{"id":"30d69efd-2d99-49eb-b7aa-dcade60a8430","arxiv_id":"1908.06702","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Floor-SP reconstructs floorplans from RGBD scans by optimizing polygonal room loops with sequential shortest-path solves, improving over FloorNet especially for non-Manhattan layouts.","lead":"This paper describes Floor-SP, a system that reconstructs building floorplans from 3D room scans by first finding rooms and then solving an optimization problem to trace each room's walls. The method works on difficult non-Manhattan layouts and outperforms the previous best system on a new dataset of 527 real apartments and houses.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 may compare Floor-SP on a new 94-scan panorama benchmark with FloorNet's published numbers from its original 155-scan smartphone-video benchmark; the paper never states that FloorNet was re-run on the same test split, so the headline Room/Room++ gains could be dataset/sensor confounds.","rationale":"I read the paper as a systems contribution whose main claim is a measured improvement over FloorNet. The optimization formulation, coordinate-descent shortest-path reduction, and honest failure discussion are all plausible, and the Mask R-CNN room-count limitation is stated clearly in Section 7.3; that limitation would cap Room++ but does not by itself invalidate the claimed improvement. The load-bearing uncertainty is whether the Table 1 FloorNet row is a matched baseline. The manuscript describes two different benchmarks (155 smartphone-video units for FloorNet vs 527/94 panorama scans for Floor-SP) and never specifies that FloorNet was trained or evaluated on the same split, nor how its RGBD-video pipeline was adapted to the top-down 4-channel input. Without that, the headline gains could be explained by dataset difficulty, scan quality, or annotation differences rather than by the room-wise shortest-path formulation. This is not an accusation; it is a missing experimental-protocol detail that is directly checkable. The paper also lacks error bars, which is secondary. If the authors confirm the baseline was re-run on the same data, I would endorse ACCEPT; as written, the claim needs that confirmation, so I move the verdict to CONDITIONAL.","tokens_in":9871,"tokens_out":6537,"duration_ms":70493,"concrete_test":"Inspect the experiment scripts or supplementary material for Table 1: was the FloorNet row produced by running FloorNet's released implementation on exactly the same 94 test scans? If yes, report the adaptation details and per-scan matched differences (e.g., paired Wilcoxon) for Room/Room++; if no, rerun FloorNet on Floor-SP's 94 test scans using its released model/code with a documented input conversion, and recompute Table 1. If FloorNet's numbers cannot be reproduced on this benchmark, restrict the headline comparison to a shared benchmark and state the dataset split explicitly.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is the significant boost over FloorNet in Table 1. The support for this claim is insecure at the measurement level. Section 2 identifies FloorNet's benchmark as 155 residential units with smartphone RGBD videos, while Section 7 evaluates Floor-SP on 94 test units from a separate set of 527 production-quality panorama RGBD scans. The text in Section 7.2 says only \"We follow FloorNet [20] and define metrics\"; it never says FloorNet was retrained or re-run on those 94 units, nor how FloorNet's video-based pipeline was adapted to the 4-channel top-down point-density/normal input. If the FloorNet row in Table 1 is taken from the original paper, then the reported gains (Room Prec/Rec 84.7/83.0 vs 81.2/72.1; Room++ 51.4/50.4 vs 42.3/37.5) compare different datasets, different sensors, different annotation protocols, and a deliberately harder Floor-SP test split with 199/667 non-Manhattan rooms. That would not demonstrate algorithmic superiority; it would only demonstrate that two numbers were measured on different benchmarks. The paper also gives no error bars or per-unit matched analysis, so even a same-benchmark comparison would benefit from paired reporting. The room-segmentation limitation that the Reader flagged is real but acknowledged and not unique to this comparison; the uncontrolled baseline is the more load-bearing threat to the central claim.