Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Pixels-to-Graph: Real-time Integration of Building Information Models and Scene Graphs for Semantic-Geometric Human-Robot Understanding

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a robot exploring an unknown environment can build a hierarchical scene graph—objects, scenes, rooms, buildings—in real time on CPU only, by moving structural understanding into the 2D image domain and bridging it…

desk verdict A credible integrated system with real-robot demos, but its quantitative segmentation claims rest on synthetic floorplans, so the transfer to real LiDAR maps is unproven. read the letter →

arxiv 2506.22593 v1 pith:LOHOXUXO submitted 2025-06-27 cs.RO cs.AIcs.CV

Antonello Longo , Chanyoung Chung , Matteo Palieri , Sung-Kyun Kim , Ali Agha , Cataldo Guaragnella , Shehryar Khattak This is my paper
classification cs.ROcs.AIcs.CV
keywords 3Dscenegraphsbuildinginformationmodelsstructuralsegmentationbird's-eyeviewmapsGANimageinpaintingCPU-onlyinferenceautonomousexplorationLiDAR-basedmapping
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

Pix2G is a pipeline for turning a robot's noisy LiDAR map and camera images into a structured scene graph during autonomous exploration. The paper's claim is that this can be done on CPU only, in real time, by reformulating structural segmentation as a 2D image problem: the LiDAR map is ortho-projected into a top-down bird's-eye view, denoised by a GAN, and instance-segmented into rooms, after which the 2D masks are projected back into 3D to segment the pointcloud and feed a four-level graph (object, scene, room, building). The result would give human operators a building-information-model-like 2D map synchronized with the robot's own 3D representation, closing a gap between human and robot situational awareness. The paper evaluates this on real autonomous runs with a legged robot in a cluttered garage and an office-like environment, and reports that adding the floorplan dataset and GAN denoising raises structural segmentation mIoU from 54.92% to 92.73% on its validation benchmark.

What carries the argument

The load-bearing mechanism is the two-stage structural pipeline. First, a Bird's-Eye View (BEV) generator scores each map column by a weighted combination of LiDAR point density and vertical height range, with an adaptive threshold updated online, so walls are recognized by being both dense and tall while ceiling-hanging objects are rejected. Second, a GAN-based image inpainting network fills gaps and removes clutter in the BEV image, then a Mask R-CNN instance segmentation model, trained on an augmented floorplan dataset, labels each room as a separate mask. The masks are back-projected through a pixel-to-voxel map to segment the 3D pointcloud and to place room nodes in the scene graph. Object and scene layers are added in parallel from camera detections and a scene classifier.

What would settle it

Take the pretrained GAN and Mask R-CNN, run them on LiDAR-derived BEV maps from a new indoor environment not seen in training, and compare the resulting room masks against hand-labeled ground truth of the actual rooms; if the mIoU is substantially lower than 92.73%, or if adjacent rooms are systematically merged or split, the central transfer claim is refuted.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the expensive part of 3D scene graph construction—separating the environment into independent rooms and structures—does not have to be done in 3D at all. It can be pulled down to 2D: project the LiDAR map into a top-down image, repair noise and missing walls with a generative inpainting network, detect rooms as instance masks with a 2D segmentation model, then use a pixel-to-voxel association to colorize the original 3D pointcloud. The paper further claims that this 2D-first formulation is what makes CPU-only real-time inference possible on a resource-constrained robot, and that it generalizes to both open garage-like spaces and narrow office-like corridors where prior room-segmentation methods that rely on topological or Manhattan-world assumptions struggle.

Load-bearing premise

The structural pipeline's accuracy rests on models trained and validated on synthetic floorplan images (CubiCasa5K) corrupted with salt-and-pepper noise, and the real-world experiments provide no quantitative ground truth; if the visual gap between those synthetic images and real LiDAR-derived bird's-eye views is large, the reported 92.73% mIoU will not hold on actual robot maps.

