{"id":"1167ec00-453d-4cf1-b4ce-2195bcbe3f80","arxiv_id":"2504.16606","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"HUG combines visibility-based block partitioning with hierarchical neural Gaussians and level-weighted supervision to improve rendering quality and speed for large-scale aerial scenes.","lead":"HUG, a new system for rendering city-scale aerial scenes, splits a large scene into blocks and trains each block with a hierarchy of detail levels. It reports faster partitioning, roughly one minute per scene, and higher rendering quality than previous large-scene Gaussian methods on five datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central SOTA claim is not established because Table 1 mixes baseline numbers from different papers with different protocols; even in that table, Sci-Art PSNR is below Mega-NeRF and Residence SSIM is a tie, so the strongest abstract claim overreaches.","rationale":"I read the paper as an engineering contribution on block-based large-scale 3DGS. The main empirical assertion is SOTA quality on five scenes and faster partitioning. The weakest point is not the algorithm's internal logic; it is the evaluation. The supplement's disclosure that most baseline numbers are quoted from other papers is the decisive fact: the headline table is a meta-analysis, not a benchmark. The authors also acknowledge the Sci-Art PSNR exception, making the abstract's unqualified SOTA sentence inaccurate. These issues are fixable by a unified re-benchmark and a softened claim (SOTA among Gaussian methods, or SOTA in most metrics). No internal inconsistency in the method's equations was found at the level needed to reject the approach. Therefore the reader's CONDITIONAL verdict remains appropriate.","tokens_in":12276,"tokens_out":4899,"duration_ms":45833,"concrete_test":"Run a unified re-benchmark: use official code/checkpoints for Mega-NeRF, 3DGS, CityGS, VastGS, Hier-GS, and OctreeGS on all five datasets, with identical train/test splits, downsampling factor (x4 for real-world, 1.6k for MatrixCity), block count, iterations, and GPU type. Train each method with at least three seeds and report mean +/- std for PSNR, SSIM, and LPIPS. If HUG's margins over CityGS/VastGS/Hier-GS are within seed noise or reverse, the SOTA claim fails; if the margins persist under the fixed protocol, the central claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the stated SOTA result, every row in Table 1 must be comparable. The supplement explicitly says that Mega-NeRF, 3DGS, and CityGS numbers are taken from the CityGS paper, OctreeGS numbers from the OctreeGS paper, VastGS from an unofficial code base, and only Hier-GS was run by the authors. This means train/test splits, image resolution, block counts, training iterations, and hardware are not controlled across the columns. The comparison is therefore not a single controlled benchmark. This is load-bearing because the reported wins are small in several cells: Residence SSIM is a tie with CityGS (0.813 vs 0.813), and on Sci-Art PSNR HUG is 21.83 versus Mega-NeRF's 25.60, which contradicts any unqualified SOTA claim in the abstract. A plausible re-run under fixed protocol could change several of these margins. The method may still be better, but the paper as written does not establish the strongest claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents HUG, a block-based reconstruction method for large-scale aerial urban scenes built on 3D Gaussian Splatting. The contributions are: (i) a visibility-based partitioning scheme that uses sparse COLMAP points and visibility masks to assign views to blocks; (ii) a hierarchical octree-based neural Gaussian representation with level-of-detail anchors, a hierarchical weighted image-supervision loss, a dynamic gradient threshold for anchor splitting, a dynamic anchor-level transition rule, and visibility-based anchor pruning; and (iii) a multi-MLP rendering pipeline that fuses independently trained blocks. The method is evaluated on the synthetic MatrixCity Small City scene and four real-world Mega-NeRF scenes (Residence, Rubble, Building, Sci-Art), reporting PSNR, SSIM, and LPIPS. The authors claim state-of-the-art results and report a roughly one-minute partitioning time, compared with over two hours for CityGS on MatrixCity.","tokens_in":12518,"tokens_out":3451,"duration_ms":34675,"significance":"If the claims were established under a controlled protocol, the paper would make a practical contribution: it combines