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Floor-SP reconstructs a floorplan from aligned panorama RGBD scans as multiple polygonal loops, one per room, by minimizing an energy that combines DNN-based corner/edge data terms, a consistency term encouraging rooms to share corners and walls, and a model-complexity penalty. The optimization is performed by room-wise coordinate descent, where each step is reduced to a shortest-path problem on a pixel graph. The paper evaluates the approach on 527 production-quality panorama scans (433 training, 94 test), deliberately making the test set challenging with 199 of 667 rooms non-Manhattan, and reports substantial improvements over FloorNet on Room and Room++ precision/recall metrics. Ablation experiments show the contribution of the data term, consistency term, and multiple coordinate-descent rounds.","tokens_in":10104,"tokens_out":4820,"duration_ms":46654,"significance":"If the reported comparison is valid, Floor-SP is a meaningful advance for floorplan reconstruction, particularly for non-Manhattan scenes, which previous optimization-based approaches handled poorly. The formulation is elegant: reducing multi-loop floorplan inference to a sequence of shortest-path problems with data and consistency terms is a creative and potentially influential idea. The release of code and data on the project website is a concrete strength that enables reproducibility and follow-up work. However, the central claim of a significant performance boost over the state of the art is currently undercut by a potentially uncontrolled comparison: the FloorNet baseline in Table 1 may be taken from the original FloorNet paper's 155-unit smartphone-video benchmark rather than re-evaluated on the same 94-unit panorama test split. If so, the reported gains are confounded by dataset, sensor, annotation protocol, and difficulty differences. The paper also explicitly acknowledges that room instance segmentation errors on the number of rooms cannot be recovered, which limits the method's robustness but is at least stated in Section 7.3.","major_comments":[{"comment":"The headline comparison against FloorNet appears to mix benchmarks. The paper states that FloorNet was evaluated on 155 residential units with smartphone RGBD videos (Section 2), while Section 7 evaluates Floor-SP on a deliberately harder test split of 94 panorama scans with 199 of 667 non-Manhattan rooms. The text in Section 7.2 only says 'We follow FloorNet [20] and define metrics'; it never states that FloorNet was retrained or re-run on the same 94-unit test set, nor how FloorNet's video-based pipeline was adapted to the 4-channel top-down input. If the FloorNet row in Table 1 is the original published numbers, then the reported gains (Room 84.7/83.0 vs. 81.2/72.1; Room++ 51.4/50.4 vs. 42.3/37.5) compare different datasets, sensors, and difficulty levels, and do not support the central claim of a significant performance boost. Please clarify the experimental setup and, if the numbers are not from a matched evaluation, re-run a comparable baseline on the same test split or temper the claim accordingly.","section":"Section 7.2, Table 1"},{"comment":"The heuristic that fixes a start-edge and removes all intersecting edges to enforce loop containment breaks the local optimality of the coordinate-descent step, as the paper acknowledges. This means the reported energy minimization is not an exact solver of the stated objective; the claim that Floor-SP 'solves a shortest path problem to optimize the room structure' is therefore qualified. Since the start-edge selection depends on a corner candidates extracted from the likelihood map, the paper should at least analyze sensitivity to this choice (e.g., how often the selected start-edge is a true wall segment) or discuss the impact on the final energy. Without such analysis, the significance of the 'optimization' contribution relative to the prior heuristic system [17] is unclear.","section":"Section 5, 'Containment constraint satisfaction'"},{"comment":"The coordinate descent is run for exactly two rounds, with no convergence criterion and no analysis of whether two rounds suffice. The claim that 'room-wise coordinate descent adds a further boost' is based on a single fixed number of iterations, and Figure 7 shows only two energy values. The paper should either run the descent to convergence and report the energy trajectory, or justify the two-round choice with evidence that additional rounds do not change the results. This is directly related to the central claim that the method optimizes the energy rather than merely applying a fixed number of heuristic passes.","section":"Section 5, 'Sequential room-wise shortest path'"}],"minor_comments":[{"comment":"No error bars, variance estimates, or per-unit paired statistics are reported for any metric in Table 1. Given that the reported Room++ differences are roughly 9 percentage points, the paper would be much stronger with confidence intervals or a paired analysis over the 94 test units to show that the differences are not driven by a few outliers.","section":"Section 7.2, Table 1"},{"comment":"The claim that the approach 'does not require corner/edge detection with thresholds' is overstated. The system uses a DRN to estimate corner and edge likelihoods, then extracts