{"id":"ed645d25-8a81-42bb-be9d-39fb6a8ecc64","arxiv_id":"2501.11097","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"URE divides floorplans into density-based unit regions and learns per-region encodings, claimed to beat pixel-wise and graph baselines on three tasks, but evidence is weakened by internal inconsistencies and an undiscussed baseline.","lead":"The paper proposes Unit Region Encoding (URE), a floorplan representation that divides a floorplan into boundary-aware unit regions using a hand-designed density map, then learns a feature vector for each region with a neural network. The authors claim this representation improves three floorplan tasks: space planning, metric learning, and generation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical evidence for URE's superiority is internally inconsistent: Table 3's 'pixelwise+voting' baseline nearly matches or beats the proposed method, and Table 2/3 report different numbers for identical settings.","rationale":"The reader's weakest_assumption identifies the axis-aligned 'squeeze' model behind the density map as the most fragile geometric premise. That is a legitimate generalization concern: Section 3.1 assumes floorplans are derived from an axis-aligned bounding box by squeeze operations, and Section 3.2 handles sloped or curved walls with an ad hoc rule, but the partition algorithm is not fully specified. However, the most load-bearing weakness in the paper as written is more immediate: the experimental validation contradicts itself. Table 2 and Table 3 are duplicates in title and structure yet report different numbers for the identical URE-Net+DeepLabv3+ configuration, meaning the quantitative basis for the central claim is not trustworthy. Furthermore, the 'pixelwise+voting' baseline in Table 3 is the correct control for isolating the contribution of the learned region encoding from the contribution of the region partition alone. It shows that URE's mIoU advantage over pixelwise voting is negligible (72.73 vs 72.62) and that URE is worse on Boundary F All (72.82 vs 73.56). The paper's assertion that URE yields 'more highly regular' segmentations is therefore not supported by its own table. These empirical problems directly attack the strongest claim of superiority, while the geometric assumption is a secondary risk about generalization. If the authors can correct the tables and demonstrate, with error bars, that URE clearly beats pixelwise+voting on internal boundary regularity and mIoU, the verdict could move toward conditional acceptance. As printed, the central claim is unsupported, so the reader's REJECT verdict stands unchanged.","tokens_in":14925,"tokens_out":6182,"duration_ms":61242,"concrete_test":"Re-run the interior space planning experiment under one fixed protocol: identical training/validation split, identical (8×8, 1m) unit region partition, identical optimizer settings, and at least three random seeds. Report mIoU and Boundary F (All and Internal) with error bars for DeepLabv3+, URE-Net+DeepLabv3+, and 'pixelwise+voting'. Also reconcile the URE-Net+DeepLabv3+ numbers that currently appear as 72.32 in Table 2 and Table 4 but 72.73 in Table 3. If 'pixelwise+voting' remains statistically tied with or better than URE on the Internal Boundary F score, the claim that learned region encodings improve segmentation regularity is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that URE achieves 'higher accuracy performance and better visual quality' across floorplan applications. The validation for this claim rests on the interior space planning experiments, and those experiments are internally inconsistent. Table 2 and Table 3 have identical captions but report different numbers for the same configuration: URE-Net+DeepLabv3+ under (8×8, 1m) has mIoU 72.32 in Table 2 and 72.73 in Table 3; Table 4's density-map ablation also uses 72.32. More critically, Table 3 introduces an unreported 'pixelwise+voting' baseline: pixel-wise DeepLabv3+ predictions are majority-voted within URE regions. This baseline reaches mIoU 72.62, nearly matching URE's 72.73, and it beats URE on Boundary F All (73.56 vs 72.82). The paper claims that URE 'constructs a more highly regular functional area segmentation' but on the All-boundary metric the simple voting baseline is better, and on mIoU the learned region encoding adds only 0.11 over voting. This directly undermines the claimed advantage of the learned region encodings and indicates that most of the reported improvement may come from the region partition itself, not from the URE-Net encoding. Since these tables are the only quantitative support for the abstract's superiority claim, the central argument currently rests on unreconciled, conflicting evidence.