{"id":"e62db1f7-2e87-414d-98fa-19aa526b4a92","arxiv_id":"2411.17942","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"An adaptive sampling plus integer programming framework improves camera placement coverage by 3-16% over random sampling and achieves the same coverage with 30-70% of the sample budget.","lead":"This paper presents an integer programming framework with two adaptive sampling strategies, E&E and TUS, for placing cameras to maximize visible space under a budget. In simulated indoor environments it reports 3-16% better coverage than random sampling, and equal coverage with only 30-70% of the sampling budget.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The flood-fill visibility step in Algorithm 3 can undercount coverage when visible regions are not connected by face-adjacent free voxels; since every reported coverage number and sampling decision is computed from this visibility map, the headline gains over random sampling may be artifacts of…","rationale":"The paper's central contribution is an algorithmic framework whose reported advantage over random sampling rests on the ability to evaluate coverage exactly enough to guide sampling. The IP formulation itself is standard (maximum k-coverage with a per-location cardinality constraint), the monotonicity argument in Proposition 4 is essentially correct, and the probability bounds in Propositions 2-3 are valid under their discretization assumptions. These parts are not where I find the load-bearing weakness.\n\nThe weakness is the flood-fill visibility oracle. Algorithm 3 is not an exact visibility algorithm: it is a BFS/FIFO expansion that assumes the set of visible, in-frustum free voxels is connected under face adjacency. The final ray-cast loop only removes false positives; it cannot add voxels that the expansion never reached. For any environment with a visible region behind a narrow opening or diagonal gap, the oracle will undercount. Since the IP objective (1) and the adaptive-sampling decisions are computed from the undercounted Vpd sets, the reported coverage improvements and budget-savings percentages could overstate the true performance of E&E relative to RS, or at least describe a different (under-resolved) problem.\n\nI agree with the reader's identification of this as the weakest assumption. I would strengthen the framing: this is not just a missing proof of connectivity; it is a correctness bug risk that a direct comparison against brute-force ray casting can settle. The paper contains no such comparison, and the claim in Section 2 that the visibility function is exact is contradicted by the algorithm's dependence on voxel-connectivity.\n\nSecondary issues (unreleased code, undefined Linear-Visibility, in-sample hyperparameter tuning) reinforce the need for the check, but I would not rest the verdict on them alone. If the brute-force comparison is clean (flood-fill and ray casting agree on the test environments), the empirical claims would be substantially supported; if not, the headline numbers need recomputation. The reader's CONDITIONAL verdict is therefore the right call, and the condition should specifically require the visibility sanity check.","tokens_in":38694,"tokens_out":6721,"duration_ms":63134,"concrete_test":"Independently reimplement the visibility oracle with brute-force ray casting: for the four benchmark rooms of Table 2 and the case-study apartment, for every sampled camera configuration (including all Table 2 trials), cast a Möller-Trumbore ray from the camera to every free-space voxel center and build Vpd_ray. Compare with Algorithm 3's Vpd_flood for each configuration; compute the symmetric difference and, if nonempty, rerun the IP with Vpd_ray and recompute Table 2's mean coverage and improvement percentages. If any visible voxel is missing and the E&E-vs-RS gaps or the 30-70% budget curves change materially, the central claim is unsupported. A minimal adversarial check is a room with a 1-voxel-wide diagonal slit between two free spaces.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 3 (Section 4.1) propagates visibility by expanding only from the camera's neighboring free-space voxels through voxels that (i) are themselves free and (ii) pass the FOV/DOF test; obstructions are filtered only after this flood-fill terminates, via one Möller-Trumbore ray cast per reached voxel. This expansion is sound only if every visible free voxel is reachable by a 6-connected chain of free voxel centers, each passing the FOV test. That condition is not proven and is false in common meshes: a visible region behind a narrow doorway, thin column, or diagonal slit can have a clear line of sight but no voxel-center chain, so the flood-fill never enqueues it. The result is systematic underestimation of Vpd, the same Vpd that feeds constraints (3) of IP (1)-(5) and drives both sampling strategies. Consequently the coverage numbers in Tables 2 and F1 and the 30-70% budget claim in Figure 8 are all computed against an unvalidated approximation, not exact visibility as asserted in Section 2 and Section 8. The paper's own appeal to sufficient granularity (Section 3.1) is not a guarantee of voxel-center connectivity; no resolution study or comparison to brute-force ray casting is provided.