corner candidates and a start-edge using those maps (Section 5), and Mask R-CNN with a lowered detection threshold (Section 6). While the method avoids explicit threshold-based corner/edge detection as a final representation, threshold-like decisions are present in the pipeline.","section":"Abstract and Section 1"},{"comment":"The figure reports energy values after each coordinate-descent round, but the y-axis is not labeled and the absolute scale of the energy is not defined. Please clarify what energy is plotted (the total objective from Section 4?) and include units or a scale so the reader can interpret the claimed improvement.","section":"Figure 7"},{"comment":"The paper should explicitly state in the conclusion that the method inherits the room-instance-segmentation errors and cannot split or create rooms, as it already does in the discussion. This is a known limitation, but it also means that the reported room-level metrics are partly determined by Mask R-CNN performance, so the paper should cite the Mask R-CNN room detection accuracy (e.g., how often the number of rooms is correct) to help the reader disentangle the contribution of the proposed optimization.","section":"Section 7.3"}],"recommendation":"major_revision","confidential_remarks":"The likely mismatch between the FloorNet baseline and the Floor-SP test benchmark is the most serious concern. If the authors can re-run FloorNet (or a re-implementation) on the same 94-unit split and still report the same advantage, the paper becomes a strong accept candidate. Otherwise, the central claim of state-of-the-art improvement is not supported. The paper's strengths (code/data release, non-Manhattan handling, clean formulation) are real, but the evaluation needs to be made apples-to-apples before publication. I also recommend the editor request the authors provide the energy-convergence analysis and sensitivity to the start-edge heuristic, as these directly bear on the 'optimization' claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Floor-SP proposes a floorplan reconstruction system that starts from Mask R-CNN room segments and solves a joint energy over multiple polygonal loops using room-wise coordinate descent, with each step reduced to a shortest-path problem. That formulation is new relative to the cited prior work, and the 527-scan production-quality dataset with a 94-scan test split is a real resource. The paper is well written, the ablations are informative, and the failure analysis in 7.3 is honest: they own the fact that a wrong room count from Mask R-CNN is unrecoverable.\n\nThe problem is Table 1. The FloorNet row appears to be taken from the original FloorNet paper, which evaluated on 155 smartphone RGBD video scans. Floor-SP is evaluated on 94 panorama scans from a different 527-unit set, with different sensors and annotation protocols. The paper never says FloorNet was retrained or rerun on the same split. If it was not, the reported gains (Room++ 51.4/50.4 vs 42.3/37.5) are not a comparison of algorithms; they are a comparison of datasets. This is the load-bearing claim of the paper, so the omission matters more than the other issues.\n\nThe rest of the concerns are secondary but real: no error bars or paired per-unit analysis, a single baseline, one train/test split, a heuristic containment step that the authors acknowledge breaks local optimality, and a fixed two rounds of coordinate descent with no convergence analysis. The room-segmentation limitation is a true limitation but reasonably scoped.\n\nI think the reader's overall verdict of accept is too generous given the measurement gap. The stress-test note is on target. The paper deserves a serious referee, but the editor should send it back with the explicit request that the authors either rerun FloorNet on their test split or substantially soften the comparison and rework the text. If the numbers are confirmed, this is a solid contribution; if not, the novelty of the formulation and dataset still makes it publishable, but the headline would have to change.","headline":"New formulation and dataset, but the headline FloorNet comparison may mix benchmarks and needs verification before the main claim is accepted.","tokens_in":10738,"tokens_out":2676,"would_cite":true,"duration_ms":25937,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A floorplan can be reconstructed from RGBD scans as a set of room loops by solving one shortest-path problem per room, and the paper shows this outperforms the prior state of the art.","keywords":["floorplan reconstruction","RGBD scans","room-wise coordinate descent","shortest path","energy minimization","instance segmentation","non-Manhattan floorplans","graph reconstruction"],"falsifier":"Take a test scan and merge two adjacent rooms into a single mask by editing the Mask R-CNN output