acceptodds
Under review as a conference paper at ICLR 2027

FlatLands: Indoor floormap completion from a single egocentric view

Abstract

A single egocentric image typically captures only a small portion of the floor, yet a complete metric traversability map of the surroundings would better serve applications such as indoor navigation. We introduce FlatLands, a dataset and benchmark for single-view indoor floormap completion in metric bird’s-eye view (BEV) conditioned on a single egocentric view of the room. The dataset contains 270,575 observations from 17,656 scene records across six metric 3D datasets, with aligned observation, visibility, validity, and ground-truth maps and both in- and out-of-distribution protocols. We compare training-free, deterministic, ensemble, and generative methods on shared geometry-derived inputs and, without fine-tuning, on maps estimated from real egocentric RGB images. Our results show that predicting floor almost everywhere can achieve high IoU when most evaluated cells are floor, while generative sample sets outperform the best single-output model under a proper scoring rule and support safer risk-weighted planning. FlatLands establishes a rigorous testbed for evaluating both floor-map accuracy and how effectively alternative completions represent the ambiguity of unseen space.

Primary area:
datasets and benchmarks
Keywords:
binary completion benchmark, computer vision for robotics, generative completion, embodied AI

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_HoWUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

I am not an expert in this domain. Based on my understanding, this paper presents FlatLands, a benchmark for completing unseen floor regions from a single egocentric view. It combines six indoor datasets into approximately 270,000 observations and compares eleven methods, including deterministic and generative models. The evaluation covers map completion, transfer to RGB-derived observations, and offline path planning. The main contribution is the dataset and evaluation framework for studying multiple plausible floor maps.

Critical strengths
  1. The paper addresses an important problem: a single view provides incomplete information about the surrounding environment. Representing several possible maps is well motivated, particularly when incorrect predictions could affect navigation.
  2. The shared representation and evaluation protocol allow different completion methods to be compared across several datasets. This could provide a useful benchmark for studying completion quality and its relationship to planning performance.
Critical weaknesses
  • The planning comparison does not clearly isolate the benefit of multiple maps. Generative models use probability-based traversal costs, while U-Net uses a thresholded map with uniform costs. Therefore, the improvement may partly come from the planning strategy rather than the generated maps.

  • The evidence for recovering multiple plausible layouts is limited. In Table 15, All-Floor achieves the same coverage as the generative methods and a lower symmetric distance. This example does not convincingly demonstrate the advantage of generating multiple solutions.

Questions for authors
  • Would the planning improvement remain if U-Net’s predicted floor probabilities were used with the same traversal costs and cell-admission rules as the generative methods?
  • Can the authors provide additional evidence that the generated samples capture distinct plausible layouts? In Table 15, a single All-Floor prediction already matches the reported coverage of the generative methods.
7 comments
AuthorsOpenReview ↗

Weakness 1: Separating the effect of the maps from the planning strategy

Thank you for the thoughtful review and for recognising the motivation for representing several possible maps. The comparison you proposed helped us isolate the benefit of multiple maps.

The planning comparison does not clearly isolate the benefit of multiple maps.

To isolate it, we fix the queries (4,000), the graph (observed floor plus all hidden valid cells), and the costs, and vary only the number of maps each generator uses, averaging over all subsets of its four saved maps. This graph admits more cells than the method comparison in Question 1.

Samples usedDiffusion successFlow Matching successFM+XAttn success
182.85%82.84%83.41%
283.30%83.51%84.02%
383.49%83.82%84.23%
483.68%83.98%84.88%

Four FM+XAttn maps raise success by 1.47 percentage points over one (paired 95% interval over physical environments 0.81 to 2.21).

Generative models use probability-based traversal costs, while U-Net uses a thresholded map with uniform costs. Therefore, the improvement may partly come from the planning strategy rather than the generated maps.

The two controls separate the effects. With the same allowed cells and costs, both U-Net probabilities (84.75%) and FM+XAttn frequencies (83.28%) improve over unit costs (81.20%; Question 1); with everything else fixed, four maps beat one (above). Retaining ambiguity, as per-cell probabilities or as multiple maps, improves planning.