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a simulation-optimization framework for camera placement: the environment is voxelized, candidate camera configurations are sampled by one of two adaptive strategies (Explore and Exploit, and Target Uncovered Spaces), visibility from each candidate is computed by a flood-fill algorithm, and a maximum-k-coverage integer program is solved iteratively over the growing candidate set. The authors prove monotonic improvements of the iterative scheme, give probability and expected-sample bounds for finding an optimal configuration by random sampling, and report coverage improvements of 3.3–16.0% over random sampling, with equivalent coverage at 30–70% of the sampling budget. A case study on the Replica apartment environment extracts practical placement insights. The mathematical programming formulation and the monotonicity argument are sound; the main unresolved issues are the unvalidated flood-fill visibility approximation, an undefined function in the TUS pseudocode, and the use of tuning environments that coincide with the final evaluation environments.","tokens_in":39066,"tokens_out":4799,"duration_ms":45768,"significance":"If the central claims are supported, the paper would be a useful practical contribution: it combines a clean integer-programming formulation with adaptive column-generation-style sampling, provides monotonicity and asymptotic-consistency guarantees, and demonstrates the approach on a realistic 3D indoor model. The theoretical bounds in Section 5 are correctly derived and are independent of the empirical results, which is a strength. However, the significance of the paper depends heavily on the credibility of the reported coverage numbers, and those numbers are currently computed with a visibility routine whose accuracy is neither proven nor compared against exact ray casting. The empirical claims also rest on a tuning setup that appears to reuse the benchmark environments. These issues must be resolved before the paper's main conclusions can be accepted.","major_comments":[{"comment":"The flood-fill visibility algorithm assumes that every visible free-space voxel is reachable from the camera by a 6-connected chain of free voxel centers that each pass the field-of-view test. The text states \"As the visibility set is a connected set\" to justify omitting large parts of the frustum, but this connectivity property is not proven and is not generally true: a visible region behind a narrow doorway, a thin column, or a diagonal slit can have a clear line of sight while no chain of face-adjacent voxel centers connects it to the camera. Since Vpd is the input to constraint (3) of IP (1)–(5) and drives all coverage statistics in Tables 2 and F1 and Figure 8, the headline coverage improvements and the 30–70% sampling-budget claim are computed from an unvalidated approximation. The appeal to \"sufficient granularity\" in Section 3.1 does not guarantee voxel-center connectivity. Please provide a correctness proof under explicit conditions, or replace the flood-fill with per-voxel ray casting, or at minimum validate the flood-fill against brute-force ray casting on the exact test environments and report the discrepancy.","section":"Section 4.3, Algorithm 2"},{"comment":"The TUS pseudocode calls the auxiliary function Linear-Visibility(p, vcenter, P), but this function is never defined anywhere in the manuscript or appendices. Footnote 2 only states that the computation is encoded in Linear-Visibility; no pseudocode, formula, or reference is given. Without this definition, the TUS algorithm is not implementable and all TUS results in Section 7, Figure 8, and Appendix F.3 depend on unspecified behavior. Please add a complete pseudocode block for Linear-Visibility and state its precise output semantics, including how it handles the case where no unobstructed camera position exists along the specified direction.","section":"Section 4.3, Algorithm 2"},{"comment":"The hyperparameters for both adaptive strategies (fexploit, VPA, APA, iterations, func, supervoxel size, strict_vis_req) are selected in Appendix F.4 by evaluating on the same four environment types — medium/large, type 1/type 2 — that are later used for the final benchmark comparison in Table 2. This selection-on-test-data procedure means the reported 3.3–16.0% improvements over random sampling are not