Unit Region Encoding: A Unified and Compact Geometry-aware Representation for Floorplan Applications

Huichao Zhang; Manyi Li; Pengyu Wang; Yaguang Wu; Zuojun Li

REVIEW 3 major objections 5 minor 2 references

Unit Region Encoding: A Unified and Compact Geometry-aware Representation for Floorplan Applications

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict A promising floorplan representation whose central empirical claims are undercut by inconsistent tables and a strong baseline the paper doesn't address. read the letter →

arxiv 2501.11097 v1 pith:PAZRSQI7 submitted 2025-01-19 cs.CV

Huichao Zhang , Pengyu Wang , Manyi Li , Zuojun Li , Yaguang Wu This is my paper · ORCID

classification cs.CV

keywords floorplanrepresentationunitregionencodinggeometry-awaredensitymapinteriorspaceplanningmetriclearninggenerationpoolingURE-Net

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

The 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.

The 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.

These 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.

Referee Report

3 major / 5 minor

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.

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 (3)

[§4.1, Tables 2–4] 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.
[§4.1, Table 3] 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.
[§3.1–3.2] 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.

minor comments (5)

[§4.1] 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.
[§3.3] 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.
[§4.3] 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.
[§4.1] 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.
[Throughout] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the unit region partition and encodings are derived from the floorplan boundary and trained with task-specific supervised losses, not from the downstream ground-truth labels.

full rationale

The paper's derivation chain is self-contained rather than circular. Section 3.1 defines the density map directly from the floorplan boundary (Eq. 1-2: f(x) = 1 / sum of nearest distances to the boundary along four directions), with no dependence on room labels or downstream predictions. Section 3.2 builds the unit region partition by clustering this density map and then slicing the density regions under a user-specified splitting strategy; the strategy is evaluated by IoU against ground-truth rooms in Table 1, but that is an evaluation of representation quality, not a definition of the representation in terms of the target. Section 3.3 trains URE-Net with standard region-wise cross-entropy or triplet losses, so the latent encodings are learned outputs rather than inputs. In the interior space planning experiments, the comparison to pixel-wise DeepLabv3+ and the ablations are empirical and do not assume the conclusion. The metric-learning application does take the given room partition as unit regions, but the encodings are still learned from dense features and the triplet loss; this is an application choice, not a self-definitional reduction. The reported discrepancies between Table 2 and Table 3 and the strong pixelwise+voting baseline are internal-consistency or empirical-evidence concerns, not circularity: they do not show that any prediction equals its input by construction. No load-bearing self-citation or imported uniqueness theorem appears; the LayoutGMN comparison cites a paper with an overlapping author, but the cited numbers are used as an external baseline and the claimed superiority is an empirical comparison rather than a derivation from that citation.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several hand-chosen parameters (grid size, minimum region size, clustering thresholds, and a grouping threshold in generation) that are selected on validation data and not fully reported. The geometric motivation depends on an unproven axis-aligned 'squeeze' model. No new physical or computational entities are postulated.

free parameters (4)

splitting strategy (M,N,h) = (8,8,1m) for planning and generation; room-based for metric learning
The grid size and minimum region size are chosen by hand and validated through ablations in Section 4.1. The choice changes mIoU by about 2 points, so the reported performance depends on this selection.
density region clustering parameters = not specified
Section 3.2 describes clustering connected locations with the same density value, but does not state the clustering algorithm, the number of clusters, or discretization thresholds. These unspecified choices affect the unit region partition and are not reproducible.
region grouping threshold in generation = not specified
In Section 4.3, unit regions are assigned to the same room when the Euclidean distance between their encodings is below a threshold. The threshold value is not reported, so the generation pipeline cannot be reproduced.
category merge = 9 to 5 categories (planning); 13 to 5 categories (generation)
The authors merge rare functional-area and room types into 'others' to simplify the tasks. This changes the label distribution and all reported metrics, and the merge criteria are subjective.