while keeping all sensor inputs identical, then rerun Floor-SP. The paper's Section 7.3 discussion predicts the final graph will contain one combined room with an incorrect shared-wall topology, because the energy has no term that can split a room proposal; observing a recovered two-room graph would refute the claimed dependence on room-count correctness.","tokens_in":9556,"feed_emoji":"🏠","tokens_out":8056,"duration_ms":78349,"temperature":0.7,"pith_summary":"Floor-SP argues that floorplan reconstruction from aligned RGBD scans can be solved as an optimization problem over one polygonal loop per room, instead of the usual bottom-up detect-then-connect pipeline. The objective mixes learned per-pixel corner and wall costs, a penalty for loops that cut through room masks, a consistency term that rewards neighboring rooms sharing corners and walls, and a complexity term that prefers fewer corners. The optimization visits rooms one at a time and solves a shortest-path problem for each, keeping already reconstructed neighbors fixed. The paper reports that this scheme reconstructs rooms with precision/recall 84.7/83.0 and Room++ 51.4/50.4 on 94 test apartments, beating FloorNet's 81.2/72.1 and 42.3/37.5, and that it handles non-Manhattan structures by allowing each room its own subset of dominant directions. A reader would care because this offers a way to avoid error accumulation from early corner and edge detection while still producing complete room loops in challenging scans.","feed_headline":"Shortest paths, room by room, rebuild floorplans from scans","feed_subtitle":"A room-wise optimizer lifts Room++ recall to 50.4 percent from the previous 37.5 on 527 real RGBD scans.","key_machinery":"The load-bearing object is the pixel-graph shortest-path reduction. Every pixel is a node; an edge between pixels carries a weight that collects one half of the corner cost at its two endpoints, the learned edge cost and room-mask penalty along the Bresenham line, a consistency discount that rewards pixels already used by other loops, and a fixed per-corner complexity charge. Fixing a start-edge and forbidding crossings of a perpendicular start-line enforces that the loop surrounds its room mask. Room-wise coordinate descent visits rooms in increasing order of area and solves this shortest-path problem for each, then repeats a second round to fix overlaps. A dominant-direction precomputation restricts candidate edges to a subset of global Manhattan frames, which is how the system extends beyond a single global Manhattan assumption.","core_discovery":"The central claim is that wall-graph inference for floorplans reduces to joint energy minimization over room loops once room segments are available. Each room's loop is a closed pixel sequence with an unknown number of corners; the paper defines data terms from DNN corner and edge likelihood maps plus a room-mask interior penalty, a consistency term over shared corner and edge pixels across all loops, and a model-complexity term counting corners. Minimizing this objective by room-wise coordinate descent, where each step is a shortest-path problem on a pixel graph with edge weights derived from the energy, yields the floorplan graph after simple loop merging. The paper's evidence is the quantitative comparison on 527 production-quality panorama RGBD scans, with 94 held out for testing: room precision/recall rises to 84.7/83.0 and Room++ to 51.4/50.4, against 81.2/72.1 and 42.3/37.5 for the prior system FloorNet.","pith_inferences":["A direct extension the paper leaves implicit is to let the optimizer correct the room count, for example by generating candidate loops from the corner and edge likelihood maps and accepting or rejecting them under the same energy; that would attack the dominant failure mode identified in Section 7.3.","The coordinate-descent plus shortest-path pattern is not tied to floorplans; any reconstruction task whose output is a set of closed contours with learned per-pixel costs could reuse the reduction, such as building footprints, cadastral parcels, or cell boundaries in imagery.","Because the paper notes that high-resolution panoramas and full 3D points are discarded, an untested extension is to feed those signals into the data terms; the reported corner-recall gap in occluded areas suggests this is where the next gain would come from.","A testable prediction following from the paper's own failure analysis: on the subset of test units where the initial room segmentation has the correct number of rooms and roughly correct masks, room-level metrics should be near the top of the table, with almost all residual failures confined to non-Manhattan geometry and segmentation-count errors."],"forward_implications":["If Floor-SP's central claim is right, floorplan reconstruction no longer has to commit to thresholded