efficient view partitioning with an LOD-structured neural Gaussian representation, and the reported partitioning speed-up is substantial. The ablation on the Rubble scene shows monotonic improvements when adding each component, which is a useful sanity check. However, the central state-of-the-art claim is not currently supported because the quantitative comparison mixes results from different papers and different codebases without a unified training/evaluation protocol. The paper also relies on a single run per method and reports no error bars, which is particularly problematic because several claimed wins are small or even negative (e.g., Residence SSIM is a tie with CityGS, and Sci-Art PSNR is lower than Mega-NeRF). The core algorithmic ideas are plausible, but the paper needs a controlled comparison or a substantially softened claim to be acceptable.","major_comments":[{"comment":"The central claim of state-of-the-art performance is not established because the baseline numbers in Table 1 are not obtained under a single controlled protocol. The supplement states that Mega-NeRF, 3DGS, and CityGS results are taken from the CityGS paper, Octree-GS results from the Octree-GS paper, VastGS results from an unofficial codebase, and only Hier-GS and HUG were run by the authors. If training splits, image resolutions, block counts, iterations, or evaluation code differ across these sources, the columns are not directly comparable. This is load-bearing because several margins are small: on Residence, HUG ties CityGS in SSIM (0.813 vs. 0.813), and on Sci-Art, HUG's PSNR (21.83) is below Mega-NeRF's (25.60), a fact acknowledged in Section 4.3. The abstract's unqualified state-of-the-art claim therefore overreaches. I ask the authors to either rerun all baselines under one protocol with the same splits and resolution, or clearly restrict the claim to the configurations actually compared.","section":"Table 1 and Supplementary Section 6"},{"comment":"The partitioning-efficiency comparison is not apples-to-apples. CityGS's reported 188-minute partitioning time includes building a coarse global 3DGS model and rendering images for all blocks, whereas HUG's one-minute partitioning uses only COLMAP sparse points and does not perform this extra work. The claim that HUG 'significantly outperforms existing methods in speed' for partitioning is therefore misleading unless the comparison isolates the same operation. Please report the time for the specific partitioning operation itself for each method, or state explicitly what each number includes.","section":"Table 2 and Section 4.3"},{"comment":"Equation (9) is unclear and may be internally inconsistent. The text says an anchor's level is 'gradually increased by 0.01,' but L is used elsewhere as an integer octree level in Eq. (7) and in the anchor-selection rule L ≤ L-hat. Adding 0.01 to an integer level and then flooring the sum is not a standard refinement mechanism, and the formula as written sums an indicator over iterations but does not define the summation index or the update schedule. Please clarify whether L is an integer or a continuous accumulator, and define the update precisely.","section":"Section 3.3, Eq. (9)"},{"comment":"The ablation study is reported on a single scene (Rubble) with a single run and no statistical variability. Given that the paper's improvements over baselines are sometimes smaller than 0.01 in SSIM or LPIPS, the absence of multiple seeds or error bars makes it difficult to judge whether the component contributions and the final gains are significant. At minimum, report multiple runs for the ablation and for the main comparison, or justify the computational cost of doing so.","section":"Table 3 and Section 4.3"}],"minor_comments":[{"comment":"The sentence 'An overview of the proposed method is illustrated in Section 2.1' should refer to Figure 2, not to Section 2.1 (Related Work).","section":"Section 3, first paragraph"},{"comment":"The text refers to 'CityGS [29]' when comparing results, but CityGS is reference [17]; [29] is City-on-Web. Please correct the citation.","section":"Section 4.3, Results Analysis"},{"comment":"There is a typo: 'hyperprameter' should be 'hyperparameter'.","section":"Section 3.3, Anchor Splitting"},{"comment":"The hierarchical loss in Eq. (8) would benefit from a one-sentence derivation of the weight choices, especially because the first two terms (λ, γ) and the summed terms (λ/2^{K+1-k}, γ/2^k) are not normalized to sum to a constant across levels. Please clarify whether this weighting is intentionally biased toward the finest level.","section":"Equation (8)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid engineering story and the partitioning speed result is attractive, but the SOTA claim is currently supported only by a mixed-protocol comparison. The editors may wish to ask for a controlled rerun of baselines or a clear downgrade of the claim. The citation mix-up (CityGS [29] vs. [17]) and the unclear Eq. (9) also need attention, but they are secondary."