Pith. sign in

REVIEW 4 major objections 5 minor 49 references

This paper introduces Assistant Placement Aria, a benchmark for virtual placement that maps every plausible placement location in 250 indoor scenes across three tasks, arguing it is the first to combine global, local, and human-centric cons

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-08-04 01:14 UTC pith:CAUE2ERC

load-bearing objection Useful task framing and a solid manual subset, but the auto-labeled 80% of the data can't encode the global constraints the paper claims to measure, and the benchmark's headline numbers mostly reflect label granularity. the 4 major comments →

arxiv 2608.00652 v1 pith:CAUE2ERC submitted 2026-08-01 cs.RO

Assistant Placement Aria: A Benchmark for Egocentric Placement Assistance

Amir Belder , Gon\c{c}alo Dias Pais , Refael Vivanti , Omri Carmi , Daniel DeTone , Oren Shrout , Ido Gattegno , Ayellet Tal This is my paper
classification cs.RO
keywords virtual placementbenchmarkhuman-centric constraintsegocentric scenesobject placementscene understandingvision-language model2D/3D annotation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that virtual placement—deciding all the places where a person would naturally put an object or sit in a room—can be studied at scale with a reusable benchmark. It introduces Assistant Placement Aria, built on 225 synthetic and 25 real egocentric indoor scenes, with over 500,000 annotations for three tasks: placing a 2D panel, suggesting a sitting spot, and positioning a TV. The benchmark provides 2D binary placement maps, 3D point clouds, and per-object text descriptions, and it splits into train, validation, and test sets with both manual and automatically generated labels. The authors argue this is the first VP benchmark that combines global, local, and human-centric constraints, and they show that existing detection and segmentation models perform poorly on the fine-grained manual labels. A sympathetic reader would care because the task sits between scene understanding and robot assistance, and data scarcity has been the main obstacle.

Core claim

The central claim is that Assistant Placement Aria is the first benchmark to explore diverse aspects of virtual placement, covering global, local, and human-centric constraints. It contains 250 scenes (225 synthetic, 25 real) with over 500,000 annotations across three tasks—2D Panel Placement, Sitting Suggestion, and TV Placement—each with 2D images, 3D point clouds, and textual object descriptions. The labeling combines expert manual tagging (50 scenes) with an automatic vision-language-model-driven method (200 scenes) that asks whether a person would place the object at each instance mask; the automatic labels are coarser, covering entire objects rather than just the usable surfaces. Basel

What carries the argument

The load-bearing mechanism is the binary placement map: for each frame, each task gets a pixel-level mask marking placeable (1) and non-placeable (0) regions, which is then reprojected into 3D using camera poses and depth to create a placement point cloud. Manual maps come from three expert taggers with majority vote; automatic maps come from prompting a vision-language model with task-specific questions (e.g., 'Could a TV screen between 43 and 65 inches be placed here?') applied to each instance mask from the segmentation, with a post-hoc check that TV placements avoid the top and bottom 20% of wall height. This two-track labeling is what lets the benchmark scale to 200 automatically labele

Load-bearing premise

The load-bearing premise is that the vision-language model's yes/no answers mirror human placement preferences; the paper's own comparison shows those answers agree with human labels only about 60–70% of the time and mark whole objects instead of usable surfaces.

What would settle it

Take a random sample of the 200 automatically labeled scenes, re-annotate them with fresh human taggers who are blind to the vision-language-model output, and measure the IoU between the new human labels and the automatic labels. If agreement falls far below the reported 0.60–0.70, or if the new humans consistently reject surfaces the VLM marked as placeable (such as table legs), the automatic portion of the benchmark is not a reliable ground truth. A second check: count the proportion of placeable masks that lie on non-surface geometry like legs or backs; if that proportion is high, the autom