Editorial extensions

If this is right

  • A robot can maintain a semantic understanding of its surroundings onboard even when communication to a base station is lost, because the graph is built locally on CPU.
  • The system works in both open-plan areas and narrow corridor environments, two regimes where previous room-segmentation approaches tend to fail.
  • Training on a large floorplan dataset instead of only in-field samples is the main source of structural accuracy: mIoU jumps from 54.92% to 84.36% in the ablation.
  • Adding GAN-based denoising and noise-corrupted training samples closes most of the remaining gap, bringing mIoU to 92.73%.
  • The four-layer graph gives a human operator a 2D building-information-model view that stays linked to the robot's 3D map, so instructions and actions can be exchanged between the two representations.

Reading between the lines

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

  • A natural testable extension is to use the trained GAN as a generator of synthetic BEV training data from partial maps, which could further reduce reliance on hand-labeled real-world samples.
  • The reported 92.73% mIoU is measured on a validation set of corrupted synthetic floorplans, not on real robot maps; a reader should treat real-world accuracy as demonstrated qualitatively rather than quantified.
  • Because the BEV and pointcloud segmentation times grow linearly with map size, a sliding-window or incremental map update is the most direct route to bounding runtime on long missions.
  • If the 2D-first formulation generalizes, it points toward a design pattern for other resource-constrained perception tasks: solve the structural part in the cheapest faithful representation, then back-project to 3D.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper presents Pixels-to-Graph (Pix2G), a CPU-only framework for building hierarchical 3D scene graphs from camera images and LiDAR maps during autonomous exploration. The structural branch generates a BEV image from the LiDAR map using an adaptive height/density threshold, denoises it with a GAN trained on augmented CubiCasa5K data, and segments rooms with Mask R-CNN; the semantic branch uses MMDetection for objects and AlexNet/Places365 for scene classification. A pixel-to-voxel association projects 2D room masks into the 3D pointcloud, and the graph layers (object, scene, room, building) are rendered in RViz. Experiments on a NeBula-Spot robot in a garage and an office show qualitative segmentation and graph generation, while a three-study ablation on CubiCasa5K reports increasing mIoU/precision/recall up to 92.73% mIoU.

Significance. The significance is moderate. If the real-world structural segmentation accuracy matched the synthetic benchmark, Pix2G would be a useful systems contribution: it combines object detection, scene classification, adaptive BEV generation, GAN inpainting, and Mask R-CNN segmentation into a CPU-runnable pipeline that produces a 4-layer scene graph and colorized pointclouds, and it demonstrates this integration on a legged robot in two environments. The ablation study is a strength in that it shows the benefit of adding CubiCasa5K data and GAN-based denoising on the chosen benchmark. However, the quantitative evaluation is confined to a synthetic distribution, the real-world evaluation is qualitative, and the object/scene layers are not evaluated at all; these gaps currently prevent the paper from supporting the strong central claims of accurate real-world structural segmentation and reliable multi-layer graph construction.