We will add both controls beside the planning results, with their protocols in the appendix.

AuthorsOpenReview ↗

Weakness 2: Whether Table 15 demonstrates distinct layouts

The evidence for recovering multiple plausible layouts is limited.

Table 15 illustrates coverage on one scene at a 0.10 Jaccard-distance tolerance. The set-level evidence is in Question 2.

In Table 15, All-Floor achieves the same coverage as the generative methods and a lower symmetric distance.

Four of the five targets lie within 0.10 of All-Floor, which accounts for its coverage and symmetric distance:

MethodTargets covered out of 5Symmetric distance ↓
All-Floor40.051
Diffusion40.054
Flow Matching40.054
FM+XAttn40.069

The disagreement-cell test and the repeated-sample comparison in Question 2 do not depend on this tolerance, and we will add both beside Table 15.

This example does not convincingly demonstrate the advantage of generating multiple solutions.

On the cells where the targets disagree, any single map reused for both targets scores exactly 50%, while sample pairs from each generator score 60.9 to 66.2% (Question 2). The sample-count control (Weakness 1) shows the same advantage in planning, where four maps beat one for every generator.

AuthorsOpenReview ↗

Question 1: Does the improvement remain when U-Net uses the same costs and admission rules?

Would the planning improvement remain if U-Net's predicted floor probabilities were used with the same traversal costs and cell-admission rules as the generative methods?

Yes, the improvement from probability-weighted costs remains. With identical allowed cells and costs, U-Net probabilities reach 84.75% success, FM+XAttn frequencies 83.28%, and unit costs 81.20%. Multiple maps add a separate gain: four maps beat one for every generator (Weakness 1).

Protocol: 4,000 fixed queries from 500 geometry-derived observations. All methods share observed floor plus the hidden valid cells that at least one FM+XAttn sample labels floor, and the goal is always admitted. Search is eight-connected Dijkstra without diagonal corner cutting. Observed floor costs 1, and hidden cells cost 1−log⁡max⁡(0.05,p)1-\log\max(0.05,p), where pp is a U-Net probability or FM+XAttn frequency; each move costs its length times the mean endpoint cost. Success requires the route to lie entirely on valid reference floor.

Cost sourceOverall successSuccess among returned routes
Unit costs81.20%84.83%
U-Net probabilities84.75%88.53%
FM+XAttn83.28%86.99%

Thank you for this question: it helped us establish that richer per-cell probabilities help in any case, whether they come from deterministic model's output probabilities or from direct outputs of generative models. This will most definitely be presented in the extended main paper.

Reviewer_DMVVOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces FlatLands, a benchmark for single-view indoor floormap completion, comprising 270,575 observations from six 3D datasets. It evaluates eleven deterministic and generative methods using completion accuracy, uncertainty-aware metrics, and offline navigation performance. Results suggest that generative methods better capture spatial uncertainty and improve risk-aware path planning.

Critical strengths

FlatLands consolidates six existing 3D datasets into a unified benchmark with 270,575 observations, providing a standardized setting for evaluating indoor floormap completion.

Critical weaknesses
  1. The necessity of predicting large unseen regions from a single image is not convincingly justified, particularly given existing multi-view reconstruction, SLAM, and pretrained vision models. The paper does not clearly establish what distinctive research problem this benchmark addresses or why its results would be useful for real-world navigation.

  2. The dataset does not clearly distinguish physical floor surfaces, obstacles, and reconstruction artifacts. Figure 2 shows unexplained holes and irregular regions in the ground truth. More fundamentally, traversability depends on the robot's size, mobility, and ability to negotiate obstacles, none of which are specified. Consequently, the reported Safe Success measures agreement with potentially ambiguous binary labels rather than demonstrated navigation safety.

Questions for authors
  1. What is the practical advantage of single-view floormap completion over existing multi-view mapping, SLAM, and pretrained vision-based approaches? What specific research gap does FlatLands address?

  2. How are the complete floormaps (F⋆F^\star) generated and validated? In Figure 2, do the numerous holes and irregular boundaries represent actual obstacles, missing geometry, or reconstruction artifacts?