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the benchmark holds up, it supplies the first large-scale training and evaluation ground for virtual placement, letting researchers measure how well a model captures human placement preferences rather than just object co-occurrence.
  • The reported baseline numbers establish that current detection and segmentation models, which treat objects as whole entities, systematically over-predict placement regions on the manually annotated subset, pointing to the need for finer-grained, affordance-aware prediction.
  • The 3D placement point clouds provide a direct route from 2D predictions to robot-executable destination surfaces, so a robot could take the output and plan a placement or sitting assistance action.
  • The per-object text descriptions and global scene captions enable language-conditioned placement models, allowing a user to ask for a specific type of spot (e.g., 'a comfortable chair') and get a mask back.
  • Because the benchmark includes both synthetic and real egocentric scenes, it can test sim-to-real transfer for placement assistance, a key step for deployment on wearable or robot egocentric cameras.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An extension the authors do not pursue is refining the automatic labels with a surface or affordance segmentation step; the reported 0.60–0.70 IoU agreement with manual labels and the tendency to mark whole objects suggest that models trained on the 200 auto-labeled scenes inherit systematic label noise.
  • The vision-language-model oracle assumption can be tested directly: sample automatically labeled scenes, have fresh human raters re-annotate them without seeing the VLM output, and compare; this would reveal how much of the benchmark's 'human-centric' signal comes from genuine human agreement rather than VLM bias.
  • A broader extension is to move from static, object-centric scenes to dynamic ones where preferred placements depend on transient context such as other people, clutter, or reachability—conditions the current benchmark does not cover.
  • Because the benchmark provides 3D point clouds and camera poses, it could be plugged into downstream planning frameworks that need a cost map over surfaces; one testable consequence is whether a planner consuming these placement maps produces more socially acceptable assistive actions than one using object-detector outputs.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces Assistant Placement Aria, a benchmark for Virtual Placement (VP) — identifying all plausible target locations for placing objects or sitting, given egocentric scene context and human-centric constraints. The benchmark includes 250 indoor scenes (225 synthetic from ASE, 25 real from AEO) with three tasks: 2D Panel Placement, Sitting Suggestion, and TV Placement. Annotations are provided as 2D binary placement masks, 3D point clouds, and per-object text descriptions. Manual annotations were created for 50 scenes; for the remaining 200 synthetic scenes, an automatic labeling procedure uses LLaVA to classify each object mask as placeable or not, with a 6DoF filter for TV height. The paper evaluates several 2D and 3D detection/segmentation baselines, reporting IoU on manual and all-scene subsets.

Significance. If the benchmark is valid, it addresses a genuine gap: existing placement datasets either predict a single target or ignore human preferences, and no large-scale dataset covers diverse VP constraints (global, local, human-centric) in egocentric 2D and 3D. The multi-task design, the combination of synthetic and real data, and the attempt to scale annotation via a VLM are all potentially valuable contributions. The paper also includes baselines that are useful starting points. However, the central value of the benchmark depends on the trustworthiness of its ground-truth labels, and the automatic labeling protocol raises serious concerns about whether the 200 automatically labeled scenes measure the task as defined. The paper is transparent about the limitations, but those limitations are load-bearing for the benchmark's main claims.