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Unit Region Encoding (URE), a geometry-aware representation for floorplans. The representation is built from a density map defined by inverse sums of directional distances to the floorplan boundary (Eq. 1), a clustering-based decomposition into density regions, and a grid-splitting step that yields compact unit regions. A URE-Net is trained to produce latent encodings for these regions, and the representation is then applied to three tasks: interior space planning (semantic segmentation of functional areas on 3D-FRONT), floorplan metric learning (triplet embeddings on RPLAN), and floorplan generation (a two-stage pipeline with a user study on RPLAN). The authors claim that URE combines the strengths of rasterized image representations and room-level graph representations, achieving higher accuracy and better visual quality than existing approaches.","tokens_in":15167,"tokens_out":6256,"duration_ms":60395,"significance":"If the claims are supported, URE would be a practically valuable unified representation: it is more compact than dense pixel maps, more geometry-aware than hand-crafted graph nodes, and adaptable to segmentation, retrieval, and generation. The paper has clear strengths: it proposes a well-motivated representation, presents quantitative results on three distinct applications, and includes ablation studies for both the splitting strategy and the density-map input. The compactness numbers in Table 1 (hundreds of times fewer units than a 256x256 pixel grid with high room-label IoU) are appealing. However, the central quantitative support is currently compromised by inconsistent tables and by an unanalyzed strong baseline, so the significance of the claimed superiority cannot be assessed until these issues are resolved.","major_comments":[{"comment":"Tables 2 and 3 have identical captions and purport to report the same experiment, but URE-Net+DeepLabv3+ under (8×8, 1m) has mIoU 72.32 in Table 2 and 72.73 in Table 3, with corresponding differences in per-class IoU and Boundary F scores. Table 4's density-map ablation repeats the Table 2 numbers (72.32), while Table 3's density-map ablation repeats the Table 3 numbers (72.73). The manuscript gives no explanation for this discrepancy; because these tables are the main evidence for the superiority claim, the reported numbers must be reconciled before the results can be trusted.","section":"§4.1, Tables 2–4"},{"comment":"The row 'pixelwise+voting' appears only in Table 3 and is never described in the text, yet it is a critical control: it applies majority voting within URE regions to pixel-wise DeepLabv3+ predictions. It reaches mIoU 72.62, nearly matching URE-Net+DeepLabv3+'s 72.73, and it exceeds URE-Net on All-boundary F-score (73.56 vs. 72.82). This indicates that most of the reported improvement over pixel-wise DeepLabv3+ comes from the region partition itself, not from the learned URE-Net encodings. The paper's central claim that the learned encoding provides 'higher accuracy performance' is therefore not supported unless the authors analyze this baseline and demonstrate what the learned encodings add beyond the partition.","section":"§4.1, Table 3"},{"comment":"The representation is constructed on the assumption that irregular floorplans are obtained from an axis-aligned bounding box by a sequence of 'squeeze' operations. For non-axis-aligned walls, Section 3.2 adds an ad hoc rule ('we take the regions connected by a piece of continuous wall into one density region') that is not formalized or evaluated. If a substantial fraction of real floorplans contain sloped or curved walls, the density map may not align with rooms and downstream performance would degrade. The authors should quantify the prevalence of such walls in 3D-FRONT and RPLAN and report performance separately for these cases, or explicitly restrict the scope of the claimed generality.","section":"§3.1–3.2"}],"minor_comments":[{"comment":"Table 3 is never cited in the text; the narrative refers only to Table 2 and Table 4. Either cite Table 3 where the pixelwise+voting baseline and the density-map ablation are discussed, or merge the tables to avoid duplicate captions with conflicting numbers.","section":"§4.1"},{"comment":"The description of the normalized density map input ('compute the inverse of the density value, filter with a Sobel operator, and divide it by 255') is unclear: the inverse of the density from Eq. (1) is the sum of directional distances, and the role of the Sobel filter in the network input is not explained.","section":"§3.3"},{"comment":"The user study would benefit from statistical analysis: with 600 votes per group, ground truth is preferred over the proposed method 326 to 237, and Graph2Plan is preferred over the proposed method 273 to 232. The wording 'close to the ground-truth' and 'comparable to Graph2Plan' is not fully supported without confidence intervals or a significance test.","section":"§4.3"},{"comment":"The Boundary F-score is described only as using 'a pre-defined distance threshold' within 1 pixel; the exact tolerance and the computation of precision/recall for internal boundaries should be specified.","section":"§4.1"},{"comment":"There are numerous typos and misspellings, including 'ususlly' (Section 4.1), 'Compairons' (Section 4.2), 'Grpah2Plan' (Section 4.3), and 'flooplans' (Section 4.2); the manuscript should be carefully proofread.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The duplicate Tables 2 and 3 with different numbers for the same configuration, together with the unreferenced pixelwise+voting baseline, make the paper look like an unedited draft. These issues are serious but fixable: the authors can rerun or re-report the experiments consistently and add an analysis of what the learned encodings contribute beyond the partition. I do not see a fundamental flaw that would force rejection, but the current evidence does not justify the paper's superiority claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The URE idea is genuinely new for floorplans: a four-directional density map, clustering into unit regions, and learned per-region encodings. The density map is a clean deterministic function of the boundary, not a fitted parameter, and the resulting partition is compact—hundreds of regions instead of tens of thousands of pixels—while preserving room structure (Table 1). The planning experiments show large boundary-F gains over pixel-wise segmentation, and the metric learning results on RPLAN beat LayoutGMN. The paper also contributes a 5,340-floorplan functional-area annotation on 3D-FRONT, which is useful.