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent engineering paper that combines known ingredients well and has one genuinely useful new idea (visibility-based partitioning) plus several plausible training tricks. The evaluation, however, does not support the unqualified SOTA claim in the abstract, because the baselines come from different papers and protocols.\n\nWhat's new: the one-threshold visibility-based view partition is simple and fast (about a minute vs hours for CityGS), and the hierarchical weighted image supervision with per-level rendering seems like a genuine fix for Octree-GS's floaters. The dynamic gradient threshold and anchor level transition are small but sensible tweaks. The ablation on Rubble is monotonic and clean. The greedy vote-based refiltering is a reasonable boundary artifact fix.\n\nWhat's soft: Table 1 is a mixed-source comparison. The supplement says Mega-NeRF, 3DGS, and CityGS numbers are quoted from the CityGS paper, OctreeGS from its paper, VastGS from an unofficial codebase, and only Hier-GS and your own runs were actually executed. That means splits, resolution, iterations, hardware are not controlled. Several margins are tiny: Residence SSIM is a tie with CityGS (0.813 vs 0.813), and on Sci-Art your PSNR (21.83) is below Mega-NeRF (25.60). So the abstract's \"state-of-the-art\" is an overclaim as written. Also, VastGS is missing from MatrixCity, and the supplement admits avoiding a specific block for visual comparisons - that's a selection-bias flag. No error bars or multiple seeds reported, and one-metric-per-dataset comparisons are fragile.\n\nNone of this is fatal. The method likely does improve on CityGS in most settings, and the partitioning speedup is real. But the paper would be much stronger if the authors reran all baselines under one protocol, or at least clearly labeled provenance per row and softened the SOTA claim. The threshold choices are many, but that is normal for this subfield and the ablation covers the main components.\n\nWho it's for: people working on city-scale 3DGS or large-scale neural rendering. It is a useful source of tricks and a data point for partitioning strategies. It deserves peer review, not desk rejection, but a serious referee should push for a controlled comparison and reproducibility details.\n\nRecommendation: send to peer review with the expectation of major revision on the evaluation.","headline":"Solid engineering with a real partitioning speedup and a clean ablation, but the SOTA claim rests on a mixed-protocol comparison and needs a unified re-benchmark.","tokens_in":13068,"tokens_out":1935,"would_cite":true,"duration_ms":17915,"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":"Visibility-based block splitting cuts large aerial scene partitioning from about two hours to about one minute, the paper claims, while hierarchical per-block Gaussian training improves rendering quality on five datasets.","keywords":["3D Gaussian Splatting","large-scale scene reconstruction","aerial urban scenes","block-based reconstruction","visibility-based partitioning","level of detail","neural Gaussians","novel view synthesis"],"falsifier":"Re-run the comparison by training every baseline on the exact same train/test splits and downsampled images used for HUG, then check whether the PSNR, SSIM, and LPIPS margins in the paper's main table survive; if quoted numbers came from different protocols, the superiority claim would not be established.","tokens_in":12044,"feed_emoji":"🏙️","tokens_out":5782,"duration_ms":55515,"temperature":0.7,"pith_summary":"Large aerial city scenes are too big for 3D Gaussian Splatting to train as a single model, so HUG splits them into blocks and reconstructs each block independently. The paper claims that partitioning by visibility, keeping only the views that actually see each block, takes about one minute per scene, where a leading block-based baseline needs over two hours. Within each block, HUG trains