major comments (4)
  1. [§III-A, Automatic labeling; Fig. 3(b); Table I] The automatic labeling protocol asks LLaVA a yes/no question on a per-object crop and then marks the entire object mask as placeable. This cannot capture the local surface-level constraints (e.g., tabletop vs. table legs) or the global scene constraints (line of sight, wall space, viewing height) that define the VP tasks in §I. The paper acknowledges this: auto labels include table legs and whole furniture items. Since these auto labels constitute 170 of 200 training scenes and 15 of 25 test scenes, the large gap between all-scene IoU and manual-subset IoU in Table I (e.g., SAM 2D-Panel: 70.7 vs 35.1) is more plausibly explained by label granularity than by VP competence. This undermines the benchmark's validity for fine-grained placement evaluation. The authors should either re-annotate test scenes manually, restrict the benchmark's claims to a 'weakly labeled' track, or provide evidenc
  2. [§III-A, Automatic labeling; §III-B, Dataset Statistics] The reported agreement between auto and manual labels is IoU 0.60–0.70 (Sitting: 0.60, 2D-Panel: 0.64, TV: 0.70). An IoU of 0.60 means substantial disagreement, not just minor boundary differences. For Sitting Suggestion, the paper notes that entire furniture items are often labeled instead of seats. Given that 80% of the benchmark and 60% of the test scenes rely on auto labels, the headline benchmark numbers are not a reliable measure of VP performance. At minimum, the paper must report results separately for manual-only scenes as the primary evaluation, and treat auto-labeled scenes as weakly supervised or auxiliary data.
  3. [§III-B, Dataset Statistics; Abstract] The claim of 'over 500,000 individual annotations' is not supported by the provided statistics. The paper states 250 scenes with an average of 1905 annotations per scene, which gives 476,250 annotations. The claim would need a clear definition of what counts as an annotation (e.g., per-task counts, 2D+3D double-counting) to be verifiable. Also, the average of 2.8 annotations per frame is difficult to reconcile with the reported range of frames per scene (93–1766); please clarify the unit of 'annotation' and provide a breakdown per task.
  4. [§IV, Table I] The exact composition of the 'Manually tagged scenes' column is unspecified. If this column includes any of the manually labeled scenes that were also used in training (the training set contains 15 real + 15 manual synthetic scenes), then the reported manual-subset IoU values are not a valid measure of generalization. The paper should state which scene IDs are in each evaluation column and ensure no overlap between training and evaluation sets.
minor comments (5)
  1. [Throughout] The name 'LLaV A' should be written consistently as 'LLaVA'.
  2. [Introduction] Typo: 'and asingle' should be 'and a single'.
  3. [§III-A] Missing spaces in several places: 'fewer than2%', 'top or bottom20%', '3D bounding box'. Please correct formatting.
  4. [Abstract and Introduction] The phrase 'the first benchmark to explore diverse aspects of VP' should be tempered or supported by a more explicit comparison to [14], since the contribution claim depends on what 'diverse' and 'aspects' mean.
  5. [Dataset release] A benchmark paper should provide a URL or availability statement for the dataset and annotation code. The current text does not mention a project page or download link.

Circularity Check

0 steps flagged

No significant circularity: this is a dataset/benchmark paper with independent manual validation, not a derivation that reduces to its inputs.

full rationale

The paper's contribution is a labeled benchmark, not a fitted model predicting its own labels. The automatic LLaVA labeling pipeline is a data-construction method; its outputs are validated against independent manual annotations (IoU 0.60–0.70, Section III-A) and the paper explicitly discloses the granularity gap (e.g., table legs included). The benchmark evaluations train standard detection/segmentation models on the labels and measure IoU on held-out scenes; this is ordinary supervised evaluation, not a self-prediction. The manual labels (50 scenes, majority vote of three human taggers) are independent ground truth and are reported separately in Table I. The benchmark's 'first VP benchmark' claim is a novelty claim, not a mathematical derivation. No self-citation chain is load-bearing: ASE and AEO are source datasets, not prior theoretical results; no uniqueness theorem or imported ansatz is used to force conclusions. The only mild concern—that LLaVA-generated labels are used as test ground truth for 15 of 25 test scenes—is a data-quality issue, not circularity, since the labels are outputs of an external VLM and are compared against human judgment rather than derived from the evaluation methods' inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The benchmark's central claims do not rest on derived equations, so the ledger records choices and assumptions behind the data. The main dependency is the use of LLaVA as an oracle for human preference on 200 of 250 scenes; this is an external model treated as ground truth. Task parameters (TV range, wall-height rule, panel size range) are author-defined constraints, not fitted values. No invented physical entities are introduced.

