{"id":"0fdd5fd3-ca20-49bc-864e-94f1b7438a3a","arxiv_id":"2507.17519","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A terrain-aware wrapper for 2D multi-UAV coverage planners, demonstrated as DARP-3D, improves 3D reconstruction accuracy by adjusting waypoint altitude and camera pitch and yaw from a pre-scanned model.","lead":"A team from Democritus University of Thrace and CERTH built a modular add-on that takes ordinary 2D drone flight paths and adjusts altitude and camera angle to follow terrain, improving 3D reconstruction of buildings and rocks. Tests in four simulated environments and one real DJI flight show better reconstruction where terrain has vertical features, with little change on flat ground.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is conditional on an unquantified pre-scan model quality; without a sensitivity analysis, the reported gains may not survive sparse or noisy initial scans.","rationale":"The reader's weakest assumption correctly identifies pre-scan fidelity, and I agree this is the most load-bearing condition for the central claim. The method cannot point cameras at surfaces that are absent from the pre-scan, and the feedback loop (pre-scan → viewpoints → reconstruction → comparison) could in principle propagate errors. The paper's own data show one testbed regression (Pallet Pile), and the conclusion acknowledges that low-elevation scenarios yield comparable results, so the 'consistently' in the abstract is an overstatement. However, the paper has genuine strengths: the algorithm is simple and modular, the simulation results show large F1 improvements in three of four testbeds, and an open-source implementation and a real-world flight are provided. These are independent supports. The main gap is the missing sensitivity analysis of the pre-scan, which is the one condition that could invalidate the mechanism in practice. This does not warrant rejection, but it does warrant the conditional verdict already given by the reader, because the central claim is not established outside a narrow, favorable regime. A single controlled test varying pre-scan quality would settle the issue.","tokens_in":8586,"tokens_out":9414,"duration_ms":101549,"concrete_test":"Run the Cathedral testbed with pre-scan models generated from 20, 50, 100, and 200 images (the real-world pre-scan used 31) and with synthetic noise (0.5 m and 1 m Gaussian) added to the pre-scan point cloud. Recompute the F1-score at 5 cm and 10 cm for DARP-3D and baseline DARP. If the DARP-3D improvement over DARP shrinks or reverses under sparse or noisy pre-scans, the central claim is conditional on a high-quality pre-scan; if improvements persist, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core mechanism (Algorithm 2) selects camera yaw/pitch toward a point in a pre-scanned 3D model that is closest to the mean altitude of points found in a downward hemisphere. The quality of that pre-scan is therefore a load-bearing prerequisite: if the initial model is sparse, biased, or noisy, the hemisphere search expands (r ← r + Δr) until it finds distant or off-surface points, and the refined viewpoints may aim at incorrect geometry, reinforcing rather than correcting pre-scan errors. Section II explicitly states the method 'leverages a pre-scanned 3D model,' and the real-world experiment used only 31 images for the pre-scan (Section III-C). The evaluation, however, never varies pre-scan density, noise, or viewpoint distribution; all quantitative results use a single pre-scan per testbed. The Pallet Pile regression in Table I (F1 -0.11 at 10 cm, precision -4.55 at 5 cm) demonstrates that the method is not uniformly beneficial and is sensitive to scene geometry and parameters, but the paper does not explore how sensitive it is to the pre-scan itself. Thus, the central claim that the approach 'consistently captures improved 3D reconstructions' is only supported under favorable, untested pre-scan conditions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a modular post-processing stage for 2D multi-UAV coverage path planners that adjusts waypoint altitudes using a pre-scanned point cloud and computes camera yaw/pitch toward a heuristic target (the point closest to the mean altitude in a downward hemisphere). The method is instantiated as DARP-3D on top of the DARP planner and evaluated in four AirSim/OpenDroneMap simulated testbeds plus a qualitative real-world DJI flight. The authors report large F1-score improvements at 5 cm and 10 cm cloud-to-cloud thresholds on the Rock, Cathedral, and Village testbeds, near-baseline performance on the low-relief Pallet Pile testbed, and a real-world demonstration that fills visible gaps.","tokens_in":8830,"tokens_out":5238,"duration_ms":59872,"significance":"If supported, this is a practically valuable and low-complexity wrapper that could upgrade existing commercial 2D planners with terrain awareness. The open-source implementation, use of standard tools (ODM, AirSim, CloudCompare), and a real-world flight test are explicit strengths that increase reproducibility and practical relevance. However, the evaluation is single-run with no error bars, the method's dependence on pre-scan quality is never characterized, and the Pallet Pile result directly contradicts the abstract's 'consistently captures improved' claim. These issues currently prevent the central claim from being fully supported, although they appear addressable within the manuscript's scope.","major_comments":[{"comment":"Table