\n\nThe soft spots are in the validation of the central claim. Table 2 and Table 3 report the same (8×8, 1m) configuration with different mIoU: 72.32 vs 72.73. Table 4 repeats 72.32. That inconsistency is never explained. More damaging, Table 3 introduces a pixelwise+voting baseline—majority-vote pixel predictions within URE regions—that scores 72.62 mIoU and 73.56 All-boundary F, beating URE-Net's 72.82 on boundary F. The paper doesn't discuss it. This suggests the learned region encoding adds little beyond the partition itself; the improvement over pixel-wise segmentation comes mostly from the region structure. That undercuts the claim of learned latent encodings as the source of the gain.\n\nThe generation user study is also reported generously: Graph2Plan gets 273 votes, URE gets 232, and the paper calls this 'comparable.' That's not superiority. And in metric learning, the unit regions are the ground-truth rooms, not the density-based partition, so that application doesn't validate the paper's core representation. The squeeze model of floorplan geometry is an idealization; the ad hoc handling of sloped or curved walls is a minor concern, not the main issue. No code, data, or error bars are provided, and the partition algorithm is underspecified.\n\nThese are load-bearing problems, but the direction is worth pursuing. The density map is not circular in any damaging way, and the representation is a reasonable alternative to raster images and room graphs. The paper deserves a serious referee, not a desk reject. A good reviewer should push the authors to reconcile the tables, put the voting baseline in the main comparison, run multiple seeds, and be honest about the user study. If the numbers hold up after that, the representation has value; as written, the central claim is not supported.","headline":"A promising floorplan representation whose central empirical claims are undercut by inconsistent tables and a strong baseline the paper doesn't address.","tokens_in":15785,"tokens_out":3668,"would_cite":false,"duration_ms":34954,"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":"The paper proposes Unit Region Encoding, a compact geometry-aware floorplan representation that outperforms rasterized images and room-level graphs across interior space planning, metric learning, and generation.","keywords":["floorplan representation","unit region encoding","geometry-aware density map","interior space planning","floorplan metric learning","floorplan generation","region pooling","URE-Net"],"falsifier":"Build a benchmark of floorplans with a high fraction of non-axis-aligned walls, compute the density-based unit regions, and measure region-to-room IoU: if the IoU falls far below the axis-aligned case, or the accuracy gains over pixel and graph baselines vanish, the central claim that the density map makes the encoding geometry-aware is false.","tokens_in":14661,"feed_emoji":"🏠","tokens_out":6567,"duration_ms":61562,"temperature":0.7,"pith_summary":"The paper proposes Unit Region Encoding (URE), a floorplan representation that sits between over-segmented raster images and room-level graphs. The idea is to divide a floorplan into a compact set of unit regions whose boundaries follow the plan's geometry, then learn a latent feature vector for each region with a network. The geometry is captured by a density map: at each point, the reciprocal of the sum of distances to the boundary along the four cardinal directions. The paper claims this representation is flexible enough to serve interior space planning, floorplan metric learning, and floorplan generation, and that it beats pixel-wise and graph-based baselines in accuracy while producing cleaner, more regular boundaries. If correct, URE offers a single encoding that makes floorplan learning tasks simpler and less dependent on post-processing.","feed_headline":"Unit regions beat pixels and graphs on floorplan tasks","feed_subtitle":"Boundary-aware regions give cleaner planning, better matching, and simpler floorplan generation.","key_machinery":"The load-bearing object is the geometry-aware density map, $f(x)=1/\\sum_{v\\in V} d(x,\\partial\\Omega,v)$, where the sum runs over the north, south, east, and west directions and each term is the distance from point $x$ to the nearest boundary point in that direction. The map operationalizes a squeeze model of floorplan formation: inward pushes of an axis-aligned bounding box create narrow high-density corridors and open low-density areas that align with rooms. Clustering this map yields density regions; a splitting strategy $(M\\times N, h)$ slices each region into unit regions while enforcing a minimum grid size, and these unit regions carry latent encodings from URE-Net via region pooling followed by a shared-weight MLP. For global tasks, the region features are combined with a symmetric max-pooling function in the spirit of PointNet.","core_discovery":"The central claim is that floorplans are better encoded on a boundary-adaptive unit region partition than on either dense pixels or hand-crafted room graphs. The partition is obtained by clustering a geometry-aware density map, defined as the inverse of the sum of nearest distances to the boundary in the four cardinal directions, and then slicing each density region into grids