free parameters (4)
  • TV size range 43–65 inches = 43–65 in (author-defined)
    Task definition for TV Placement; no empirical justification beyond 'typical TV sizes' (§III-A).
  • TV height rule excluding top/bottom 20% of wall height = 20% threshold
    Used to filter LLaVA TV labels; chosen by the authors to encode ergonomics, not derived from data (§III-A).
  • 2D panel size range 8–40 cm = 8–40 cm (author-defined)
    Abstracts small man-made objects into one generic panel; bounds chosen to cover a cup to a large laptop (§III-A).
  • Point cloud resampling to 500,000 points per frame = 500k points
    Choice for memory and computational efficiency, without convergence analysis (Appendix VI).
axioms (4)
  • domain assumption LLaVA's answers to prompted questions align with human placement preferences
    Automatic ground truth for 200 scenes is defined by LLaVA yes/no answers (§III-A); the paper reports only partial validation via auto-vs-manual IoU 0.60–0.70.
  • domain assumption Binary pixel-level placeability is an adequate representation of virtual placement
    All tasks are labeled as binary maps (placeable=1/not=0), collapsing continuous preference or comfort into a hard label (§III-A).
  • domain assumption ASE and AEO scenes are representative of indoor human placement contexts
    The benchmark is built entirely on these two egocentric datasets; no coverage analysis across room types is given beyond dataset descriptions (§III).
  • domain assumption Majority vote of three human taggers yields correct annotations
    Manual labels come from three expert taggers with majority vote; inter-annotator agreement is only described as disagreements in fewer than 2% of annotations, without per-pixel or kappa statistics (§III-A).

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Assistant Placement Aria: A Benchmark for Egocentric Placement Assistance." pith.science (2026). https://pith.science/paper/CAUE2ERC

@misc{pith2026260800652,
  author       = {Pith},
  title        = {Pith review of: Assistant Placement Aria: A Benchmark for Egocentric Placement Assistance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CAUE2ERC}},
  note         = {Machine review of arXiv:2608.00652}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Human assistance in robotics spans around several tasks such as navigation, object manipulation, and placement, where a key challenge is selecting target destinations that align with human intentions or preferences. We focus on this challenge in the context of Virtual Placement (VP), the task of identifying all plausible target locations given scene context and human-centric constraints. This differs from traditional placement tasks that typically focus on a single, predefined target location. The VP problem is complex, as it requires both global and local reasoning about the scene's geometry, semantics, and plausibility. To address this gap, we introduce {\bf Assistant Placement Aria}, the first benchmark to explore diverse aspects of VP, including global, local, and human-centric constraints. It contains both synthetic and real indoor scenes annotated for three tasks: (i)~2D Panel Placement, (ii)~Sitting Suggestion, and (iii)~TV Placement. Each scene includes 2D images, a 3D point cloud, and a textual description of the objects within the scene. By contributing this benchmark, we aim to encourage further research in this underexplored and challenging field that is critically dependent on relevant data. We also evaluate several foundation models for object detection and segmentation on our benchmark.

Figures

Figures reproduced from arXiv: 2608.00652 by Amir Belder, Ayellet Tal, Daniel DeTone, Gon\c{c}alo Dias Pais, Ido Gattegno, Omri Carmi, Oren Shrout, Refael Vivanti.