an octree-level hierarchy of neural Gaussians with a weighted supervision scheme that keeps coarse and fine anchors active, reducing zoom artifacts. On one synthetic city and four real-world aerial datasets, the paper reports the best PSNR, SSIM, and LPIPS scores among the compared methods. A sympathetic reader would care because this is a path to memory-feasible, real-time rendering of whole cities from aerial imagery.","feed_headline":"One-minute scene splitting brings 3D Gaussian rendering to city scale","feed_subtitle":"By training hierarchical Gaussian blocks, HUG lifts rendering quality across five large aerial scenes.","key_machinery":"The central object is a hierarchical neural Gaussian representation: anchors placed at nodes of an octree built from each block's sparse point cloud, with the octree depth set by the spread of camera distances, and each anchor decoding into several neural Gaussians via a small MLP. Its level of detail is carried by distance-based anchor selection, so coarse levels supply base colors and the finest level supplies detail. The argument is carried by two additional mechanisms: a visibility-based view partition that assigns only informative images to each block and masks irrelevant pixels, and a hierarchical weighted image supervision that renders each level separately and applies level-dependent L1 and SSIM weights, keeping every anchor trained and suppressing the floaters that appear in a plain octree-Gaussian baseline.","core_discovery":"On its own terms, the discovery is that block-based reconstruction of large-scale aerial scenes does not need redundant overlap or a slow global pre-pass. HUG divides COLMAP sparse points into uniform blocks, assigns to each block only the training images whose reprojected visible point count exceeds a threshold, and derives a visibility mask that focuses loss on in-block regions. Each block is trained as an octree-organized set of anchors that produce neural Gaussians through a small MLP; level selection by camera distance gives a level-of-detail structure. A hierarchical weighted image loss supervises every level separately, a dynamic gradient threshold and anchor level transition refine detail where gradients demand it, and visibility-based pruning removes under-used anchors. At fusion time, a greedy vote keeps anchors whose generated Gaussians mostly fall inside the block, and a multi-MLP rasterizer renders all blocks seamlessly. The paper reports top PSNR, SSIM, and LPIPS on MatrixCity Small City and on the Residence, Rubble, Building, and Sci-Art scenes from Mega-NeRF.","pith_inferences":["The fixed visible-point threshold for assigning views is a natural target for adaptivity: a per-block threshold based on point density or coverage could make the partition robust to low-texture areas, which the paper itself flags as a limitation of relying on COLMAP points.","The level-weighted supervision recipe, supervising every octree level separately with geometrically decaying weights, looks transferable to other LOD radiance-field or Gaussian architectures as a generic anti-float training strategy.","If pose-free structure-from-motion replaces COLMAP, the same visibility partition could apply to unstructured drone video, since nothing in the partition itself requires calibrated poses beyond what the sparse reconstruction provides.","Separate MLPs per block open a route to per-block appearance or lighting models, which could compensate for the absence of dynamic-object and lighting-decoupling handling that the paper lists as future work."],"forward_implications":["The one-minute visibility-based partition makes block decomposition cheap enough to rerun as scenes change, lowering the total cost of city-scale reconstruction from hours of preprocessing plus training to roughly training time alone.","Because each block is trained without large redundant boundary regions and with fewer views, memory per block drops, which is what makes parallel training on multiple GPUs practical for a whole city.","The hierarchy gives free level of detail: nearby views get fine neural Gaussians, distant views use coarse ones, so real-time fly-throughs do not pay full detail cost for the whole scene.","Reporting best SSIM and LPIPS on four real aerial datasets implies the method removes the multicolored zoom-in artifacts that appear in octree-based Gaussian rendering.","The multi-MLP rasterizer means the benefits of neural Gaussians are no longer limited to single-block scenes; trained