I reports a single point estimate per testbed with no error bars, no repeated trials, and no statistical tests. Because the central claim is that DARP-3D 'consistently captures improved 3D reconstructions,' the evaluation must show that the observed gains are not due to run-to-run variation in AirSim/ODM or ICP alignment; add at least three independent runs with different random seeds and report mean/std or paired statistical tests.","section":"§III-B, Table I"},{"comment":"The method's key camera-angle mechanism depends on the pre-scanned model, but the paper never varies pre-scan quality. Algorithm 2 grows the hemisphere radius until some point is found, so with a sparse or noisy pre-scan (the real-world pre-scan uses only 31 images) it can select a distant or off-surface target and reinforce initial model errors. A sensitivity study varying pre-scan density, noise, and viewpoint distribution is needed to establish that the reported gains are robust rather than conditional on favorable pre-scan conditions.","section":"§II-B, Algorithm 2; §III-C"},{"comment":"The Pallet Pile testbed contradicts the abstract's 'consistently captures improved' statement: precision drops by 4.55 percentage points at 5 cm and F1 drops by 0.11 percentage points at 10 cm. The text acknowledges the exception, but the abstract and conclusions should be qualified, or the paper should explicitly identify the terrain regimes where the method is and is not expected to help.","section":"Table I; abstract"},{"comment":"The equal-image-count fairness statement is ambiguous. If the DARP-3D final model is reconstructed only from the 200 second-stage images while the DARP baseline uses 250 images, the comparison is not equal in reconstruction input; if the 50 pre-scan images are included, DARP-3D receives privileged prior information not available to the baseline. Please state exactly which image set produces each compared model and justify why the comparison is fair.","section":"§III-B"},{"comment":"The method introduces several free parameters (z_offset, tol/Δtol, r0/Δr, step, δz) and provides no sensitivity analysis. The Pallet Pile discussion itself attributes its regression to 'sensitivity to parameters definition.' Without varying these parameters, it is unclear whether the reported improvements are robust or are the result of favorable per-testbed tuning.","section":"§III-B; Algorithms 1 and 2"}],"minor_comments":[{"comment":"There are frequent typographical inconsistencies such as 'UA V' instead of 'UAV', 'W GS84' instead of 'WGS84', and 'xof f set' instead of 'x_offset' in the Figure 3 caption.","section":"Throughout"},{"comment":"The claim of 'O(n) linear time complexity' is not fully supported because Algorithm 2 includes a while-loop that expands the search radius and each SphereSearch has a KD-tree query cost; please state the assumptions and the worst-case complexity more carefully.","section":"§II-B"},{"comment":"The real-world validation is qualitative only; adding a cloud-to-cloud comparison for the real-world models, even with the same thresholds, would substantially strengthen the demonstration of practical effectiveness.","section":"§III-C"},{"comment":"The explanation of precision and recall under cloud-to-cloud thresholds is confusing ('the higher the value the more lenient the measurement becomes'); please define the two directions explicitly and state which cloud is the reference.","section":"§III-B"},{"comment":"The 5 cm and 10 cm thresholds are admittedly chosen arbitrarily; a short sensitivity check with additional thresholds (e.g., 2 cm and 20 cm) would make the reader more confident that the conclusions are not threshold artifacts.","section":"§III-B"}],"recommendation":"major_revision","confidential_remarks":"This is a promising systems-oriented paper that fits a venue valuing practical integration and reproducible evaluations. The main concerns are not with novelty or the core idea but with the strength of the evaluation evidence: single-run results, uncharacterized pre-scan sensitivity, and a negative testbed that contradicts the headline claim. These are fixable with additional experiments and a more careful fairness description."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper; here's where I land. It is a solid, modest engineering contribution: a modular wrapper that takes a standard 2D coverage planner (DARP), adjusts waypoint altitude to follow terrain via a KD-tree search over a pre-scanned point cloud, and aims the camera at the point nearest the local mean altitude inside a downward hemisphere. The result, DARP-3D, beats fixed-altitude DARP by large F1 margins on three of four simulated testbeds and shows a visible improvement in a qualitative real-world DJI flight. The open-source release is real evidence and makes the work reproducible, and the comparison against DARP with a fixed 60-degree gimbal is a practical baseline many papers omit.\n\nWhat's genuinely new is modest: the specific recipe of altitude-following plus hemisphere-based camera targeting as a planner-agnostic wrapper, instantiated and validated. I have no problem calling that a legitimate incremental result. The body is also honest about scope — the conclusions explicitly say low-relief scenes like Pallet Pile show no benefit, even though the abstract's 'consistently captures improved reconstructions' overstates it (precision there actually drops 4.55pp at 5cm).