Figure 1
Figure 1. Figure 1: Virtual placement. Examples from our benchmark combining all three placement tasks in a single visualization: (a) a synthetic scene and (b) a real one. The tasks include (1) 2D Panel Placement, (2) Sitting Suggestion, and (3) TV Placement. Both are annotated with binary placement maps (white = placeable, black = non-placeable), and a VLM generates per-frame text descriptions including object characteristic… view at source ↗
Figure 2
Figure 2. Figure 2: Ground truth example. Each frame was annotated separately for each task. (a) The original RGB image. (b) The only plausible place to place a 2D-Panel is on the desk at the bottom left corner of the image. (c) A person would sit on the couch. (d) The TV should be placed high enough over the wall and in front of the couch to enable comfortable watching. range of object sizes and spatial reasoning challenges,… view at source ↗
Figure 3
Figure 3. Figure 3: Annotation examples. Each task is shown with 4 manually annotated frames (a), followed by one automatically generated annotation (b), and one 3D point cloud annotation (c) reprojected from the annotated frames. The manual annotations highlight the high quality and precision of the human tagging. The automatic tagging is a bit less precise, as it covers entire objects. (a) Assistance placement questions (b)… view at source ↗
Figure 4
Figure 4. Figure 4: Example VP task prompts and scene description. The questions posed to LLaVA for the VP annotations and the text description of each frame. Text descriptions. Since textual descriptions are valuable for a range of scene understanding tasks, we also employ LLaVA to generate descriptions for each object in every frame. Leveraging these descriptions may improve VP performance in text-aware systems. For each ob… view at source ↗
Figure 5
Figure 5. Figure 5: Number of frames per scene. The number of frames within each scene. eligible. For TV Placement, 4.1% of the synthetic data and only 1% of the real data were eligible. This difference is not surprising, as the real data contains more images of walls, leaving less vacant space for televisions. Overall, the average number of annotations per scene is 1905, with each frame containing an average of 2.8 annotatio… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative results. Three examples of the placement performed by different 2D models. 3D virtual placement, leveraging the benchmark to advance human-centric scene understanding. VI. APPENDIX On the prompting process for the scene’s textual description. To generate the description of each object, we first identify a representative image patch that can serve as input for text generation. During preprocessi… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 4 linked inside Pith

  1. [1]

    Nyc-indoor-vpr: A long-term indoor visual place recognition dataset with semi-automatic annotation,

    D. Sheng, A. Yang, J.-R. Rizzo, and C. Feng, “Nyc-indoor-vpr: A long-term indoor visual place recognition dataset with semi-automatic annotation,” inIEEE Int’l Conf. Robotics and Automation (ICRA), 2024, pp. 14 853–14 859

    2024

  2. [2]

    Oriented surface reachability maps for robot placement,

    T. Birr, C. Pohl, and T. Asfour, “Oriented surface reachability maps for robot placement,” inIEEE Int’l Conf. Robotics and Automation (ICRA), 2022, pp. 3357–3363

    2022

  3. [3]

    Time-optimal tcp and robot base placement for pick-and-place tasks in highly constrained environments,

    A. Wachter, A. Kugi, and C. Hartl-Nesic, “Time-optimal tcp and robot base placement for pick-and-place tasks in highly constrained environments,” inIEEE/RSJ Int’l Conf. Intelligent Robots and Systems (IROS), 2024, pp. 2251–2257

    2024

  4. [4]

    Optimal object placement for minimum discontinuity non-revisiting coverage task,

    T. Yang, J. V . Miro, Y . Wang, and R. Xiong, “Optimal object placement for minimum discontinuity non-revisiting coverage task,” inIEEE Int’l Conf. Robotics and Automation (ICRA), 2021, pp. 8422–8428

    2021

  5. [5]

    Learning object placement via dual-path graph completion,

    S. Zhou, L. Liu, L. Niu, and L. Zhang, “Learning object placement via dual-path graph completion,” inEuropean Conf. Computer Vision (ECCV). Springer, 2022, pp. 373–389

    2022

  6. [6]

    Interactive object placement with reinforcement learning,

    Q. Meng and Q. Liu, “Interactive object placement with reinforcement learning,” 2023

    2023

  7. [7]

    Text2place: Affordance- aware text guided human placement,

    R. Parihar, H. Gupta, S. VS, and R. V . Babu, “Text2place: Affordance- aware text guided human placement,” inEuropean Conf. Computer Vision (ECCV). Springer, 2024, pp. 57–77

    2024

  8. [8]

    Topnet: Transformer-based object placement network for image compositing,

    S. Zhu, Z. Lin, S. Cohen, J. Kuen, Z. Zhang, and C. Chen, “Topnet: Transformer-based object placement network for image compositing,” inIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 1838–1847

    2023

  9. [9]

    Context- aware synthesis and placement of object instances,

    D. Lee, S. Liu, J. Gu, M.-Y . Liu, M.-H. Yang, and J. Kautz, “Context- aware synthesis and placement of object instances,” 2018

