Indoor 3D scanners and multi-view stereo, which recovers depth by matching pixels across overlapping photos, leave holes wherever walls are blank, shiny, or hidden behind furniture. Floorplan reconstruction turns these patchy scans into clean vector layouts: rooms as closed polygons, walls as straight segments. Two families of methods compete here. Fast neural networks predict the layout in one pass but cannot guarantee that it is physically valid. Exact integer programs guarantee closed, non-overlapping rooms but become too slow to run as buildings grow. We propose to combine them. A feedforward network gives a complete first draft. A search procedure then repeatedly erases a small cluster of neighbouring rooms, chosen where the scan evidence is weakest or the rooms disagree most, and re-solves only that cluster exactly while the rest of the building stays frozen.
The search procedure is Large Neighborhood Search (LNS), an optimization method that alternates "destroy" and "repair" steps on a full solution. It is standard in vehicle routing and mixed-integer programming, but no one has used it in structured indoor reconstruction. The frozen surroundings are what make the small exact solve fast. Architecture supplies a natural shape for the neighbourhoods: rooms share walls, and walls mostly follow a few dominant directions. We use a loose version of the Manhattan and Atlanta-world priors, meaning walls prefer a small set of vertical-plane orientations but are not forced onto them. The claim is that this loop keeps the speed of neural prediction on easy scenes. On hard real scans and on buildings larger than the network's room budget, it should produce layouts that are both more accurate and always topologically valid.
Core Claim
A feedforward floorplan predictor followed by LNS with exact, small integer-program repairs should beat three alternatives on real, occluded, non-Manhattan scans and on scenes too large for fixed query budgets. The first alternative is the network plus heuristic post-processing. The second is a single global integer program given the same time. The third is room-by-room coordinate descent. The key ingredient is joint re-optimization of two to four adjacent rooms against frozen context, with candidate walls allowed to enter during repair.
This is distinct from each closest method in a specific way. Floor-SP re-solves one room at a time as a shortest path with all others fixed, so a shared wall that is wrong for two rooms at once can never be fixed. It averages 785 seconds per scene. MonteFloor runs a Monte Carlo tree search that builds layouts constructively from a fixed pool of room proposals. It never takes an existing full layout apart, and it cannot create walls that are missing from the pool. Building Floorplan Reconstruction Based on Integer Linear Programming writes a loose-Manhattan energy with confidence and topology terms and solves it once, globally. We reuse its energy almost unchanged as the repair objective, but only ever solve it on small regions. RoomFormer and CAGE predict polygons directly in about 0.01 seconds. RoomFormer is capped at 20 rooms and 40 vertices, and CAGE needs heuristic edge intersection to close polygons. We use them as the initializer, not the final answer.
Research Program
The representation is a planar graph of candidate wall segments over a 2D projection of the scan. Each room is a cycle in this graph. The candidate pool is the union of three sources: edges predicted by the network, wall planes found by RANSAC (a robust plane-fitting method), and lines through observed corners along the scene's dominant horizontal directions. These directions are estimated Atlanta-style, with gravity plus several free horizontal orientations. The third source is what lets a repair add a wall that plane fitting missed because the wall was occluded. That was an explicit failure of the ILP paper.
Let binary variable $x_e$ indicate that candidate segment $e$ is used as a wall. The repair solves
$$\min_{x} \; \sum_e x_e \big( \lambda_d\, D_e + \lambda_a\, A_e + \lambda_c \big) \quad \text{s.t. rooms are closed, non-overlapping, and share coincident walls,}$$
subject to all variables outside the destroyed region being fixed. The terms are as follows. $D_e$ is a data cost from point density and free-space evidence along $e$. $A_e$ is the angular deviation of $e$ from its nearest dominant direction, which is the loose-Atlanta term. $\lambda_c$ is a per-segment complexity cost. The $\lambda$ weights are shared across all scenes. The hard constraints are what feedforward methods lack.
The destroy operators are where multi-view stereo failure modes enter. One operator targets rooms whose walls sit in evidence gaps: low density, or high disagreement between depth hypotheses, which matches the "information-limited uncertainty" separated out in Joint Modeling of Corruption-Driven and Information-Limited Uncertainty for Robust 3D Gaussian Splatting. A second targets pairs of adjacent rooms whose shared wall has high energy. A third targets junctions where the network's topology is invalid. A fourth is random, as a control. Following Balans, a multi-armed bandit, which learns online which option pays off most, chooses among the operators. Following the finding in Anytime Multi-Agent Path Finding via Large Neighborhood Search that fast repairs covering many neighbourhoods beat slow optimal ones, each repair runs under a short time cap. Neighbourhood size, from one to four rooms, is itself a bandit arm, because the right destroy size is known to be empirical.
We will build this on SCIP for the repair and CAGE and RoomFormer as initializers. We will then measure energy-versus-wall-clock curves, final accuracy, and the validity of every output.