major comments (5)
  1. [Section IV-B3, Table I] The only quantitative evaluation of the structural segmentation (Table I) is computed on a validation set of dirty CubiCasa5K samples, i.e., synthetic floorplans corrupted with salt-and-pepper noise, which matches the augmentation used for training Study C. This is a distribution-matched benchmark, not a test of transfer to real LiDAR-derived BEV maps, so the reported 92.73% mIoU does not by itself support the paper's claim of accurate structural segmentation in real unknown environments. The authors should add a held-out real-world evaluation with manually annotated ground-truth masks from the garage and office runs, reporting mIoU, precision, and recall per map.
  2. [Sections IV-B1 and IV-B2, Figs. 5 and 8] Real-world structural segmentation results are qualitative only: the garage experiment reports 'three large areas' and the office experiment reports a room count of 21, with no ground-truth masks, no intersection-over-union or precision/recall numbers, and no error bars. Because the room and building layers of the scene graph are constructed directly from these segmentation masks (Section III-C), the central claim of the system's practical utility in unknown environments is not quantitatively established. This is fixable by annotating the real maps and reporting per-room and per-building metrics.
  3. [Section III-C, Object layer] The object layer of the scene graph relies on a heuristic depth estimate based on bounding-box sizes ('no depth estimation module has been employed'), but the paper reports no evaluation of object detection accuracy or object localization error. Without such measurements, the object-level layer of the claimed multi-layer graph is unverified, and the graph cannot be stated to represent object positions reliably.
  4. [Section IV-B1, Figs. 6 and 7] The 'real-time' claim is not quantified: the text says the BEV and 3D segmentation latencies grow linearly with map size and 'still allows real-time operation,' but no numeric latencies, map sizes, or a target real-time rate are given. The authors should report concrete execution times, CPU/memory peaks, and the map size at which the pipeline meets (or violates) the real-time requirement.
  5. [Section IV-B3, Study C] The comparison between Study B and Study C changes two factors simultaneously: the addition of salt-and-pepper augmentation and the introduction of GAN preprocessing. Since both are intended to improve robustness, the reported improvement from 84.36% to 92.73% mIoU cannot be attributed to either component individually; an ablation with augmentation only and GAN only is needed to support the specific robustness claims.
minor comments (6)
  1. [Equation (1)] The first term ∑_{p∈B_i} p is not well-defined because p is a 3D point, not a scalar; it should be a count or an explicit density measure, and the normalization/units of the 0.4 and 0.6 weights are not specified.
  2. [Section II, Related work] There are several typos and citation issues, including 'to to reconstruct' and 'hierachical'; please proofread the manuscript carefully.
  3. [Reference [7]] Reference [7] appears to cite an unpublished 'Pix2pix gan for image-to-image translation' by Henry, Natalie, and Madsen rather than the original Pix2Pix paper (Isola et al., CVPR 2017); this should be corrected.
  4. [Equations (1) and (2)] The adaptive threshold parameters (weights 0.4/0.6 and EMA coefficient 0.85) are fixed without a sensitivity analysis; a short sensitivity study would strengthen the robustness claim.
  5. [Section IV-B2] The office experiment's room count (21) is reported as a success criterion, but without a ground-truth room count or an overlap measure this number alone does not indicate accuracy.
  6. [Abstract] The abstract states that the method is 'quantitatively and qualitatively evaluated during real-world experiments,' but the quantitative metrics in Table I are on synthetic CubiCasa5K samples; the wording should distinguish synthetic quantitative evaluation from qualitative real-world evaluation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the held-out CubiCasa5K evaluation is standard benchmark practice, and the central contribution does not reduce to its inputs.

full rationale

The derivation chain in Pix2G is self-contained against external benchmarks and does not exhibit any step where an output is defined in terms of the claimed prediction or where a fitted parameter is renamed as a prediction. The BEV generation uses an explicit adaptive threshold score (Eqs. 1-2); the GAN denoiser is trained on CubiCasa5K plus in-field samples corrupted with salt-and-pepper noise; the Mask R-CNN is trained on an augmented CubiCasa5K dataset and evaluated on a held-out validation set of dirty CubiCasa5K samples (Section IV-B3, Table I). This is a conventional train/validation split on an external dataset, and the 92.73% mIoU is a learned outcome, not a quantity forced by construction. Real-world NeBula-Spot experiments provide qualitative evidence; the lack of quantitative real-world ground truth is a validation/generalization weakness, not circularity. Self-citations to LOCUS 2.0, NeBula-Spot, ROSE, and the inspection planner are used as system components for SLAM, locomotion, state estimation, and planning; they do not carry the paper's core claims about structural segmentation or scene-graph generation. No load-bearing self-citation chain, uniqueness theorem, or ansatz-smuggling is present. Hence no circular step is identified.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. The central claim rests on several hand-tuned numerical coefficients and domain assumptions about LiDAR map structure and synthetic training data. The free parameters are chosen heuristically, and the domain assumptions are stated explicitly in the method sections. The most consequential assumption is that synthetic floorplan corruption approximates real LiDAR noise, which is not demonstrated.

