gaussian i by T[i] times a per-Gaussian information gain. This is a heuristic weighting rather than a consequence of ShannNMI. The paper provides no argument for why a rendering weight should act as a probability weight in an information-theoretic sum, and the independence assumptions in the model do not yield this structure.","section":"§IV-C, Eq. (13)"},{"comment":"The expected GauSS-MI assumes L=0 for novel viewpoints, i.e., perfect future rendering, so that δ^{-1}=0 and the information gain reduces to -log(P(r[i])). This assumption removes any dependence on the candidate observation's predicted error and reduces the metric to a self-information of the current reliability estimate, which is not a predictive mutual information. Additionally, the sensor model P(z[j]|M[j]) in Eq. (18) is referenced to [10] but never specified, so the metric is not fully defined. Finally, the termination threshold φ=75% in Eq. (24) is set after observing the PSNR curve in §VI-F, as the authors themselves state in §VIII; this post hoc selection limits the claims about autonomous termination. These issues together mean that the empirical results, while suggestive of a useful heuristic, do not validate the stated theoretical contribution.","section":"§IV-D, Eq. (19) and §V-C / §VIII"}],"minor_comments":[{"comment":"The row labeled 'hyperparameter on cumulative transmittance' lists 'λL 7.0' but the notation in the text and Eqs. (7)-(10) uses λT for this hyperparameter; the table should read λT = 7.0.","section":"Table II"},{"comment":"The notation σ(n)_f = 0 for n = 1,2,3 is unclear; it would be clearer to state that the higher derivatives of the viewpoint state are constrained to zero, the same for all components.","section":"Eq. (22)"},{"comment":"The figure caption says 'the hyperparameters λ are omitted for simplicity,' but the axis and curves would be easier to interpret if the chosen λ values (λL=1.7, λT=7.0) were stated in the caption or legend.","section":"Figure 2"},{"comment":"The inverse sensor model is introduced as P(r[i]|Zk) = 1 / ((λ_L L_k)^{λ_T T[i]} + 1), but the derivation from sensor noise characteristics is not given. The paper would benefit from either a justification from a physical noise model or an explicit statement that this is a heuristic choice.","section":"§IV-B, Eq. (7)"}],"recommendation":"reject","confidential_remarks":"The empirical system appears to work well and the results are suggestive, but the paper's central theoretical claim that GauSS-MI is Shannon mutual information is not supported by the derivation. The errors in Eq. (14) and Appendix A are not typographical; they change the meaning of the metric. I would be open to a revised manuscript that reframes GauSS-MI as a heuristic information-gain criterion, with the theory corrected or removed, but as submitted the main contribution is not valid. The rejection is based on this mathematical issue, not on the experimental scope or novelty."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Xie et al. build a fast, practical next-best-view selector for active 3D Gaussian Splatting reconstruction, and they back it with a real system: online mapper, per-Gaussian reliability updates, view/motion planning, and termination logic, tested against FisherRF, FUEL, and NARUTO in simulation plus two real scenes. The runtime advantage is credible (5.55 ms vs 11.66 ms per view), and the AUSE sparsification results suggest the uncertainty score tracks actual rendering error. The per-Gaussian reliability model with the loss-based inverse sensor model is a reasonable construction, and the final score in Eq. (20) is a cheap and sensible heuristic for guiding a camera to reduce visual uncertainty. As a systems paper, this is a solid contribution with open-sourced code.\n\nThe soft spot is the theoretical packaging. The paper claims GauSS-MI is Shannon mutual information, and that claim does not survive contact with the equations. Eq. (14) only keeps the s=1 term for a binary reliable/unreliable Gaussian; the s=0 term is dropped even though it is generally nonzero and negative, so the sum is not MI. Appendix A then inverts posterior odds: Eq. (26) gives P(r|Z) = (1+o)/o, which is greater than 1 for any finite o; it should be o/(1+o). That is not a cosmetic typo, it feeds directly into the derived f(delta,o). The T[i] weighting introduced in Eq. (13) makes sense as a rendering weight but has no information-theoretic justification as a probability weight. So the mathematical derivation does not establish the central claim. The paper would be acceptable if it honestly framed GauSS-MI as an expected-entropy heuristic; the experiments could still be evaluated on their own terms.\n\nOther concerns are real but less severe. The hyperparameters (lambda_L, lambda_T, w_I, w_J, tau, phi) are hand-set with no sensitivity analysis. The termination threshold phi is chosen after observing the PSNR curve, which is post hoc. The sensor model P(z|M) is referenced but not specified. Comparisons are single runs without error bars. These are fixable with reporting and ablation, but they limit how strongly the empirical results can be read.