\n\nThe soft spots, in proportion. First, evaluation: one run per testbed, no repeats, no error bars, no statistics. With F1 differences ranging from flat to +44 points, a few repeated trials would separate signal from noise. Second, and more load-bearing: the method depends on the pre-scan model being a decent proxy for true geometry, and that dependence is never stress-tested. The hemisphere search expands until it finds points; with a sparse 31-image pre-scan (used in the real flight), the target can be a distant or off-surface point, and nothing in the paper shows how the method degrades as pre-scan density or noise worsens. This is why the central claim stays conditional. Note the comparison itself is not circular — the final reconstruction is measured against ground truth, not the pre-scan — but the method's success is still conditioned on that first model. Third, the 'any planner' generality is a design claim, not a demonstrated one; only DARP is tested.\n\nMinor: thresholds and the 50/200 image split are arbitrary, but they are matched across methods, which is the fair part.\n\nWho this is for: practitioners building multi-UAV mapping workflows who want a drop-in terrain-aware upgrade to 2D planners. It deserves a serious referee; the concerns are addressable with more experiments and a sensitivity analysis, not structural redesign.","headline":"A practical, modest wrapper that turns a 2D coverage planner into a terrain-aware 3D mapper, with an open-source implementation and honest scope reporting — but single-run evaluation and an untested dependence on pre-scan quality keep the central claim conditional.","tokens_in":9419,"tokens_out":3926,"would_cite":false,"duration_ms":41829,"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":"Terrain-aware refinement of 2D coverage paths yields 3D reconstructions with F1 gains up to 44 points on vertical testbeds, matching baseline on flat terrain.","keywords":["multi-UAV coverage path planning","terrain-aware path planning","3D reconstruction","camera viewpoint optimization","DARP-3D","aerial photogrammetry","pre-scanned 3D model"],"falsifier":"Run DARP-3D on a testbed whose pre-scan deliberately omits a major vertical or occluded face, then compare its 5 cm and 10 cm cloud-to-cloud F1 recall on that face against baseline DARP using the same total image count; if the occluded face's recall does not improve, the claim that pre-scanned geometry guides better viewpoints is falsified.","tokens_in":8349,"feed_emoji":"🚁","tokens_out":7926,"duration_ms":81526,"temperature":0.7,"pith_summary":"This paper tries to establish that a standard two-dimensional coverage path planner can be upgraded for three-dimensional mapping simply by post-processing its waypoints: resetting altitude to follow the terrain and aiming each camera at nearby structure in a rough pre-scan. The authors wrap the well-known DARP planner this way and call the result DARP-3D. Across four simulated testbeds, DARP-3D improves cloud-to-cloud F1 scores over baseline DARP on the three with prominent vertical features, with gains up to 44.26 percentage points at the 5 cm threshold, while staying essentially equal on flat terrain. A real-world flight on commercial multicopter hardware produced visibly more complete models, especially in previously empty areas. If this holds, existing commercial 2D planning tools could gain 3D reconstruction capability without replacing their core coverage logic.","feed_headline":"Terrain-aware drone flights lift 3D model accuracy by up to 44 points","feed_subtitle":"Re-aiming camera angles from a rough pre-scan lets a standard 2D planner capture vertical structures far better.","key_machinery":"DARP (Divide Areas for Optimal Multi-Robot Coverage Path Planning) supplies the baseline 2D paths. The load-bearing mechanism is the viewpoint refinement step, built on two searches: a KD-tree search finds terrain height under each waypoint and sets altitude to that height plus a safety offset, and a second search expands a downward-facing hemisphere of radius $r$ around the waypoint until it finds points from the pre-scanned point cloud, then computes yaw and pitch toward the point whose height is closest to the average height of all points found. This aim-at-the-average-height-point rule is what converts nadir views into oblique views of vertical structure, and the paper reports it runs in linear time per path.","core_discovery":"The central claim is that the reason 2D planners produce poor 3D reconstructions is not the coverage logic but the fixed altitude and downward-only camera, and that both can be corrected externally. DARP-3D uses the pre-scanned 3D model to rewrite each waypoint's altitude to the local terrain elevation plus a user offset, inserting extra waypoints where elevation changes exceed a margin, then searches a downward-facing hemisphere of growing radius around each waypoint to find scene points and aims the camera at the point closest to the cluster's average altitude. Aiming obliquely at vertical surfaces lets photogrammetry reconstruct faces that nadir imagery misses. The paper reports that DARP-3D beats DARP and a fixed 60-degree gimbal variant on the Rock, Cathedral, and Village testbeds at 5 cm and 10 cm thresholds, and roughly matches them on the flat Pallet Pile testbed.","pith_inferences":["Our inference: the same refinement could be applied to any downward-looking camera platform, not just rotorcraft, as long as the camera cannot point above the platform; otherwise the