blocks from different machines can be merged and rendered seamlessly."],"supporting_citations":[{"why":"Supplies the base 3D Gaussian Splatting representation, differentiable rendering, and the optimization losses that HUG extends.","marker":"[11]"},{"why":"CityGS is the main large-scale block-based Gaussian baseline and the source of several quoted comparison numbers.","marker":"[17]"},{"why":"Octree-GS provides the octree LOD anchor structure and the zoom-in artifact problem that HUG's hierarchical supervision targets.","marker":"[26]"},{"why":"Scaffold-GS contributes the anchor-to-neural-Gaussian decoding scheme that HUG's hierarchical representation builds on.","marker":"[18]"},{"why":"MatrixCity supplies the synthetic city-scale dataset used for the main quantitative evaluation.","marker":"[14]"},{"why":"Mega-NeRF provides the four real-world aerial datasets and the Mega-NeRF baseline for comparison.","marker":"[33]"},{"why":"COLMAP supplies the sparse point cloud that the visibility-based partitioning and octree initialization depend on.","marker":"[27]"},{"why":"Hier-GS is the hierarchical Gaussian baseline for very large datasets that HUG compares against and says its design improves on.","marker":"[12]"}],"fun_headline_variants":["HUG: fast block-based Gaussian splatting for city scenes","Visibility-based partitioning accelerates 3D Gaussian training","Hierarchical training boosts quality of urban Gaussian rendering","HUG scales 3D Gaussian splatting to aerial city-scale datasets","Block-level neural Gaussians render large scenes in minutes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art comparison assumes that the numbers reported for several prior methods, which are quoted from those methods' own papers rather than rerun by the authors, were produced with the same training splits, image resolution, and evaluation protocol as HUG's runs.","fun_headline_variants_meta":{"raw":{"variants":["HUG: fast block-based Gaussian splatting for city scenes","Visibility-based partitioning accelerates 3D Gaussian training","Hierarchical training boosts quality of urban Gaussian rendering","HUG scales 3D Gaussian splatting to aerial city-scale datasets","Block-level neural Gaussians render large scenes in minutes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1612,"prompt_tokens":906,"completion_tokens":706,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":624}},"tokens_in":522,"tokens_out":706,"duration_ms":7164,"temperature":1.0,"reasoning_tokens":624,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:59:09.850022+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the comparison by training every baseline on the exact same train/test splits and downsampled images used for HUG, then check whether the PSNR, SSIM, and LPIPS margins in the paper's main table survive; if quoted numbers came from different protocols, the superiority claim would not be established.","supporting_citations":[{"cited_title":"3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4), 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the base 3D Gaussian Splatting representation, differentiable rendering, and the optimization losses that HUG extends."},{"cited_title":"Citygaus- sian: Real-time high-quality large-scale scene rendering with gaussians","cited_arxiv_id":null,"evidence_quote":"CityGS is the main large-scale block-based Gaussian baseline and the source of several quoted comparison numbers."},{"cited_title":"Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians, 2024","cited_arxiv_id":null,"evidence_quote":"Octree-GS provides the octree LOD anchor structure and the zoom-in artifact problem that HUG's hierarchical supervision targets."},{"cited_title":"Mega-nerf: Scalable construction of large- scale nerfs for virtual fly-throughs","cited_arxiv_id":null,"evidence_quote":"Mega-NeRF provides the four real-world aerial datasets and the Mega-NeRF baseline for comparison."},{"cited_title":"Structure-from-motion revisited","cited_arxiv_id":null,"evidence_quote":"COLMAP supplies the sparse point cloud that the visibility-based partitioning and octree initialization depend on."},{"cited_title":"A hierarchical 3d gaussian representation for real-time ren- dering of very large datasets.ACM Transactions on Graphics (TOG), 43(4):1–15, 2024","cited_arxiv_id":null,"evidence_quote":"Hier-GS is the hierarchical Gaussian baseline for very large datasets that HUG compares against and says its design improves on."}],"review_version":1}