{"id":"c00bd25b-cef5-4fd7-a0f1-7364c1763df3","arxiv_id":"2505.10578","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ExploreGS combines drone exploration, smart image-pair selection, a MASt3R neural network for point clouds, and 3D Gaussian Splatting to reconstruct scenes from RGB images on an edge device.","lead":"A new drone framework, ExploreGS, uses ordinary color cameras to build a detailed 3D model of a scene while flying, without expensive laser scanners. In tests it rebuilt a warehouse and an office in about a minute, with visual quality close to slower state-of-the-art methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quality comparison may be uninformative: reported PSNR appears insensitive to missing scene coverage, and no completeness metric is provided.","rationale":"The reader's weakest assumption focused on the BoW selector retaining enough overlapping image pairs for every part of the scene; my concern is adjacent but more fundamental: the evaluation protocol does not measure whether the final model is complete. The paper's own simulation result suggests PSNR is not sensitive to missing scene content, so the headline numerical comparison may be an artifact of how PSNR was computed. I still view the method as plausible, and the right remedy is conditional acceptance: release code and data, add repeated trials with error bars, report a scene-completeness metric and coverage-aware PSNR, and run baselines on the same hardware. That is why the verdict remains conditional rather than changing to accept or reject. I do not allege any selective reporting; the ambiguity in the text is enough to make the claim under-supported.","tokens_in":8043,"tokens_out":7284,"duration_ms":72966,"concrete_test":"Re-run the simulation and real-world evaluations with a coverage-aware protocol: render a depth map from each trained 3DGS model for every held-out view, mark pixels whose rendered depth is within, say, 5% of the reference depth (from the Gazebo ground truth or the D455/occupancy grid), and report (i) full-image PSNR, (ii) covered-pixel PSNR, and (iii) the fraction of pixels covered. Also sweep tau in {0.01, 0.03, 0.1} and thrin in {0.02, 0.04, 0.1} and plot coverage versus pair count. If InstantSplat's full-image PSNR drops below Ours once unmodeled pixels are included, or if coverage falls sharply at the paper's thresholds, the claimed quality/speed tradeoff is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is a full 3DGS scene model, comparable in quality to state-of-the-art methods, at roughly five times lower cost. For that claim to hold, the quality metric must actually measure quality of the complete scene. The paper reports only global PSNR (Tables II and III), and in Section IV-B it states that InstantSplat 'reconstructs only a quarter of the scene' while scoring 24.12 dB, higher than the proposed method's 23.91 dB. If the evaluation views included the unmodeled region, a model missing 75% of the scene could not achieve that PSNR; spherical-harmonic adaptability can change the colors of rendered Gaussians, but it cannot invent correct pixels for geometry that was never reconstructed. The natural conclusion is that the reported PSNR is computed on pixels or views that are covered, or that the comparison set does not exercise the missing region. Either way, the table does not support the claim of comparable quality on the full scene. The BoW admission filter in Section III-B, Eq. (5), is a plausible source of coverage loss: a texture-poor or low-overlap segment can fall below the similarity threshold and never enter the pair set, and the paper gives no coverage or completeness metric, no failure cases, and no sensitivity analysis for tau and thrin. This is precisely where the 'complete model' claim is most insecure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents ExploreGS, an RGB-only 3D scene reconstruction framework for drones. It replaces LiDAR-based point cloud acquisition with MASt3R, a vision model that generates dense point clouds from selected image pairs, and uses a Bag-of-Words (BoW) based match selector to reduce the number of image pairs by 84-95%. The resulting point cloud initializes a 3D Gaussian Splatting model that is trained in about one minute. Experiments in a Gazebo simulation and one real-world room compare PSNR and runtime against InstantSplat, Splatfacto, Mip-NeRF, Bio-NeRF, and NeRF, reporting a factor-of-five speedup with PSNR within about 0.2-0.8 dB of InstantSplat.","tokens_in":8247,"tokens_out":5446,"duration_ms":49051,"significance":"The proposed pipeline is a practical integration of recent components (MASt3R, 3DGS, and VINS-style BoW) into a drone system with on-board edge inference, which is timely and potentially useful. A strength is the real-world validation on a physical quadrotor. The 84-95% data reduction and the reported speedup on an edge processor are potentially valuable. However, the quantitative support for the headline claim of comparable quality is not yet convincing: PSNR is reported without completeness or coverage metrics, and the paper itself notes that a higher-PSNR competitor reconstructs only part