    2018

  10. [10]

    Scanrefer: 3d object localization in rgb-d scans using natural language,

    D. Z. Chen, A. X. Chang, and M. Nießner, “Scanrefer: 3d object localization in rgb-d scans using natural language,”16th European Conference on Computer Vision (ECCV), 2020

    2020

  11. [11]

    Jiang, A. Tagliasacchi, M. Pollefeys, and T. Funkhouser, “Openscene: 3d scene understanding with open vocabularies,

    S. Peng, K. Genova, C. “. Jiang, A. Tagliasacchi, M. Pollefeys, and T. Funkhouser, “Openscene: 3d scene understanding with open vocabularies,” inIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 815–824

    2023

  12. [12]

    Pla: Language- driven open-vocabulary 3d scene understanding,

    R. Ding, J. Yang, C. Xue, W. Zhang, S. Bai, and X. Qi, “Pla: Language- driven open-vocabulary 3d scene understanding,” inIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 7010–7019

    2023

  13. [13]

    Scanqa: 3d question answering for spatial scene understanding,

    D. Azuma, T. Miyanishi, S. Kurita, and M. Kawanabe, “Scanqa: 3d question answering for spatial scene understanding,” inIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 19 129–19 139

    2022

  14. [14]

    Seeing the unseen: Visual common sense for semantic placement,

    R. Ramrakhya, A. Kembhavi, D. Batra, Z. Kira, K.-H. Zeng, and L. Weihs, “Seeing the unseen: Visual common sense for semantic placement,” 2024

    2024

  15. [15]

    Project aria: A new tool for egocentric multi-modal ai research,

    J. Engel, K. Somasundaram, M. Goesele, A. Sun, A. Gamino, A. Turner, A. Talattof, A. Yuan, B. Souti, B. Meredith,et al., “Project aria: A new tool for egocentric multi-modal ai research,”arXiv preprint arXiv:2308.13561, 2023

    Pith/arXiv arXiv 2023

  16. [16]

    Efm3d: A benchmark for measuring progress towards 3d egocentric foundation models,

    J. Straub, D. DeTone, T. Shen, N. Yang, C. Sweeney, and R. Newcombe, “Efm3d: A benchmark for measuring progress towards 3d egocentric foundation models,” 2024

    2024

  17. [17]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” in Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  18. [18]

    Learning object place- ments for relational instructions by hallucinating scene representations,

    O. Mees, A. Emek, J. Vertens, and W. Burgard, “Learning object place- ments for relational instructions by hallucinating scene representations,” inIEEE Int’l Conf. Robotics and Automation (ICRA), Paris, France, 2020

    2020

  19. [19]

    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,” inIEEE/CVF Int’l Conf. Computer Vision (ICCV), 2023, pp. 4015– 4026

    2023

  20. [20]

    Anyloc: Towards universal visual place recognition,

    N. Keetha, A. Mishra, J. Karhade, K. M. Jatavallabhula, S. Scherer, M. Krishna, and S. Garg, “Anyloc: Towards universal visual place recognition,”IEEE Trans. Robot. Automat., vol. 9, no. 2, pp. 1286–1293, 2023

    2023

  21. [21]

    Siglip 2: Multilingual vision- language encoders with improved semantic understanding, localization, and dense features,

    M. Tschannen, A. Gritsenko, X. Wang, M. F. Naeem, I. Alabdulmohsin, N. Parthasarathy, T. Evans, L. Beyer, Y . Xia, B. Mustafa, O. Hénaff, J. Harmsen, A. Steiner, and X. Zhai, “Siglip 2: Multilingual vision- language encoders with improved semantic understanding, localization, and dense features,” 2025

    2025

  22. [22]

    Second: Sparsely embedded convolutional detection,

    Y . Yan, Y . Mao, and B. Li, “Second: Sparsely embedded convolutional detection,”Sensors, vol. 18, no. 10, p. 3337, 2018

    2018

  23. [23]