unbiased estimates of performance on new environments. Please either use a separate validation set for hyperparameter selection, or present a nested cross-validation/tuning protocol, or explicitly characterize the results as in-sample tuned performance. The current presentation overstates the generalizability of the empirical claims.","section":"Section 7.1, Table 2"},{"comment":"Each comparison in Table 2 uses only five trials per condition, and the reported ranges overlap even for headline cases: for Large 1 High, the RS range is [4583, 4814] and the E&E range is [4722, 4983], so the 3.33% mean improvement is not clearly separated from noise. No confidence intervals, paired tests, or adjustments for multiple comparisons are provided. For a stochastic simulation study, this is insufficient support for the abstract's quantitative claims of 3.3–16.0% improvement and 30–70% budget savings. Please report per-condition confidence intervals or paired significance tests, and discuss the overlap explicitly.","section":"Section 7.1, Table 2"}],"minor_comments":[{"comment":"The proof contains a typo: \"P Dk+1 \\ P Dk+1\" should be \"P Dk+1 \\ P Dk\". In addition, constraint (4) is restated with quantification over (p,d) in P Dk, whereas the original IP (1)–(5) quantifies over p in P; the notation should be aligned.","section":"Section 5.2"},{"comment":"The caption and surrounding text for the angle perturbation allowance sensitivity analysis repeatedly refer to \"voxel perturbation allowance\" (VPA) when the quantity being varied is the angle perturbation allowance (APA). Please correct the captions and text to avoid confusing the two hyperparameters.","section":"Appendix F.4.2"},{"comment":"The derivation of |D| = O(1/epsilon^2) uses a volume-counting argument over representable floating-point values. Single-precision floating-point values are not uniformly distributed over the unit sphere, so the cardinality count is heuristic; if the lemma is only meant as an order-of-magnitude modeling bound, this should be stated explicitly.","section":"Appendix E"},{"comment":"The rejection loop for pnew (\"while pnew = ∅ or pnew /∈ P\") may be problematic when perturbations are continuous and P is a discrete voxel set; clarify how pnew is projected onto P and state the termination behavior of this loop.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the core IP-plus-adaptive-sampling idea is promising, but the empirical section currently rests on an unvalidated visibility approximation, an undefined auxiliary function in TUS, and tuning on the test environments. These are fixable with additional validation, held-out tuning, and statistical reporting, so I do not recommend rejection; however, the current version should not be accepted without addressing these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a useful engineering paper, not a theory breakthrough. The contribution is two adaptive sampling schemes (E&E and TUS) wrapped around a modified maximum k-coverage IP, plus a flood-fill visibility routine. The IP formulation is standard, the monotonicity proof (Prop 4) is correct, and the probability bounds (Props 2–3) are correct but weak, as the authors admit. The empirical claims are plausible: E&E beats random sampling by 3.3–16% in their custom rooms and REPLICA case study, and reaches random-sampling coverage with 30–70% of the budget. The case study insights (cameras along wall faces rather than corners, corridor-focused placement under tight budgets) are genuinely interesting.\n\nThe soft spots are real. First, Algorithm 3's flood-fill assumes visible voxel centers form a 6-connected set. The paper asserts this and appeals to 'sufficient granularity,' but gives no proof, no resolution study, and no comparison to brute-force ray casting. In meshes with narrow or diagonal openings, visible regions can be missed, so Vpd is undercounted. Since every coverage number and sampling decision flows from that map, the headline gains are computed against an unvalidated approximation. The Section 2 claim that the visibility calculation is 'exact' is too strong. This needs either a proof under stated grid conditions, a conservative connectivity check, or a brute-force benchmark on at least one environment.\n\nSecond, Algorithm 2 calls Linear-Visibility, which is never defined; the appendix defines Rotate-Along, Sample-Spherical-Cap, and Sample-Random-Configurations, but not this function. Third, hyperparameters are tuned on the same custom room generator used for final evaluation; that is in-sample tuning, and a held-out sensitivity check would strengthen the results. Code is promised but not released.