of the scene. No code or dataset release is mentioned, and all quantitative results are single runs. I found no circularity: the compared systems are external and the thresholds are hand-set, so the circularity burden is effectively zero.","major_comments":[{"comment":"Section IV-B states that InstantSplat 'reconstructs only a quarter of the scene' yet obtains a higher PSNR (24.12 dB) than ExploreGS (23.91 dB), and attributes this to spherical-harmonic adaptability. This is a direct admission that the reported PSNR is not a reliable indicator of complete-scene reconstruction quality. Since the central claim is comparable quality of the complete 3D model, the evaluation must include completeness and coverage metrics (e.g., the fraction of evaluation views or pixels backed by valid geometry, or accuracy against a ground-truth scan) and report PSNR on the full evaluation set as well as on covered subsets. Without this, Tables II and III do not support the claim of comparable quality.","section":"IV-B, Table II"},{"comment":"All PSNR and timing entries are single runs on a self-built Gazebo warehouse and one real-world room. No error bars, repeated trials, or standard datasets are reported, so the 0.2-0.8 dB difference to InstantSplat is within plausible run-to-run variation. The authors should report mean and standard deviation over multiple runs and include a standard multi-view or sparse-view dataset (e.g., DTU, Tanks and Temples, or the benchmark used by InstantSplat) to allow external comparison.","section":"IV-B, Tables II-III"},{"comment":"The BoW admission filter with thresholds tau=0.03 and thrin=0.04 determines which image pairs are passed to MASt3R, and the paper reports an 84-95% reduction in image pairs. However, no sensitivity analysis is provided and no coverage or completeness metric is given to show that discarded images do not correspond to under-reconstructed regions. Texture-poor or low-overlap segments could be filtered out, causing holes in the final model. The authors should report the number of selected keyframes and pairs per scene and analyze the effect of varying tau and thrin on both runtime and reconstruction completeness.","section":"III-B, Eqs. (3)-(5)"},{"comment":"The timing comparison is not apples-to-apples. The simulation for the proposed method runs on an RTX 3060, while all comparative algorithms run on a server with 4 RTX 4090 GPUs. While this may understate the speed advantage of the proposed method, the authors should specify exactly which stages are included in 'Time' (data acquisition, pair selection, point cloud generation, global alignment, 3DGS training) and report timings on the same hardware, or at least with identical GPU models; otherwise the '5 times faster' claim cannot be independently verified.","section":"IV-A, Tables II-III"}],"minor_comments":[{"comment":"The word 'leverags' should be 'leverages'.","section":"Abstract"},{"comment":"There is an extra space before the comma in 'occupancy grid ,'; please fix the formatting.","section":"III-A"},{"comment":"The confidence map q was defined as per-pixel earlier in the section, but Eq. (10) uses q as a scalar weight in the sum over pairs; clarify whether q is a pixel-wise weight, an aggregated score, or something else.","section":"III-C, Eq. (10)"},{"comment":"'Comparisions' should be 'Comparisons'.","section":"IV-B"},{"comment":"The phrase 'less than five-sixth of the computational time' is confusing: 50 seconds is about one-seventh of 6 minutes, so 'less than one-sixth' or a precise ratio would be clearer.","section":"IV-C"},{"comment":"The notation for the threshold is inconsistent: 'thrin' appears in Eqs. (3)-(5) and in the experimental setup, but the surrounding text also uses 'thr_in'; please define it once and use it consistently.","section":"Section III-B"},{"comment":"The captions state which reconstruction is 'Ours' and which is 'InstantSplat', but they do not indicate which image shows the complete model and which shows the partial one; please add this information to the captions.","section":"Fig. 6 and Fig. 9"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's own admission that InstantSplat's higher PSNR corresponds to an incomplete model is the central problem; a revision that adds completeness and coverage metrics, repeated runs, a standard dataset, and a sensitivity analysis for the BoW thresholds would address the main concerns. With those additions, the paper could be a suitable fit for this venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Ricardo —\n\nYou asked about ExploreGS (2505.10578). Quick take: it's a credible system integration — frontier exploration, BoW pair selection, MASt3R, and 3DGS running on an edge drone — and the pair-selection rule (Eqs. 3–5) is the only genuinely new piece. The paper is honest about not introducing new theory. What it does well: the pipeline is coherent, the data reduction numbers (84–95%) are the kind of claim that matters for onboard processing, and running on an Orin NX is a concrete demonstration that the low-overhead story is real.