free parameters (5)
  • Density weight alpha = 0.4
    Hand-chosen weight in Eq. (1) for the point-density term of the BEV score; no sensitivity analysis or theoretical justification is provided.
  • Height-difference weight beta = 0.6
    Hand-chosen weight in Eq. (1) for the height-difference term; complements alpha without stated basis.
  • EMA smoothing factor = 0.85
    Hand-chosen factor in Eq. (2) for the adaptive threshold; controls how fast the threshold adapts to new scores, with no sensitivity analysis.
  • GAN loss weights = 0.2 generator / 0.8 L1
    Hand-chosen balance in Eq. (3) between adversarial loss and L1 reconstruction loss; no ablation on this weighting is reported.
  • BEV padding = 12 px per side
    Fixed padding added to the BEV image to preserve boundaries; the effect of this choice on segmentation is not studied.
assumptions (4)
  • domain assumption Walls are the densest and tallest structures in a LiDAR map
    Justifies the adaptive BEV filter in Section III-B. In cluttered environments with dense ceiling objects, this may fail, though the height-difference term is meant to mitigate that.
  • ad hoc to paper CubiCasa5K floorplan images, augmented with salt-and-pepper noise, represent the noise characteristics of real LiDAR BEV images
    Used to train and validate the GAN denoiser and Mask R-CNN. No evidence links synthetic salt-and-pepper noise to the actual artifacts of LiDAR orthoprojections.
  • ad hoc to paper Heuristic depth from bounding box sizes can localize detected objects
    In Section III-C, object depth is estimated from bounding box sizes rather than learned or measured, which can be inaccurate and affects object-to-room linking.
  • domain assumption Adjacent rooms belong to the same building and open spaces separate buildings
    In Section III-C (Building layer), this rule is used to aggregate room-level information to building level; it may fail in multi-building complexes connected by corridors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pixels-to-Graph: Real-time Integration of Building Information Models and Scene Graphs for Semantic-Geometric Human-Robot Understanding." pith.science (2026). https://pith.science/paper/LOHOXUXO

@misc{pith2026250622593,
  author       = {Pith},
  title        = {Pith review of: Pixels-to-Graph: Real-time Integration of Building Information Models and Scene Graphs for Semantic-Geometric Human-Robot Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LOHOXUXO}},
  note         = {Machine review of arXiv:2506.22593}
}
read the original abstract

Autonomous robots are increasingly playing key roles as support platforms for human operators in high-risk, dangerous applications. To accomplish challenging tasks, an efficient human-robot cooperation and understanding is required. While typically robotic planning leverages 3D geometric information, human operators are accustomed to a high-level compact representation of the environment, like top-down 2D maps representing the Building Information Model (BIM). 3D scene graphs have emerged as a powerful tool to bridge the gap between human readable 2D BIM and the robot 3D maps. In this work, we introduce Pixels-to-Graph (Pix2G), a novel lightweight method to generate structured scene graphs from image pixels and LiDAR maps in real-time for the autonomous exploration of unknown environments on resource-constrained robot platforms. To satisfy onboard compute constraints, the framework is designed to perform all operation on CPU only. The method output are a de-noised 2D top-down environment map and a structure-segmented 3D pointcloud which are seamlessly connected using a multi-layer graph abstracting information from object-level up to the building-level. The proposed method is quantitatively and qualitatively evaluated during real-world experiments performed using the NASA JPL NeBula-Spot legged robot to autonomously explore and map cluttered garage and urban office like environments in real-time.

Figures

Figures reproduced from arXiv: 2506.22593 by the authors.