Baselines
The naive reference point is the initializer alone: CAGE's output with its own heuristic post-processing. It sets the floor on accuracy. It also shows how often feedforward output breaks hard constraints, measured as self-intersections, overlaps, or unclosed rooms. On Structured3D, CAGE reports room F1 of 99.1% and corner F1 of 91.7%. On cross-dataset SceneCAD it reports room IoU of 85.6%, against 74.0 for RoomFormer. CAGE is the strongest competitor on accuracy and speed, and we will run the released model unchanged on identical density maps. Beating it means higher room and corner scores on SceneCAD and the Floor-SP test set, with zero constraint violations. On Structured3D we expect only parity, since there is little headroom above 99.1%.
The strongest search-based competitor is MonteFloor. It reports room precision of 0.96 and corner recall of 0.77 on Structured3D in 71 seconds, against 0.89, 0.73, and 785 seconds for Floor-SP. We will compare at MonteFloor's 71-second budget and at matched budgets of 10 and 30 seconds. Beating it means equal or better room precision and higher corner recall at 71 seconds or less. On the Floor-SP test set, Floor-SP reports a Room++ score of 51%. No per-metric MonteFloor accuracy on that set was found on Lacuna, so we will rerun it where the code allows.
Two further baselines use our own energy, candidate pool, and warm start, so that only the search strategy differs. The first is a single global SCIP solve of the full energy, with SCIP's built-in primal heuristics enabled and the same wall-clock limit. The ILP paper reports no numbers on Lacuna; its GibLayout and FloorNet results are qualitative only. The second is Floor-SP-style coordinate descent, which is our LNS restricted to one-room neighbourhoods. Two ablations isolate the key ingredients. One uses random rather than evidence-targeted destroy. The other removes corner-line candidates, which tests whether adding new walls matters.
Why not just…?
Why not just use CAGE with better post-processing? This is a real threat on clean synthetic data, and we expect no gain on Structured3D. The case rests on two regimes. The first is real, occluded scans, where RoomFormer already drops to 74.0 room IoU on SceneCAD. The second is buildings with more rooms than the query budget. We will build the second from Structured3D houses stitched into multi-unit floors of 30 to 80 rooms, where the network must tile its input and seams appear. If CAGE plus snapping matches us on SceneCAD and on the stitched scenes, the proposal has failed, and we will report that.
Why not just give the global integer program the same time? SCIP already runs local-branching and neighbourhood heuristics internally, so this baseline could close most of the gap. Balans found that an outer LNS layer cut SCIP's primal gap by more than 75% on hard MIPLIB instances, but floorplan energies may be easier than those instances. Energy-versus-time curves, compared across scene sizes with the same warm start, settle the question directly. We expect the global solve to match us on small apartments and to fall behind as the candidate count grows.
Why not just run Floor-SP's room-by-room descent from a good initialization? Starting from CAGE rather than Mask R-CNN proposals may remove most of Floor-SP's local-minimum problem. The one-room-neighbourhood ablation tests exactly this. Joint multi-room destruction must lower final energy and raise corner recall specifically on scenes with non-orthogonal shared walls. Nearly half of the Floor-SP scans contain such walls.
Proposed Experiments
The main evaluation covers Structured3D, SceneCAD, the 100-scene Floor-SP test set, and the stitched large-building set. It reports room, corner, and angle scores, the constraint-violation rate, and runtime. The decisive checks are three. First, on SceneCAD and Floor-SP scans, LNS must improve corner F1 over CAGE with zero invalid layouts. Second, on stitched buildings, accuracy must degrade gracefully while CAGE and the global solve do not. Third, per-operator attribution must show that evidence-gap destruction produces most of the accepted improvements. If random destruction does just as well, the link to multi-view stereo failure modes is not supported.
A secondary check tests whether the repaired layout helps dense reconstruction. We will pass the recovered wall planes as the planar prior to AtlasGS on ScanNet scenes, in place of its segmentation-derived wall normals. We then measure surface F-score against ManhattanSDF's reported 0.602 from Neural 3D Scene Reconstruction with the Manhattan-world Assumption. No ScanNet F-score for AtlasGS itself was found on Lacuna, so we will rerun it. A gain here would show that fixing structure discretely addresses the segmentation sensitivity both continuous methods share.
References
- Floor-SP: Inverse CAD for Floorplans by Sequential Room-wise Shortest Path
- MonteFloor: Extending MCTS for Reconstructing Accurate Large-Scale Floor Plans
- Building Floorplan Reconstruction Based on Integer Linear Programming
- Connecting the Dots: Floorplan Reconstruction Using Two-Level Queries
- CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction
- Joint Modeling of Corruption-Driven and Information-Limited Uncertainty for Robust 3D Gaussian Splatting
- Balans: Multi-Armed Bandits-based Adaptive Large Neighborhood Search for Mixed-Integer Programming Problems
- Anytime Multi-Agent Path Finding via Large Neighborhood Search
- AtlasGS: Atlanta-world Guided Surface Reconstruction with Implicit Structured Gaussians
- Neural 3D Scene Reconstruction with the Manhattan-world Assumption