\n\nThe soft spots are mostly in the validation, and one is serious. In the simulation comparison, InstantSplat gets a higher PSNR (24.12 vs 23.91) while reconstructing only a quarter of the scene. The authors wave this away with spherical-harmonic adaptability, but SH cannot invent correct pixels for geometry that was never built. The natural reading is that PSNR was computed on covered views or on a set that does not exercise the missing region. Either way, Table II does not support the \"comparable quality\" claim for the full scene. That is a real hole. If they want to claim completeness, they need a completeness metric, not just global PSNR.\n\nAlso minor-to-moderate: every number is a single run with no error bars; baselines run on a much bigger machine (4x 4090 vs a 3060 or Orin), so the time comparison is confounded; no code or data release; no sensitivity analysis of tau and thrin, which control coverage. The BoW admission filter could plausibly drop texture-poor segments, and no failure cases are shown.\n\nThe central idea holds up. The circularity burden is essentially zero — MASt3R and 3DGS are external, thresholds are chosen, not fitted. The paper just needs a better experimental protocol.\n\nWho this is for: people building autonomous drone reconstruction systems, especially with edge constraints. It deserves a serious referee because the integration is non-trivial and the claims are testable, but I would expect heavy revision before acceptance: release code and data, run repeated trials, fix the coverage/PSNR inconsistency, and report completeness.\n\nMy recommendation: send it to peer review, with a clear message that the validation section as written does not support the headline claim.","headline":"A credible edge-drone reconstruction pipeline whose central quality claim is undercut by a flawed PSNR comparison that ignores scene completeness.","tokens_in":8833,"tokens_out":2143,"would_cite":false,"duration_ms":21643,"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":"ExploreGS reconstructs complete 3D Gaussian Splatting scenes from RGB drone video in about one minute, roughly five times faster than prior methods, while in the reported tests PSNR stays within about a decibel of the fastest baseline.","keywords":["3D Gaussian Splatting","drone reconstruction","vision-based reconstruction","Bag-of-Words image selection","MASt3R","on-board 3D reconstruction","RGB-only reconstruction"],"falsifier":"Run the drone through a scene containing a long textureless corridor or a large low-overlap loop and measure the completeness of the final 3DGS model against a lidar ground truth; if the selected image pairs leave holes where MASt3R had no overlap, the completeness gap will show up as missing geometry and a sharp PSNR drop in those views.","tokens_in":7799,"feed_emoji":"🎥","tokens_out":4435,"duration_ms":40793,"temperature":0.7,"pith_summary":"This paper proposes ExploreGS, a drone-deployed pipeline that turns ordinary RGB video into a photorealistic 3D Gaussian Splatting model without lidar and without offline structure-from-motion. The central claim is that replacing lidar point clouds with a neural vision model, and pairing images through Bag-of-Words selection, cuts reconstruction time to about one minute on an edge processor while keeping reconstruction quality comparable. The authors validate this in a simulated warehouse and a real indoor flight, reporting PSNR within about 0.8 dB of the fastest baseline and a reduction of 84 to 95 percent in redundant image pairs. The result, if it holds, makes same-flight, on-board 3D reconstruction practical for small drones.","feed_headline":"A drone-only camera pipeline rebuilds 3D scenes five times faster","feed_subtitle":"No lidar: selective image pairing plus a vision model cuts compute so the drone finishes in about a minute.","key_machinery":"The load-bearing component is the Match Selector, a Bag-of-Words similarity filter with thresholds tau=0.03 and thrin=0.04 that dynamically normalises scores against a recent historical frame and admits only image pairs with overlapping views. This step cuts the visual inference workload from thousands of image pairs to fewer than 180 for a 60-frame sequence, which is what makes on-board 3DGS training feasible. The vision model MASt3R then converts each admitted pair into a dense local point cloud, and a global alignment step jointly optimises camera extrinsics and scale factors before Gaussian Splatting training begins.","core_discovery":"The central discovery is that the expensive parts of 3D reconstruction, lidar scanning and exhaustive image matching, can be replaced by a cheap selective pairing step plus a learned stereo model. ExploreGS lets the drone autonomously explore an unknown environment while capturing RGB frames, uses FAST and BRIEF features with a Bag-of-Words dictionary to admit only a small set of overlapping image pairs, feeds those pairs to MASt3R to produce local dense point clouds, globally aligns them, and initializes 3DGS training on the merged cloud. The reported result is complete scene reconstruction in roughly one minute on a Jetson Orin NX edge processor. While InstantSplat sometimes reaches a slightly higher PSNR, the paper reports that it