Figure 1
Figure 1. 3D scene graph with human-readable semantic BIM gener [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of the proposed approach, from input data to the creation of structurally segmented 2D BIM map and the 3D [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of using the proposed adaptive height threshold [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Architecture of the generator network of the proposed GAN: [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The pointcloud segmentation from 2D masks. The small [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: System performance results at runtime as executed on BS [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: CPU and Memory usage of the proposed system as executed [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Output graph after a testing run in urban office environment. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. BIM Informed Visual SLAM for Construction Environments

    cs.RO 2025-09 conditional novelty 5.0 of 10

    Adding BIM wall correspondences as fixed-node constraints in a visual SLAM back-end reduces average ATE by 23.71% and map RMSE by 7.14% on the authors' collected construction and office sequences.

Reference graph

Works this paper leans on

38 extracted references · 32 canonical work pages · cited by 1 Pith paper

  1. [1]

    Connecting semantic building information models and robotics: An application to 2d lidar-based localization,

    R. Hendrikx, P. Pauwels, E. Torta, H. P. Bruyninckx, and M. van de Molengraft, “Connecting semantic building information models and robotics: An application to 2d lidar-based localization,” in IEEE international conference on robotics and automation (ICRA) , 2021

    work page 2021

  2. [2]

    Automated facility inspection using robotics and bim: A knowledge-driven approach,

    J. Chen, W. Lu, Y . Fu, and Z. Dong, “Automated facility inspection using robotics and bim: A knowledge-driven approach,” Advanced Engineering Informatics, vol. 55, p. 101838, 2023

    work page 2023

  3. [3]

    Building multiversal semantic maps for mobile robot operation,

    J.-R. Ruiz-Sarmiento, C. Galindo, and J. Gonzalez-Jimenez, “Building multiversal semantic maps for mobile robot operation,” Knowledge- Based Systems, vol. 119, pp. 257–272, 2017

    work page 2017

  4. [4]

    Hydra: A real-time spatial perception system for 3d scene graph construction and optimization,

    N. Hughes, Y . Chang, and L. Carlone, “Hydra: A real-time spatial perception system for 3d scene graph construction and optimization,” arXiv preprint arXiv:2201.13360 , 2022

    arXiv 2022

  5. [5]

    3d dynamic scene graphs: Actionable spatial perception with places, objects, and humans,

    A. Rosinol, A. Gupta, M. Abate, J. Shi, and L. Carlone, “3d dynamic scene graphs: Actionable spatial perception with places, objects, and humans,” arXiv preprint arXiv:2002.06289 , 2020

    arXiv 2002

  6. [6]

    Hier- archical Open-V ocabulary 3D Scene Graphs for Language-Grounded Robot Navigation,

    A. Werby, C. Huang, M. B ¨uchner, A. Valada, and W. Burgard, “Hier- archical Open-V ocabulary 3D Scene Graphs for Language-Grounded Robot Navigation,” in Proceedings of Robotics: Science and Systems , Delft, Netherlands, July 2024

    work page 2024

  7. [7]

    Pix2pix gan for image-to-image translation,

    J. Henry, T. Natalie, and D. Madsen, “Pix2pix gan for image-to-image translation,” Research Gate Publication , pp. 1–5, 2021

    work page 2021

  8. [8]

    Image seg- mentation using morphological operations,

    D. Chudasama, T. Patel, S. Joshi, and G. I. Prajapati, “Image seg- mentation using morphological operations,” International Journal of Computer Applications, vol. 117, no. 18, 2015

    work page 2015

Show all 38 references
  1. [9]

    Image segmentation using deep learning: A survey,

    S. Minaee, Y . Boykov, F. Porikli, A. Plaza, N. Kehtarnavaz, and D. Terzopoulos, “Image segmentation using deep learning: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 44, no. 7, pp. 3523–3542, 2022

    2022

  2. [10]

    Deep learning techniques—r-cnn to mask r-cnn: a survey,

    P. Bharati and A. Pramanik, “Deep learning techniques—r-cnn to mask r-cnn: a survey,” Computational Intelligence in Pattern Recognition: Proceedings of CIPR 2019 , pp. 657–668, 2020

    2019

  3. [11]