\n\nNone of this sinks the paper. The relative comparison against random sampling uses the same visibility machinery on both sides, so the ranking may survive even if absolute coverage is biased. But the exactness claim and the absolute coverage numbers cannot be taken at face value until the flood-fill is validated. I would send it to review, expecting major revision: define Linear-Visibility, add a correctness argument for Algorithm 3 or benchmark it against full ray casting, and either release code or add a reproducibility appendix.","headline":"A practical adaptive-sampling framework for camera placement with solid IP monotonicity, but the flood-fill visibility step is unvalidated and one TUS function is missing—worth refereeing with revisions.","tokens_in":39545,"tokens_out":3078,"would_cite":false,"duration_ms":30503,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C10","90C27","68U05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Adaptive sampling plus integer programming can place cameras to cover 3.3–16.0% more space than random sampling at the same budget.","keywords":["camera placement","maximum k-coverage","integer programming","adaptive sampling","visibility computation","flood-fill","surveillance","simulation optimization"],"falsifier":"Place a thin wall with a narrow doorway between two open volumes, discretize it at the paper's voxel resolution, and compare the flood-fill visible set with a brute-force ray cast from every voxel center; any missing voxel in the flood-fill output would show that the connected-component assumption fails and the reported coverage gains are not reliable for such scenes.","tokens_in":38536,"feed_emoji":"📹","tokens_out":10022,"duration_ms":82905,"temperature":0.7,"pith_summary":"The paper tries to establish that optimal camera placement in 3D environments can be made practical by iteratively generating candidate camera configurations with adaptive sampling and then solving a modified maximum k-coverage integer program to select the best network under a camera budget. The claimed payoff is that the Explore and Exploit strategy beats random sampling by 3.3–16.0% in coverage at equal budget, while Target Uncovered Spaces adds 6.9–9.1% in open, low-budget settings. Both adaptive strategies reportedly match the random-sampling baseline's coverage using only 30–70% of the sampling budget, which matters because visibility calculations dominate the computational cost. A sympathetic reader should care because the framework turns an NP-hard continuous problem into deployable placement decisions with measured gains over the standard random baseline.","feed_headline":"Adaptive sampling beats random camera placement by up to 16%","feed_subtitle":"Same coverage with 30-70% of the sampling budget means cheaper surveillance setups and faster redeployment.","key_machinery":"The load-bearing machinery is a feedback loop among three components. The optimizer is a modified maximum k-coverage integer program: each camera configuration is a set covering some free-space voxels, the program chooses at most k configurations subject to a cost budget and one-camera-per-neighborhood constraints, and the objective counts uniquely covered voxels. The visibility engine is a flood-fill algorithm that starts from voxels adjacent to the camera and expands to connected neighbors inside the field of view, only ray-casting for obstructions after the flood fill; this exploits the assumption that visible voxels form a single connected set. The sampling strategies generate new columns for the IP: E&E adds random exploratory configurations plus local perturbations of the previous optimal network, while TUS samples camera positions aimed at supervoxels with the largest uncovered-voxel counts.","core_discovery":"The central claim is that a simulation-optimization loop can solve the camera placement problem in practice: sample candidate camera configurations, compute which free-space voxels each one sees, solve a maximum k-coverage integer program to pick the best subset, and use the previous solution to guide the next round of sampling. The Explore and Exploit strategy samples randomly for breadth and then adds configurations jittered around the current best network's positions and orientations; Target Uncovered Spaces samples supervoxels by how many uncovered voxels they contain and aims cameras at those blind spots. Across synthetic rooms of two sizes, two obstruction patterns, and two camera budgets, the authors report that E&E raises mean coverage by 3.3–16.0% over random sampling and TUS raises it by 6.9–9.1% in open, low-budget conditions, while both reach the random baseline's coverage with 30–70% of the sample budget. The paper also proves that coverage improves monotonically as samples are added, and demonstrates the method on a real apartment, where the optimized network tends to favor wall-face, near-ceiling placements over corner placements.","pith_inferences":["An extension the authors