assumptions (4)

domain assumption Axis-aligned squeeze model: every floorplan is obtainable from an axis-aligned bounding box by inward boundary pushes.
Section 3.1 introduces this assumption to motivate the four-directional density map. It is not proven and is violated by slanted or curved walls.
domain assumption Density map clustering yields meaningful regions: connected locations with equal density values form regions that align with room structure.
Section 3.2 relies on this for the unit region partition. The paper provides empirical IoU in Table 1, but the clustering is underspecified and no formal justification is given.
standard math Region-wise RoI pooling and a shared MLP produce adequate encoding features for all applications.
This is standard deep learning machinery used in Section 3.3 and is not questioned by the paper.
domain assumption RPLAN and 3D-FRONT annotations are reliable ground truth for room and functional-area labels.
The paper uses RPLAN from Wu et al. 2019 and its own newly collected 3D-FRONT functional annotations. The new annotations are not released, so their quality cannot be independently checked.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Unit Region Encoding: A Unified and Compact Geometry-aware Representation for Floorplan Applications." pith.science (2026). https://pith.science/paper/PAZRSQI7

@misc{pith2026250111097,
  author       = {Pith},
  title        = {Pith review of: Unit Region Encoding: A Unified and Compact Geometry-aware Representation for Floorplan Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PAZRSQI7}},
  note         = {Machine review of arXiv:2501.11097}
}

read the original abstract

We present the Unit Region Encoding of floorplans, which is a unified and compact geometry-aware encoding representation for various applications, ranging from interior space planning, floorplan metric learning to floorplan generation tasks. The floorplans are represented as the latent encodings on a set of boundary-adaptive unit region partition based on the clustering of the proposed geometry-aware density map. The latent encodings are extracted by a trained network (URE-Net) from the input dense density map and other available semantic maps. Compared to the over-segmented rasterized images and the room-level graph structures, our representation can be flexibly adapted to different applications with the sliced unit regions while achieving higher accuracy performance and better visual quality. We conduct a variety of experiments and compare to the state-of-the-art methods on the aforementioned applications to validate the superiority of our representation, as well as extensive ablation studies to demonstrate the effect of our slicing choices.

Figures

Figures reproduced from arXiv: 2501.11097 by the authors.

**Figure 1.** We present the Unit Region Encoding of floorplans, which is a unified and compact geometry-aware encoding representation for various applications. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** The visualization of the proposed density map and unit region partition under different splitting strategies. The last picture shows the uniform partition [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** The statistics and a visualization of the collected functional area [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Qualitative comparison of the interior space planning application. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 6.** Figure 6: The ground-truth floorplan and the generated results of our method [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 5.** Figure 5: The floorplan retrieval comparison between our method and Layout [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[2]

In European Conference on Computer Vision

House-gan: Relational generative adversarial networks for graph-constrained house layout generation. In European Conference on Computer Vision . Springer, 162–177. Nelson Nauata, Sepidehsadat Hosseini, Kai-Hung Chang, Hang Chu, Chin-Yi Cheng, and Yasutaka Furukawa. 2021. House-GAN++: Generative Adversarial Layout Refinement Network towards Intelligent Com...

work page 2021
[2020]

ACM Trans- actions on Graphics (TOG) 39, 4 (2020), 118–1

Graph2plan: Learning floorplan generation from layout graphs. ACM Trans- actions on Graphics (TOG) 39, 4 (2020), 118–1. Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. 2017. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE confer- ence on computer vision and pattern recognition . 1125–1134. Arun Ku...

work page 2020

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [2]

In European Conference on Computer Vision

House-gan: Relational generative adversarial networks for graph-constrained house layout generation. In European Conference on Computer Vision . Springer, 162–177. Nelson Nauata, Sepidehsadat Hosseini, Kai-Hung Chang, Hang Chu, Chin-Yi Cheng, and Yasutaka Furukawa. 2021. House-GAN++: Generative Adversarial Layout Refinement Network towards Intelligent Com...

work page 2021

[2] [2020]

ACM Trans- actions on Graphics (TOG) 39, 4 (2020), 118–1

Graph2plan: Learning floorplan generation from layout graphs. ACM Trans- actions on Graphics (TOG) 39, 4 (2020), 118–1. Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. 2017. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE confer- ence on computer vision and pattern recognition . 1125–1134. Arun Ku...

work page 2020