    Faster r-cnn: Towards real- time object detection with region proposal networks,

    S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real- time object detection with region proposal networks,” Advances in neural information processing systems , vol. 28, 2015

    2015

  4. [12]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo et al. , “Segment anything,” arXiv:2304.02643, 2023

    2023 arXiv

  5. [13]

    Learning transferable visual models from natural language supervision,

    A. Radford et al. , “Learning transferable visual models from natural language supervision,” in International conference on machine learn- ing, 2021

    2021

  6. [14]

    Language-driven semantic segmentation,

    B. Li, K. Q. Weinberger, S. J. Belongie, V . Koltun, and R. Ranftl, “Language-driven semantic segmentation,” CoRR, 2022. [Online]. Available: https://arxiv.org/abs/2201.03546

    2022 arXiv

  7. [15]

    3d point cloud segmentation: A survey,

    A. Nguyen and B. Le, “3d point cloud segmentation: A survey,” in IEEE Conference on Robotics, Automation and Mechatronics , 2013

    2013

  8. [16]

    A unified point-based framework for 3d segmentation,

    H.-Y . Chiang, Y .-L. Lin, Y .-C. Liu, and W. H. Hsu, “A unified point-based framework for 3d segmentation,” in 2019 International Conference on 3D Vision (3DV) , 2019, pp. 155–163

    2019

  9. [17]

    Robust structure identification and room segmentation of cluttered indoor environments from occupancy grid maps,

    M. Luperto, T. P. Kucner, A. Tassi, M. Magnusson, and F. Amigoni, “Robust structure identification and room segmentation of cluttered indoor environments from occupancy grid maps,” IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 7974–7981, 2022

    2022

  10. [18]

    V oronoi random fields: Extract- ing topological structure of indoor environments via place labeling

    S. Friedman, H. Pasula, and D. Fox, “V oronoi random fields: Extract- ing topological structure of indoor environments via place labeling.” in IJCAI, vol. 7, 2007, pp. 2109–2114

    2007

  11. [19]

    The polygonal 3d layout reconstruction of an indoor environment via voxel-based room segmentation and space partition,

    F. Yang, Y . Li, M. Che, S. Wang, Y . Wang, J. Zhang, X. Cao, and C. Zhang, “The polygonal 3d layout reconstruction of an indoor environment via voxel-based room segmentation and space partition,” ISPRS International Journal of Geo-Information , 2022

    2022

  12. [20]

    Segmentation of structural elements from 3d point cloud using spatial dependencies for sustainability studies,

    J. Ntiyakunze and T. Inoue, “Segmentation of structural elements from 3d point cloud using spatial dependencies for sustainability studies,” Sensors, vol. 23, no. 4, p. 1924, 2023

    1924

  13. [21]

    Automatic room segmentation from unstructured 3-d data of indoor environments,

    R. Ambrus ¸, S. Claici, and A. Wendt, “Automatic room segmentation from unstructured 3-d data of indoor environments,” IEEE Robotics and Automation Letters , vol. 2, no. 2, pp. 749–756, 2017

    2017

  14. [22]

    Learning indoor layouts from simple point-clouds,

    M. T. Mahmood and M. E. Ali, “Learning indoor layouts from simple point-clouds,” arXiv preprint arXiv:2108.03378 , 2021

    2021 arXiv

  15. [23]

    House-gan++: Generative adversarial layout refine- ment network towards intelligent computational agent for professional architects,

    N. Nauata et al., “House-gan++: Generative adversarial layout refine- ment network towards intelligent computational agent for professional architects,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021

    2021

  16. [24]

    Wallplan: synthesizing floorplans by learning to generate wall graphs,

    J. Sun, W. Wu, L. Liu, W. Min, G. Zhang, and L. Zheng, “Wallplan: synthesizing floorplans by learning to generate wall graphs,” ACM Transactions on Graphics (TOG) , vol. 41, no. 4, pp. 1–14, 2022

    2022

  17. [25]

    Reconstruction and prediction of the layout of indoor environments from two-dimensional metric maps,