    Pointpillars: Fast encoders for object detection from point clouds,

    A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” in IEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR). Computer Vision Foundation / IEEE, June 2019, pp. 12 697–12 705

    2019

  24. [24]

    Lps-net: Lightweight parameter-shared network for point cloud-based place recognition,

    C. Liu, G. Chen, and R. Song, “Lps-net: Lightweight parameter-shared network for point cloud-based place recognition,” inIEEE Int’l Conf. Robotics and Automation (ICRA), 2024, pp. 448–454

    2024

  25. [25]

    Pointrcnn: 3d object proposal generation and detection from point cloud,

    S. Shi, X. Wang, and H. Li, “Pointrcnn: 3d object proposal generation and detection from point cloud,” inIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2019, pp. 770–779

    2019

  26. [26]

    Visual place recognition: A survey,

    S. Lowry, N. Sünderhauf, P. Newman, J. J. Leonard, D. Cox, P. Corke, and M. J. Milford, “Visual place recognition: A survey,”IEEE Trans. Robot. Automat., vol. 32, no. 1, pp. 1–19, 2016

    2016

  27. [27]

    Learning to place new objects in a scene,

    Y . Jiang, M. Lim, C. Zheng, and A. Saxena, “Learning to place new objects in a scene,”The International Journal of Robotics Research (IJRR), vol. 31, no. 9, pp. 1021–1043, 2012

    2012

  28. [28]

    Optimization beyond the convolution: Generalizing spatial relations with end-to-end metric learning,

    P. Jund, A. Eitel, N. Abdo, and W. Burgard, “Optimization beyond the convolution: Generalizing spatial relations with end-to-end metric learning,” inIEEE Int’l Conf. Robotics and Automation (ICRA). IEEE, 2018, pp. 1–7

    2018

  29. [29]

    Cut, paste and learn: Surprisingly easy synthesis for instance detection,

    D. Dwibedi, I. Misra, and M. Hebert, “Cut, paste and learn: Surprisingly easy synthesis for instance detection,” inIEEE Int’l Conf. Computer Vision (ICCV), 2017, pp. 1301–1310

    2017

  30. [30]

    Learning object arrangements in 3d scenes using human context,

    Y . Jiang, M. Lim, and A. Saxena, “Learning object arrangements in 3d scenes using human context,”arXiv preprint arXiv:1206.6462, 2012

    Pith/arXiv arXiv 2012

  31. [31]

    Learning the spatial semantics of manipulation actions through preposition grounding,

    K. Zampogiannis, Y . Yang, C. Fermuller, and Y . Aloimonos, “Learning the spatial semantics of manipulation actions through preposition grounding,” inIEEE Int’l Conf. Robotics and Automation (ICRA). IEEE, 2015, pp. 1389–1396

    2015

  32. [32]

    OPA: object placement assessment dataset,

    L. Liu, B. Zhang, J. Li, L. Niu, Q. Liu, and L. Zhang, “OPA: object placement assessment dataset,”CoRR, vol. abs/2107.01889, 2021

    Pith/arXiv arXiv 2021

  33. [33]

    Making images real again: A comprehensive survey on deep image composition,

    L. Niu, W. Cong, L. Liu, Y . Hong, B. Zhang, J. Liang, and L. Zhang, “Making images real again: A comprehensive survey on deep image composition,”arXiv preprint arXiv:2106.14490, 2021

    arXiv 2021

  34. [34]

    Tf-icon: Diffusion-based training- free cross-domain image composition,

    S. Lu, Y . Liu, and A. W.-K. Kong, “Tf-icon: Diffusion-based training- free cross-domain image composition,” inIEEE/CVF Int’l Conf. Computer Vision (ICCV), 2023, pp. 2294–2305

    2023

  35. [35]

    Context- aware synthesis and placement of object instances,

    D. Lee, S. Liu, J. Gu, M.-Y . Liu, M.-H. Yang, and J. Kautz, “Context- aware synthesis and placement of object instances,” inAdvances in Neural Information Processing Systems (NeurIPS), 2018, pp. 10 393– 10 403