can cover only part of the scene when the initial point cloud is wrong, whereas ExploreGS produces the full model.","pith_inferences":["A testable extension is to run the same pipeline on scenes with heavily repetitive texture or large textureless walls, where BoW similarity scores may not reflect true 3D overlap; the pair selector should be compared against ground-truth covisibility.","The reported speed-up depends on MASt3R's forward passes dominating the pipeline; on drones with weaker GPUs the real-time pairing step could become the bottleneck, so profiling the Match Selector separately would locate where the fivefold gain lives.","The same selective-pairing idea could apply to other dense stereo models or to online 3DGS training during flight, which the paper lists as future work."],"forward_implications":["Drone 3D reconstruction no longer requires lidar payloads or a second surveying flight, so the same flight that explores also produces the final model.","The 84 to 95 percent reduction in image pairs lowers on-board compute enough that 3DGS training runs on an edge processor in about a minute.","The pipeline sidesteps classic SfM failure in low-overlap or texture-poor sequences, because pair selection and dense geometry come from a learned model rather than epipolar feature matching.","If the approach transfers to other scenes, inspection and mapping tasks could get same-flight photorealistic models instead of post-processed lidar fusion."],"supporting_citations":[{"why":"Supplies MASt3R, the neural vision model that turns overlapping image pairs into dense local point clouds and camera poses, replacing lidar.","marker":"[15]"},{"why":"Provides 3D Gaussian Splatting, the scene representation and differentiable renderer that the merged point cloud initializes.","marker":"[16]"},{"why":"Supplies the pre-trained visual vocabulary and dynamic normalization scheme used by the Match Selector.","marker":"[19]"},{"why":"Provides the Bag-of-Words binary image retrieval method for fast feature-similarity assessment.","marker":"[20]"},{"why":"Supplies FAST corner detection, the keypoint extraction step feeding the pair selector.","marker":"[21]"},{"why":"Supplies BRIEF binary descriptors that encode keypoint neighborhoods for BoW matching.","marker":"[22]"},{"why":"Supplies the ATSP formulation used to sort exploration viewpoints and reduce flight time.","marker":"[17]"},{"why":"Provides InstantSplat, the fastest baseline and the main comparison for reconstruction quality and runtime.","marker":"[23]"}],"fun_headline_variants":["Drone camera and bag-of-words rebuild full 3D scenes in one minute","Vision-based drone reconstruction skips lidar, runs on board in real time","Selective image pairs plus learned stereo give complete drone 3D models","ExploreGS: low-overhead 3D scene reconstruction for drones from RGB only"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The Match Selector's fixed similarity thresholds (tau=0.03 and thrin=0.04) are assumed to keep enough overlapping views of every region for MASt3R to reconstruct and align the full scene, but the paper reports no coverage metric, no failure cases, and no sensitivity analysis for these thresholds.","fun_headline_variants_meta":{"raw":{"variants":["Drone camera and bag-of-words rebuild full 3D scenes in one minute","Vision-based drone reconstruction skips lidar, runs on board in real time","Selective image pairs plus learned stereo give complete drone 3D models","ExploreGS: low-overhead 3D scene reconstruction for drones from RGB only"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000578,"raw_usage":{"total_tokens":2667,"prompt_tokens":829,"completion_tokens":1838,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":1753}},"tokens_in":445,"tokens_out":1838,"duration_ms":16721,"temperature":1.0,"reasoning_tokens":1753,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:39:47.107440+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the drone through a scene containing a long textureless corridor or a large low-overlap loop and measure the completeness of the final 3DGS model against a lidar ground truth; if the selected image pairs leave holes where MASt3R had no overlap, the completeness gap will show up as missing geometry and a sharp PSNR drop in those views.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering,","cited_arxiv_id":null,"evidence_quote":"Provides 3D Gaussian Splatting, the scene representation and differentiable renderer that the merged point cloud initializes."},{"cited_title":"Bags of binary words for fast place recognition in image sequences,","cited_arxiv_id":null,"evidence_quote":"Provides the Bag-of-Words binary image retrieval method for fast feature-similarity assessment."},{"cited_title":"Machine learning for high-speed corner detection,","cited_arxiv_id":null,"evidence_quote":"Supplies FAST corner detection, the keypoint extraction step feeding the pair selector."},{"cited_title":"Brief: Binary robust independent elementary features,","cited_arxiv_id":null,"evidence_quote":"Supplies BRIEF binary descriptors that encode keypoint neighborhoods for BoW matching."}],"review_version":1}