under a configurable splitting strategy. A trained URE-Net maps the floorplan mask plus the density map to a dense feature field and pools it per region to produce the encoding. The paper demonstrates the claim in three applications: region-wise interior space planning improves mean IoU and especially internal boundary regularity; triplet-based metric learning reaches 99.48 percent IoU-based accuracy; and a two-stage generation pipeline produces floorplans rated comparable to Graph2Plan by professional designers.","pith_inferences":["The four-directional density map is essentially an axis-aligned distance transform, so the method's advantage should shrink on floorplans dominated by slanted or curved walls; this is a testable prediction, and the paper's continuous-wall merging rule is only a partial mitigation.","URE behaves like a learned mid-level superpixel representation for floorplans; the same region-pooling recipe could transfer to other domains with boundary-aligned semantics, such as indoor scene layouts, building footprint vectorization, or document layout analysis, if a suitable density map is defined.","Because the partition depends only on geometry and not on semantic labels, URE could support cross-dataset or semi-supervised learning: the same unit regions can be reused with different label sets, an extension the paper does not explore."],"forward_implications":["For interior space planning, region-wise prediction with URE-Net raises internal Boundary F scores from 23.69 for DeepLabv3+ to 63.19, meaning functional-area boundaries come out far more regular without post-processing.","Adding the normalized density map as an input channel improves both pixel-wise and region-wise segmentation, by about 0.5 and 0.7 mIoU respectively, confirming that the geometric cue is doing real work.","In floorplan metric learning, URE-Net with the density map reaches 99.48 percent IoU-based triplet accuracy, above LayoutGMN's 97.54 percent, indicating room-wise learned encodings carry more shape information than hand-crafted graph nodes.","In floorplan generation, a simple two-step pipeline, wall mask via Pix2Pix then room instances via URE, produces plans that professional designers rate close to ground truth and comparable to Graph2Plan, without tedious vectorization.","The partition is far more compact than raster images: on 3D-FRONT, the 8 by 8, 0.5 meter strategy reaches 99.98 percent region-to-room IoU with roughly 125 regions on average, versus 24,137 pixels in a 256 by 256 image."],"supporting_citations":[{"why":"Supplies the 3D-FRONT dataset used for interior space planning training and testing and for the region-to-room IoU measurements.","marker":"[Fu et al. 2021]"},{"why":"Supplies the RPLAN dataset used for metric learning and generation, and serves as a baseline data-driven generation method.","marker":"[Wu et al. 2019]"},{"why":"LayoutGMN is the main metric-learning baseline whose training protocol and reported numbers the paper follows for comparison.","marker":"[Patil et al. 2021]"},{"why":"Graph2Plan is the baseline compared in the floorplan generation user study.","marker":"[Hu et al. 2020]"},{"why":"DeepLabv3+ is used as both a pixel-wise segmentation baseline and the backbone dense encoding network inside URE-Net.","marker":"[Chen et al. 2018]"},{"why":"PointNet supplies the symmetric max-pooling aggregation used to turn the set of region encodings into a whole-floorplan embedding.","marker":"[Qi et al. 2017]"},{"why":"Pix2Pix is used as the first-stage wall mask generator in the floorplan generation pipeline.","marker":"[Isola et al. 2017]"},{"why":"DefGrid is cited for the Boundary F-score evaluation convention used to measure boundary regularity in space planning.","marker":"[Gao et al. 2020]"}],"fun_headline_variants":["Boundary-adaptive regions outdo pixels and graphs on floorplan tasks","Unit region encoding: compact, flexible, and better for floorplan AI","Floorplan encoding that beats rasterized images and room graphs","Sliced unit regions yield cleaner planning and sharper generation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The representation assumes that floorplans are mostly formed by squeezing an axis-aligned box, so the four-directional density map lines up with rooms; if a floorplan has many slanted or curved walls, the density-based partition can drift away from the true room boundaries.","fun_headline_variants_meta":{"raw":{"variants":["Boundary-adaptive regions outdo pixels and graphs on floorplan tasks","Unit region encoding: compact, flexible, and better for floorplan AI","Floorplan encoding that beats rasterized images and room graphs","Sliced unit regions yield cleaner planning and sharper generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000202,"raw_usage":{"total_tokens":1338,"prompt_tokens":859,"completion_tokens":479,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":475,"completion_tokens_details":{"reasoning_tokens":406}},"tokens_in":475,"tokens_out":479,"duration_ms":5113,"temperature":1.0,"reasoning_tokens":406,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:39:03.576520+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a benchmark of floorplans with a high fraction of non-axis-aligned walls, compute the density-based unit regions, and measure region-to-room IoU: if the IoU falls far below the axis-aligned case, or the accuracy gains over pixel and graph baselines vanish, the central claim that the density map makes the encoding geometry-aware is false.","supporting_citations":[],"review_version":1}