  3. What constitutes a traversable floor, and for which type of robot? How are small objects, movable obstacles, carpets, and spaces beneath furniture handled when constructing the binary ground truth?

  4. If the ground-truth labels do not account for robot-specific traversability, how should the reported Safe Success be interpreted as a measure of navigation safety? What evidence supports its relevance to real-world navigation?

7 comments
AuthorsOpenReview ↗

Weakness 1: Why predict unseen floor from one view?

Thank you for the thorough review. Your questions prompted a new footprint evaluation.

The necessity of predicting large unseen regions from a single image is not convincingly justified, particularly given existing multi-view reconstruction, SLAM, and pretrained vision models.

A robot often has to choose its next move or observation before the relevant floor is seen. SLAM and multi-view reconstruction integrate only the measurements already taken (Cadena et al., 2016), and pretrained depth or segmentation models describe only visible surfaces; the same visible evidence can fit different hidden layouts. Completion estimates the unseen floor, a role studied in Occupancy Anticipation (Ramakrishnan et al., 2020) and MapEx (Ho et al., 2025).

The paper does not clearly establish what distinctive research problem this benchmark addresses or why its results would be useful for real-world navigation.

FlatLands benchmarks the prediction of unseen floor, including alternative layouts when one observation does not determine a unique answer. It evaluates single-map accuracy, prediction-set quality, held-out-source generalisation, and transfer to image-derived inputs across six indoor datasets, and links completion errors to route validity and goal reachability through planning. We will make this clearer with the expanded revision.

AuthorsOpenReview ↗

Weakness 2: Ambiguous labels and robot-specific traversability

The dataset does not clearly distinguish physical floor surfaces, obstacles, and reconstruction artifacts.

The target F⋆F^\star marks known floor. The validity mask VV excludes cells with no reconstructed geometry; every other cell is labelled floor or non-floor. The unobserved mask UU marks cells not seen in the input. Training and evaluation use only Reval=U⊙VR_{\mathrm{eval}}=U\odot V, so cells without geometry are never scored (Section 3, Appendix B.3).

Figure 2 shows unexplained holes and irregular regions in the ground truth.

Holes inside VV are cells not labelled floor: obstacles such as furniture, and cells without a floor label, which are non-floor by construction. Cells with no reconstructed geometry lie outside VV and are not scored. We will add a Figure 2 panel showing how source geometry, labels, and masks align.

More fundamentally, traversability depends on the robot's size, mobility, and ability to negotiate obstacles, none of which are specified.

Table 6 models a planar robot occupying one 3.9 cm grid cell (Appendix A.5); Question 3 adds circular robots of any radius. We will release code, model weights, and scene caches from which users can build navmeshes for their own robot dimensions.

Consequently, the reported Safe Success measures agreement with potentially ambiguous binary labels rather than demonstrated navigation safety.

Safe Success is valid-route success: whether a planned route reaches its goal on valid reference floor for the specified robot model. It measures the planning consequences of completion errors (Question 4).

AuthorsOpenReview ↗

Question 1: What practical advantage does completion provide?

What is the practical advantage of single-view floormap completion over existing multi-view mapping, SLAM, and pretrained vision-based approaches?

Completion lets a planner act on likely floor before it is observed. We test this in a closed-loop sensing simulation: 240 trials in 240 physical environments, one controller, a synthetic floor sensor, and supplied hidden goals. The controller moves only on confirmed floor and scans before entering unknown cells; each scan updates the observation. Only the cost of unknown cells differs between conditions: unit costs, U-Net probabilities, or FM+XAttn sample frequencies. Budgets are ten extra scans and 300 cells of travel.

Goals reached out of 240, by scan budget:

Planning costs2 scans4 scans6 scans8 scans10 scans
Unit costs83156181194200
FM+XAttn83162183202212
U-Net92167194198203

With ten scans, completion-derived costs reach 212 goals with FM+XAttn and 203 with U-Net, against 200 with unit costs. Completion complements SLAM: SLAM supplies the observed map that completion extends.

What specific research gap does FlatLands address?

FlatLands provides a common benchmark for unseen-floor completion that compares single-map accuracy and prediction-set quality across six indoor datasets, independent of the sensing front end and the robot (Weakness 1).