corner and edge detections; an optimizer can place and connect corners from learned cost maps.","Because every room proposal is guaranteed a loop, room-level recall no longer collapses when a corner detector fires weakly on occluded walls, which is the mechanism behind the recall gains over FloorNet.","The ablation table in the paper shows that removing the consistency term lowers Room++ precision/recall from 51.4/50.4 to 49.4/48.5, so shared-wall agreement is a source of the gain rather than an artifact of post-processing.","The dominant-direction approximation means non-Manhattan rooms do not need one global Manhattan frame; each room can pick its own subset of up to four frames, which is why the system reports strong results on the non-Manhattan rooms in the test set.","Two rounds of coordinate descent can repair overlaps and holes that a single round leaves, which is why the paper fixes the optimization to two rounds."],"supporting_citations":[{"why":"FloorNet is the state-of-the-art baseline whose metric protocol and dataset evaluation Floor-SP adopts and outperforms.","marker":"[20]"},{"why":"Mask R-CNN supplies the room instance segments that reduce floorplan inference to multi-loop reconstruction and also creates the main failure mode.","marker":"[12]"},{"why":"Prior shortest-path loop reconstruction and containment heuristic that Floor-SP extends from one loop to multiple rooms.","marker":"[2]"},{"why":"Dilated Residual Networks produce the corner, edge, and wall-direction likelihood maps used in the data terms.","marker":"[29]"},{"why":"Structured indoor modeling is the room-segmentation-then-shortest-path pipeline that inspires Floor-SP's coordinate-descent design.","marker":"[17]"}],"fun_headline_variants":["Room-wise shortest paths solve floorplan reconstruction","Floorplan from scans: sequential shortest paths per room","No corner thresholds: DP guides room loops to floorplans","Floor-SP: dynamic programming for room-wise floorplans","RGBD to floorplan: each room a shortest path, then merge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, conceded in the paper's Section 7.3 discussion, is that the initial room-segmentation step already determined the correct number of rooms and a roughly correct mask for each one; if a room is missed or two rooms are merged into one segment, no later stage can split or create rooms, so the final floorplan topology is wrong regardless of how well the shortest-path optimization runs.","fun_headline_variants_meta":{"raw":{"variants":["Room-wise shortest paths solve floorplan reconstruction","Floorplan from scans: sequential shortest paths per room","No corner thresholds: DP guides room loops to floorplans","Floor-SP: dynamic programming for room-wise floorplans","RGBD to floorplan: each room a shortest path, then merge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000275,"raw_usage":{"total_tokens":1619,"prompt_tokens":898,"completion_tokens":721,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":639}},"tokens_in":514,"tokens_out":721,"duration_ms":7205,"temperature":1.0,"reasoning_tokens":639,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:36:50.910442+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a test scan and merge two adjacent rooms into a single mask by editing the Mask R-CNN output while keeping all sensor inputs identical, then rerun Floor-SP. The paper's Section 7.3 discussion predicts the final graph will contain one combined room with an incorrect shared-wall topology, because the energy has no term that can split a room proposal; observing a recovered two-room graph would refute the claimed dependence on room-count correctness.","supporting_citations":[{"cited_title":"Floornet: A unified framework for floorplan reconstruction from 3d scans","cited_arxiv_id":null,"evidence_quote":"FloorNet is the state-of-the-art baseline whose metric protocol and dataset evaluation Floor-SP adopts and outperforms."},{"cited_title":"Girshick","cited_arxiv_id":null,"evidence_quote":"Mask R-CNN supplies the room instance segments that reduce floorplan inference to multi-loop reconstruction and also creates the main failure mode."},{"cited_title":"Piecewise planar and compact floorplan reconstruction from images","cited_arxiv_id":null,"evidence_quote":"Prior shortest-path loop reconstruction and containment heuristic that Floor-SP extends from one loop to multiple rooms."},{"cited_title":"Funkhouser","cited_arxiv_id":null,"evidence_quote":"Dilated Residual Networks produce the corner, edge, and wall-direction likelihood maps used in the data terms."},{"cited_title":"Struc- tured indoor modeling","cited_arxiv_id":null,"evidence_quote":"Structured indoor modeling is the room-segmentation-then-shortest-path pipeline that inspires Floor-SP's coordinate-descent design."}],"review_version":1}