hemisphere constraint would need to change.","Our inference: relying on a pre-scan creates a possible bias loop, because surfaces the pre-scan misses will rarely be aimed at; feeding the improved model back into the pipeline for a second iteration is a natural next test.","Our inference: the paper's own future-work note points to abrupt yaw and pitch changes between waypoints causing overlap gaps; measuring whether smooth angle interpolation improves the 10 cm F1 score more than the 5 cm score would help separate overlap failures from pure coverage failures."],"forward_implications":["The wrapper is planner-agnostic: any fixed-altitude 2D coverage planner that outputs waypoints can receive the same altitude and camera-angle refinement.","The improved reconstruction comes at the same total image count as the baseline, because the first-stage pre-scan uses far fewer images and is not part of the final model.","The largest gains appear on terrain with significant vertical features; users mapping flat, low-relief areas should expect results similar to the baseline rather than large improvements.","The refinement runs offline with linear complexity and completed in under two minutes in dense test configurations, making it practical for field mission planning.","The refined plans can be loaded onto commercial drone hardware through standard mission APIs, as the real-world test demonstrates."],"supporting_citations":[{"why":"Supplies the baseline 2D multi-robot coverage planner whose paths the extension modifies.","marker":"[9]"},{"why":"Provides the open photogrammetry pipeline that turns captured images into the initial rough model and the final comparison models.","marker":"[28]"},{"why":"Supplies the drone simulator used for all four synthetic testbed flights.","marker":"[29]"},{"why":"Bridges DARP's planned paths to the simulator so flights can be executed and images captured.","marker":"[30]"},{"why":"Performs point-cloud sampling and alignment used to compare reconstructed models with ground truth.","marker":"[31]"},{"why":"Defines the cloud-to-cloud distance metric from which the reported precision, recall, and F1 scores are computed.","marker":"[36]"},{"why":"Provides the commercial multi-UAV mission planning platform used in the real-world flight test.","marker":"[25]"}],"fun_headline_variants":["Terrain-aware altitude and camera fix 2D UAV planners","Modular terrain-aware upgrade for existing 2D path planners","DARP-3D: terrain-aware UAV coverage for better 3D models","Pre-scan terrain to aim drone camera for vertical surfaces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the pre-scanned rough 3D model is faithful enough that aiming cameras at points in it improves reconstruction of the true scene; a sparse, biased, or erroneous first scan would cause the refined viewpoints to reinforce the model's mistakes.","fun_headline_variants_meta":{"raw":{"variants":["Terrain-aware altitude and camera fix 2D UAV planners","Modular terrain-aware upgrade for existing 2D path planners","DARP-3D: terrain-aware UAV coverage for better 3D models","Pre-scan terrain to aim drone camera for vertical surfaces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00026,"raw_usage":{"total_tokens":1566,"prompt_tokens":897,"completion_tokens":669,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":595}},"tokens_in":513,"tokens_out":669,"duration_ms":8163,"temperature":1.0,"reasoning_tokens":595,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:46:03.387989+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DARP-3D on a testbed whose pre-scan deliberately omits a major vertical or occluded face, then compare its 5 cm and 10 cm cloud-to-cloud F1 recall on that face against baseline DARP using the same total image count; if the occluded face's recall does not improve, the claim that pre-scanned geometry guides better viewpoints is falsified.","supporting_citations":[{"cited_title":"DARP: Divide areas algorithm for optimal multi- robot coverage path planning","cited_arxiv_id":null,"evidence_quote":"Supplies the baseline 2D multi-robot coverage planner whose paths the extension modifies."},{"cited_title":"OpenDroneMap/ODM GitHub Page 2020; https:// github.com/OpenDroneMap/ODM","cited_arxiv_id":null,"evidence_quote":"Provides the open photogrammetry pipeline that turns captured images into the initial rough model and the final comparison models."},{"cited_title":"AirSim: High-Fidelity Visual and Physical Simulation for Autonomous Vehicles","cited_arxiv_id":null,"evidence_quote":"Supplies the drone simulator used for all four synthetic testbed flights."},{"cited_title":"Improving time and energy efficiency in multi-UA V coverage operations by optimizing the UA Vs’ initial posi- tions","cited_arxiv_id":null,"evidence_quote":"Bridges DARP's planned paths to the simulator so flights can be executed and images captured."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Performs point-cloud sampling and alignment used to compare reconstructed models with ground truth."},{"cited_title":"https://www","cited_arxiv_id":null,"evidence_quote":"Defines the cloud-to-cloud distance metric from which the reported precision, recall, and F1 scores are computed."},{"cited_title":"Cooperative multi-UA V coverage mis- sion planning platform for remote sensing applications","cited_arxiv_id":null,"evidence_quote":"Provides the commercial multi-UAV mission planning platform used in the real-world flight test."}],"review_version":1}