Reviewer_zF8zOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces FlatLands, a new dataset and benchmark for single-view floor map completion. The dataset contains large scenes from six indoor datasets, including in-distribution and out-of-distribution scenes, and focuses on evaluating unobserved areas that are either occluded or outside the camera's view. Furthermore, the authors evaluate training-free, deterministic, ensemble, and generative approaches on FlatLands. The overall results demonstrate that generative models outperform deterministic models.

Critical strengths
  1. This paper provides a well-defined benchmark for evaluating predictions of unobserved areas.
  2. The authors offer abundant experimental results and detailed descriptions of the dataset and experimental settings.
Critical weaknesses
  1. The authors provide ScanNet++ as an out-of-distribution dataset. However, in Table 2, the All-Floor result illustrates high performance on the OOD dataset. Although the authors already provided challenging cases, I wonder whether those challenging cases are from the in-distribution or out-of-distribution dataset. If the cases are from the in-distribution dataset, would it be possible to provide some challenging case results from the OOD dataset?
  2. Based on my understanding, this paper focuses on the unobserved areas, but in real situations, robots mostly rely on both observed and unobserved areas to make decisions. I suggest that evaluations containing complete floor maps can be more comprehensive.
Questions for authors
  1. In Figure 6, I wonder why the UMR starts to increase after rcond > 0.5. Could you please explain?
6 comments
AuthorsOpenReview ↗

Weakness 1: Challenging OOD cases and All-Floor

Thank you for the careful review and for your positive assessment of the benchmark and its experiments.

The authors provide ScanNet++ as an out-of-distribution dataset. However, in Table 2, the All-Floor result illustrates high performance on the OOD dataset.

All-Floor's IoU equals the floor fraction of the hidden valid region, so its 0.854 on ScanNet++ reflects high floor prevalence. The hard subsets test cases where this shortcut is weak.

Although the authors already provided challenging cases, I wonder whether those challenging cases are from the in-distribution or out-of-distribution dataset.

The Structurally Challenging Subset contains both: 1,221 ID and 165 OOD observations.

If the cases are from the in-distribution dataset, would it be possible to provide some challenging case results from the OOD dataset?

The table below reports them separately. Both subsets have at most 20% of target floor observed and less than 50% floor in the hidden valid region; the 165 OOD cases come from 74 ScanNet++ environments.

MethodHard ID IoU ↑Hard OOD IoU ↑
All-Obstacle0.0000.000
All-Floor0.1860.232
Uniform Random0.1470.176
NN Propagation0.1820.268
U-Net0.4690.466
PConv-UNet0.4580.455
LaMa0.3240.271
LaMa-Ensemble0.1820.230
Diffusion0.4070.460
Flow Matching0.3660.394
FM+XAttn0.4000.412

We will add these tables to the appendix and reference them in the main paper.

AuthorsOpenReview ↗

Weakness 2: Evaluating complete floor maps

Based on my understanding, this paper focuses on the unobserved areas, but in real situations, robots mostly rely on both observed and unobserved areas to make decisions. I suggest that evaluations containing complete floor maps can be more comprehensive.

We computed IoU over all valid cells, observed and hidden, with inputs and predictions unchanged. Each entry gives hidden-region IoU → whole-map IoU for the 12,129 ID and 16,214 ScanNet++ test observations and the 165 hard ScanNet++ cases. Generators use their first sample.

MethodIDScanNet++Hard ScanNet++
All-Obstacle0.000 → 0.3430.000 → 0.3710.000 → 0.139
All-Floor0.661 → 0.7090.854 → 0.8980.232 → 0.257
Uniform Random0.367 → 0.5300.453 → 0.6440.176 → 0.217
NN Propagation0.620 → 0.6930.779 → 0.8460.268 → 0.317
U-Net0.763 → 0.8230.870 → 0.9120.466 → 0.533
PConv-UNet0.758 → 0.8190.870 → 0.9120.455 → 0.525
LaMa0.623 → 0.7220.672 → 0.7790.271 → 0.361
LaMa-Ensemble0.661 → 0.7100.853 → 0.8980.230 → 0.255
Diffusion0.656 → 0.7450.734 → 0.8240.460 → 0.516
Flow Matching0.667 → 0.7440.771 → 0.8440.394 → 0.451
FM+XAttn0.659 → 0.7430.725 → 0.8170.412 → 0.472

Whole-map IoU is higher for every method because observed cells, taken from the geometry-derived input, match the reference. U-Net and PConv-UNet remain best. We keep hidden-region IoU as the main metric because it isolates completion errors.