\n\nWho should read this? Robotics and scene-capture people who want a deployable active-reconstruction view selector will get value from the system and the heuristic. Someone looking for a rigorous information-theoretic criterion will be misled by the current presentation. I would send it to peer review because the empirical work deserves airing and the math is revisable, but I would flag the MI framing as load-bearing and ask the authors to either correct the derivation or explicitly downgrade the claim to a heuristic.","headline":"A useful active-reconstruction heuristic whose Shannon-MI framing is not supported by the math; the system and experiments are worth a look, but the central derivation needs substantial correction.","tokens_in":19663,"tokens_out":2100,"would_cite":true,"duration_ms":22122,"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":"This paper claims that the visual information a new viewpoint would contribute to a 3D Gaussian Splatting map can be computed in real time as a Shannon mutual information, and that an active reconstruction system selecting views by this…","keywords":["active 3D reconstruction","next best view selection","3D Gaussian Splatting","Shannon mutual information","uncertainty quantification","view planning","visual fidelity","real-time mapping"],"falsifier":"Run the same active reconstruction pipeline on a fixed set of scenes with $\\lambda_L$ and $\\lambda_T$ swept over a grid (for example 0.5 to 5.0 and 1 to 15) while keeping everything else constant, and record both the rank order of candidate viewpoints and the final PSNR; if the chosen views and final quality change strongly across the grid, the reliability model is not carrying the claimed information.","tokens_in":18570,"feed_emoji":"🎥","tokens_out":7185,"duration_ms":68891,"temperature":0.7,"pith_summary":"This paper claims that the visual value of a candidate camera viewpoint for a 3D Gaussian Splatting map can be computed in real time as a Shannon mutual information between the map and the would-be observation. The key move is to treat each Gaussian as a Bernoulli random variable that is reliable for rendering, update that reliability from the residual between rendered and observed images, and then rasterize the resulting information gain over candidate pixels. If this works, active reconstruction systems can pick the next best view by directly reducing visual uncertainty instead of only chasing geometric completeness. The paper further claims that a complete active system built on this metric achieves higher visual fidelity and better efficiency than volumetric or radiance-field baselines in simulated and real-world scenes.","feed_headline":"A 5-millisecond score picks the next best 3D view","feed_subtitle":"Gaussian-by-Gaussian render error becomes a mutual-information score that chooses the views worth capturing.","key_machinery":"The load-bearing mechanism is the Gaussian Splatting Shannon Mutual Information identity, Eq. (20): for a candidate viewpoint the expected information gain is obtained by rasterizing $-\\log P(r[i])$, the per-Gaussian surprise that the current model is reliable, weighted by the cumulative transmittance $T[i]$ along each ray and by a per-pixel measurement prior derived from rendered luminance. The update side is carried by a Bayesian log-odds recursion $l_{1:k}=l_{1:k-1}-\\lambda_T T[i]\\log(\\lambda_L L_k)$, which folds the latest render loss into each Gaussian's reliability without storing past images. Because evaluation is one rasterization pass, the score's cost is $O(N_p N_g N_c)$ per decision step and does not grow with the number of already seen views.","core_discovery":"The central discovery is an explicit, closed-form mutual-information score for 3D Gaussian Splatting maps. For each Gaussian the paper maintains the log odds of the event that it renders correctly, updated by a binary Bayesian filter whose inverse sensor model maps a per-pixel color-depth loss $L_k$ and the Gaussian's transmittance $T[i]$ into a reliability probability $P(r[i]|Z_k)=1/((\\lambda_L L_k)^{\\lambda_T T[i]}+1)$. For an unobserved viewpoint, the expected Shannon mutual information between the map and the observation then collapses to Eq. (20), $I(r;z)=\\sum_j P(z[j]|M[j])\\sum_{i\\in N[j]}-T[i]\\log P(r[i])$, where $P(z[j]|M[j])$ is a sensor prior from the rendered luminance and the sum runs over Gaussians along each pixel ray. The paper argues that maximizing this quantity selects the next best view that most reduces expected map entropy, and that the same quantity provides a principled termination criterion.","pith_inferences":["The inverse sensor model in Eq. (7) is a hand-chosen function whose hyperparameters $\\lambda_L=1.7$, $\\lambda_T=7.0$ are tuned constants; fitting them to the actual Poisson-Gaussian noise statistics of the camera would turn the reliability probability into a calibrated sensor model instead of a heuristic.","The independence assumption on Gaussian reliabilities ignores correlation between overlapping splats; extending the Bernoulli prior to a spatial Markov or covariance model could improve uncertainty ranking in texture-rich regions without changing the rasterization logic.","The same derivation maps onto other differentiable scene representations: any renderer that can produce a per-primitive transmittance and a per-pixel loss gives an analogous rasterized mutual-information score, so the criterion is not specific to Gaussian splatting.","Because the metric directly reports residual visual uncertainty, its value could serve as an online stopping signal or a confidence map for downstream tasks such as robotic inspection, not just as a view-ranking score."],"forward_implications":["Next-best-view selection can target visual fidelity directly: views are ranked by expected reduction of rendering uncertainty, not by how much unseen geometry they reveal.","The per-step