    M. Luperto and F. Amigoni, “Reconstruction and prediction of the layout of indoor environments from two-dimensional metric maps,” Engineering Applications of Artificial Intelligence , 2022

    2022

  18. [26]

    3-d scene graph: A sparse and semantic representation of physical environments for intelligent agents,

    U.-H. Kim, J.-M. Park, T.-J. Song, and J.-H. Kim, “3-d scene graph: A sparse and semantic representation of physical environments for intelligent agents,” IEEE transactions on cybernetics , 2019

    2019

  19. [27]

    Bridging 2d and 3d segmentation networks for computation-efficient volumetric medical image segmentation: An empirical study of 2.5 d solutions,

    Y . Zhang, Q. Liao, L. Ding, and J. Zhang, “Bridging 2d and 3d segmentation networks for computation-efficient volumetric medical image segmentation: An empirical study of 2.5 d solutions,” Comput- erized Medical Imaging and Graphics , vol. 99, p. 102088, 2022

    2022

  20. [28]

    Manhattan world: Compass direc- tion from a single image by bayesian inference,

    J. M. Coughlan and A. L. Yuille, “Manhattan world: Compass direc- tion from a single image by bayesian inference,” in IEEE international conference on computer vision , 1999

    1999

  21. [29]

    Mmdetection: Open mmlab detection toolbox and benchmark,

    K. Chen et al. , “Mmdetection: Open mmlab detection toolbox and benchmark,” ArXiv, vol. abs/1906.07155, 2019

    1906 arXiv

  22. [30]

    Imagenet classifica- tion with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classifica- tion with deep convolutional neural networks,” Advances in neural information processing systems , vol. 25, 2012

    2012

  23. [31]

    Places: A 10 million image database for scene recognition,

    B. Zhou, A. Lapedriza, A. Khosla, A. Oliva, and A. Torralba, “Places: A 10 million image database for scene recognition,” IEEE Transac- tions on Pattern Analysis and Machine Intelligence , 2017

    2017

  24. [32]

    Locus 2.0: Robust and computationally efficient lidar odometry for real-time 3d mapping,

    A. Reinke et al., “Locus 2.0: Robust and computationally efficient lidar odometry for real-time 3d mapping,” IEEE Robotics and Automation Letters, 2022

    2022

  25. [33]

    Semantic image inpainting with deep generative models,

    R. A. Yeh et al. , “Semantic image inpainting with deep generative models,” in IEEE conference on computer vision and pattern recog- nition, 2017

    2017

  26. [34]

    Cu- bicasa5k: A dataset and an improved multi-task model for floorplan image analysis,

    A. Kalervo, J. Ylioinas, M. H ¨aiki¨o, A. Karhu, and J. Kannala, “Cu- bicasa5k: A dataset and an improved multi-task model for floorplan image analysis,” in 21st Scandinavian Conference (SCIA) , 2019

    2019

  27. [35]

    Mask r-cnn,

    K. He, G. Gkioxari, P. Doll ´ar, and R. Girshick, “Mask r-cnn,” in IEEE international conference on computer vision , 2017

    2017

  28. [36]

    Autonomous spot: Long-range autonomous ex- ploration of extreme environments with legged locomotion,

    A. Bouman et al. , “Autonomous spot: Long-range autonomous ex- ploration of extreme environments with legged locomotion,” in IEEE International Conference on Intelligent Robots and Systems , 2020

    2020

  29. [37]

    Rose: Robust state estimation via online covari- ance adaption,

    S. Fakoorian et al., “Rose: Robust state estimation via online covari- ance adaption,” in The International Symposium of Robotics Research (ISRR), 2022

    2022

  30. [38]

    Efficient line-of-sight viewpoint sampling in complex environments for autonomous surface inspection,

    S. Moon et al., “Efficient line-of-sight viewpoint sampling in complex environments for autonomous surface inspection,” in IEEE Interna- tional Symposium on Safety Security Rescue Robotics (SSRR) , 2024

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.