We will add this comparison to the appendix and summarise it in the main results.

AuthorsOpenReview ↗

Question 1: Conditioning ratio and UMR

In Figure 6, I wonder why the UMR starts to increase after rcond > 0.5. Could you please explain?

Each bin holds different observations, so the rise does not mean that revealing more of a scene worsens completion. Above 0.5, the bins contain less hidden floor and far fewer observations, which raises the mismatch rate of any method that predicts floor where floor is common; All-Floor shows this directly.

All nine bins, from the 28,343 test observations (percentages):

Conditioning ratioObservationsHidden floorAll-Floor UMRU-Net UMRDiffusion UMRFM+XAttn UMR
[0.1,0.2)5,77869.9530.0512.0223.3023.82
[0.2,0.3)4,93078.5521.4510.7922.1823.12
[0.3,0.4)4,85983.3416.6610.2822.2023.08
[0.4,0.5)5,74686.4413.569.4622.2821.42
[0.5,0.6)4,64285.4614.5410.6821.3318.73
[0.6,0.7)1,75979.6820.3214.6221.5319.92
[0.7,0.8)49271.8328.1719.5020.5721.01
[0.8,0.9)12361.3838.6226.8525.1625.37
[0.9,1.0)1447.9352.0724.4117.0619.44

We will add the bin counts to the appendix and the Figure 6 caption.

Thank you again for these helpful suggestions.

Comments on the paper

13 comments
AuthorsOpenReview ↗

Thank you for the reviews

We thank all four reviewers for their careful reading and constructive feedback, and for recognising the value of the benchmark. Their questions prompted additional controls, reported in our replies to each review.

AuthorsOpenReview ↗

Weakness 1: Confounding factors in the planning comparison

Thank you for the detailed and constructive review, and for recognising the benchmark design and the value of MES. The controls you suggested separate the two sources of the planning improvement.

The planning comparison has some confounding factors. The sample-based planner differs from the U-Net planner in both admissible support and traversal costs. Therefore, it is not clear whether the improvement comes from multiple generative samples or from the overall planning design.

We added two controls. The first fixes endpoints, allowed cells, search, and cost formula, and changes only the per-cell score source (Question 2). The second fixes the graph and costs, and changes only the number of generated maps.

More controlled baselines, such as planning directly with U-Net probabilities, would strengthen this result.

On 4,000 queries, U-Net probabilities give 84.75% success, FM+XAttn frequencies 83.28%, and unit costs 81.20%. Using four FM+XAttn maps instead of one raises success by 1.47 percentage points (paired 95% interval 0.81 to 2.21). Both controls show that retaining ambiguity, as per-cell probabilities or as multiple maps, improves planning.

We will add both controls beside the planning results, with their protocols in the appendix.

AuthorsOpenReview ↗

Weakness 2: Training seeds and the influence of ZInD

Most learned models use only one training seed, while some differences between Diffusion, Flow Matching, and FM+XAttn are small.

We add two U-Net seeds and an update-matched second seed for plain Flow Matching (Question 1). U-Net ScanNet++ IoU spans 0.868 to 0.872, and the two Flow Matching runs differ by about 0.003 hard-OOD MES. Generators reach ID MES of 0.161 to 0.174 against U-Net's 0.237, while U-Net leads single-output IoU.

Also, ZInD provides nearly half of the observations and uses a different target-generation pipeline.

Excluding ZInD preserves all 11 MES ranks and U-Net's lead in IoU, mismatch rate, and F1 (Question 3).

Therefore, I think conclusions about small model differences should be made carefully.

The MES gap between generators and U-Net, and U-Net's lead in single-output IoU, hold across seeds (Question 1). We will add the seed comparisons to the appendix.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.