computation stays nearly constant as the map grows, since GauSS-MI evaluates only candidate views and folds prior observations into per-Gaussian probabilities during the map update.","A simple, transparent termination rule becomes possible: stop when a threshold fraction (75% in the paper) of Gaussians reaches average reliability above 0.7; the reported PSNR saturates near that point.","On fixed frame budgets the metric picks views that give higher PSNR than a random policy and is competitive with or better than a Fisher-information baseline, while running at roughly 182 fps in the reported setup.","In full active reconstruction, the system reaches visual quality comparable to a NeRF-based active method with far fewer captured frames and shorter paths."],"supporting_citations":[{"why":"Basis for expressing total mutual information as a sum over measurement beams; supplies the Shannon mutual information framework the metric derives from.","marker":"[18]"},{"why":"Defines 3D Gaussian Splatting rendering and the transmittance/blending equations the probabilistic model builds on.","marker":"[19]"},{"why":"Provides the online Gaussian-splatting SLAM used as the reconstruction backbone in the active system.","marker":"[24]"},{"why":"The Fisher-information view-selection baseline that GauSS-MI is compared against in efficiency and uncertainty quantification.","marker":"[15]"},{"why":"Supplies the Poisson-Gaussian sensor noise model used to build the measurement prior $P(z[j]|M[j])$.","marker":"[10]"},{"why":"The quadrotor simulator used for the simulation scenes and evaluation.","marker":"[32]"},{"why":"A volumetric frontier/exploration baseline used for the active-reconstruction comparison.","marker":"[43]"},{"why":"A NeRF-based active reconstruction baseline used for visual-quality and efficiency comparison.","marker":"[9]"}],"fun_headline_variants":["5-ms Shannon score chooses the next 3D viewpoint","Gaussian-level uncertainty becomes a 5-millisecond view score","Closed-form mutual information metric for active 3D scanning","Shannon mutual information quantifies the best next view in 3D","Real-time next-view selection via per-Gaussian mutual information"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Eq. (7), a hand-chosen inverse sensor model with fixed hyperparameters $\\lambda_L=1.7$ and $\\lambda_T=7.0$: if this function does not actually describe how likely a Gaussian is to be reliable for rendering, the whole uncertainty model and the mutual-information score built on it are not measuring what they claim.","fun_headline_variants_meta":{"raw":{"variants":["5-ms Shannon score chooses the next 3D viewpoint","Gaussian-level uncertainty becomes a 5-millisecond view score","Closed-form mutual information metric for active 3D scanning","Shannon mutual information quantifies the best next view in 3D","Real-time next-view selection via per-Gaussian mutual information"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001326,"raw_usage":{"total_tokens":5416,"prompt_tokens":986,"completion_tokens":4430,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":4343}},"tokens_in":602,"tokens_out":4430,"duration_ms":34404,"temperature":1.0,"reasoning_tokens":4343,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:21:02.891266+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same active reconstruction pipeline on a fixed set of scenes with $\\lambda_L$ and $\\lambda_T$ swept over a grid (for example 0.5 to 5.0 and 1 to 15) while keeping everything else constant, and record both the rank order of candidate viewpoints and the final PSNR; if the chosen views and final quality change strongly across the grid, the reliability model is not carrying the claimed information.","supporting_citations":[{"cited_title":"On mutual information-based control of range sensing robots for mapping applications","cited_arxiv_id":null,"evidence_quote":"Basis for expressing total mutual information as a sum over measurement beams; supplies the Shannon mutual information framework the metric derives from."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Defines 3D Gaussian Splatting rendering and the transmittance/blending equations the probabilistic model builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the online Gaussian-splatting SLAM used as the reconstruction backbone in the active system."},{"cited_title":"Fisherrf: Active view selection and mapping with radiance fields using fisher information","cited_arxiv_id":null,"evidence_quote":"The Fisher-information view-selection baseline that GauSS-MI is compared against in efficiency and uncertainty quantification."},{"cited_title":"Practical poissonian-gaussian noise modeling and fitting for single-image raw-data","cited_arxiv_id":null,"evidence_quote":"Supplies the Poisson-Gaussian sensor noise model used to build the measurement prior $P(z[j]|M[j])$."},{"cited_title":"Flightmare: A flexible quadrotor simulator","cited_arxiv_id":null,"evidence_quote":"The quadrotor simulator used for the simulation scenes and evaluation."},{"cited_title":"FUEL: Fast uav exploration using incremen- tal frontier structure and hierarchical planning","cited_arxiv_id":null,"evidence_quote":"A volumetric frontier/exploration baseline used for the active-reconstruction comparison."},{"cited_title":"Naruto: Neural active reconstruction from uncer- tain target observations","cited_arxiv_id":null,"evidence_quote":"A NeRF-based active reconstruction baseline used for visual-quality and efficiency comparison."}],"review_version":1}