do not test: the same adaptive-sampling-plus-IP loop should transfer to other coverage-maximizing sensor problems, such as thermal cameras, LiDAR, or acoustic sensors, wherever coverage can be evaluated on a discrete grid.","Because coverage improves monotonically, warm-started re-optimization can track small environmental changes without losing the incumbent solution; this makes the framework usable for rapid redeployments, though the paper does not quantify that operational gain.","The 30–70% sample-budget saving suggests a practical scheduling benefit: stopping early gives a coverage guarantee relative to the random baseline, so a deployment team could trade final coverage for speed in time-critical situations.","The flood-fill visibility assumption is the main risk: in scenes with thin walls, narrow doorways, or clutter that breaks the connected chain of visible voxels, the coverage matrix fed to the IP would be wrong, and a ray-casting audit on such scenes would show how much the reported gains depend on that assumption."],"forward_implications":["Using E&E instead of one-shot random sampling raises final mean coverage by 3.3–16.0% across the tested room sizes, obstruction patterns, and camera budgets, with the largest gains in open, low-budget scenarios.","The adaptive strategies reach the random-sampling benchmark's coverage after spending only 30–70% of the sample budget, so equivalent coverage can be achieved with substantially fewer expensive visibility computations.","The greedy heuristic finds a marginally worse solution than the integer program in 14 of 40 instances and differs by about 1%, so it can stand in for the IP on very large problems or warm-start the IP for a provably optimal finish.","In real indoor layouts, the model's solutions place cameras along wall faces near ceilings rather than at corners, and with tight budgets they favor central positions that see into corridors and doorways."],"supporting_citations":[{"why":"Defines the budgeted maximum coverage problem that the paper modifies into its camera-placement integer program.","marker":"Khuller et al., 1999"},{"why":"Analyzes the greedy approach to maximum k-coverage, which the paper uses as the optimization benchmark.","marker":"Hochbaum and Pathria, 1998"},{"why":"Provides the area-flooding technique the visibility algorithm extends to voxel grids.","marker":"Levoy, 1981"},{"why":"Supplies the fast ray-triangle intersection used for the final obstruction check in visibility calculation.","marker":"Möller and Trumbore, 1997"},{"why":"Quadric decimation reduces the real-world mesh from millions of triangles to a tractable size for the case study.","marker":"Garland and Heckbert, 1997"},{"why":"Provides the real indoor apartment environment used in the case study and placement insights.","marker":"Straub et al., 2019"}],"fun_headline_variants":["Adaptive sampling lifts camera coverage by up to 16%","Same coverage with 70% less sampling for camera networks","Explore-Exploit sampling: up to 16% better camera coverage","Adaptive search beats random for camera placement by 16%","Optimized camera placement with 30-70% less sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the voxels visible from a camera form one connected component in the voxel grid at the chosen resolution, so the flood-fill visibility routine can reach every visible voxel by stepping through visible neighbors without ray-casting each one.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive sampling lifts camera coverage by up to 16%","Same coverage with 70% less sampling for camera networks","Explore-Exploit sampling: up to 16% better camera coverage","Adaptive search beats random for camera placement by 16%","Optimized camera placement with 30-70% less sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000579,"raw_usage":{"total_tokens":2786,"prompt_tokens":1063,"completion_tokens":1723,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":679,"completion_tokens_details":{"reasoning_tokens":1636}},"tokens_in":679,"tokens_out":1723,"duration_ms":11757,"temperature":1.0,"reasoning_tokens":1636,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:41:21.693648+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Place a thin wall with a narrow doorway between two open volumes, discretize it at the paper's voxel resolution, and compare the flood-fill visible set with a brute-force ray cast from every voxel center; any missing voxel in the flood-fill output would show that the connected-component assumption fails and the reported coverage gains are not reliable for such scenes.","supporting_citations":[{"cited_title":"(1981) Area flooding algorithms","cited_arxiv_id":null,"evidence_quote":"Provides the area-flooding technique the visibility algorithm extends to voxel grids."}],"review_version":1}