    2018

  36. [36]

    Octopus: Open-vocabulary content tracking and object placement using semantic understanding in mixed reality,

    L. Yoffe, A. Sharma, and T. Höllerer, “Octopus: Open-vocabulary content tracking and object placement using semantic understanding in mixed reality,” 2023

    2023

  37. [37]

    Predart: Towards automatic oracle prediction of object placements in augmented reality testing,

    T. Rafi, X. Zhang, and X. Wang, “Predart: Towards automatic oracle prediction of object placements in augmented reality testing,” in Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering, ser. ASE ’22. New York, NY , USA: Association for Computing Machinery, 2023

    2023

  38. [38]

    Laion-5b: An open large-scale dataset for training next generation image-text models,

    C. Schuhmann, R. Beaumont, R. Vencu, C. Gordon, R. Wightman, M. Cherti, T. Coombes, A. Katta, C. Mullis, M. Wortsman,et al., “Laion-5b: An open large-scale dataset for training next generation image-text models,”Advances in Neural Information Processing Systems (NeurIPS), vol. 35, pp. 25 278–25 294, 2022

    2022

  39. [39]

    Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation,

    M. Khanna, Y . Mao, H. Jiang, S. Haresh, B. Shacklett, D. Batra, A. Clegg, E. Undersander, A. X. Chang, and M. Savva, “Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation,” inIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2024, pp. 16 384– 16 393

    2024

  40. [40]

    Sun rgb-d: A rgb-d scene understanding benchmark suite,

    S. Song, S. P. Lichtenberg, and J. Xiao, “Sun rgb-d: A rgb-d scene understanding benchmark suite,” inIEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2015, pp. 567–576

    2015

  41. [41]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3d reconstructions of indoor scenes,” inIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2017

    2017

  42. [42]

    Matterport3d: Learning from rgb-d data in indoor environments,

    A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niessner, M. Savva, S. Song, A. Zeng, and Y . Zhang, “Matterport3d: Learning from rgb-d data in indoor environments,”Int’l Conf. 3D Vision (3DV), 2017

    2017

  43. [43]

    The Replica dataset: A digital replica of indoor spaces,

    J. Straub, T. Whelan, L. Ma, Y . Chen, E. Wijmans, S. Green, J. J. Engel, R. Mur-Artal, C. Ren, S. Verma, A. Clarkson, M. Yan, B. Budge, Y . Yan, X. Pan, J. Yon, Y . Zou, K. Leon, N. Carter, J. Briales, T. Gillingham, E. Mueggler, L. Pesqueira, M. Savva, D. Batra, H. M. Strasdat, R. D. Nardi, M. Goesele, S. Lovegrove, and R. Newcombe, “The Replica dataset...

    Pith/arXiv arXiv 1906

  44. [44]

    Aria digital twin: A new benchmark dataset for egocentric 3d machine perception,

    X. Pan, N. Charron, Y . Yang, S. Peters, T. Whelan, C. Kong, O. Parkhi, R. Newcombe, and Y . C. Ren, “Aria digital twin: A new benchmark dataset for egocentric 3d machine perception,” inIEEE/CVF Int’l Conf. Computer Vision (ICCV), 2023, pp. 20 133–20 143

    2023

  45. [45]

    Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding,

    M. Roberts, J. Ramapuram, A. Ranjan, A. Kumar, M. A. Bautista, N. Paczan, R. Webb, and J. M. Susskind, “Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding,” inIEEE/CVF Int’l Conf. Computer Vision (ICCV), 2021, pp. 10 912–10 922

    2021

  46. [46]

    Openrooms: An open framework for photorealistic indoor scene datasets,

    Z. Li, T.-W. Yu, S. Sang, S. Wang, M. Song, Y . Liu, Y .-Y . Yeh, R. Zhu, N. Gundavarapu, J. Shi,et al., “Openrooms: An open framework for photorealistic indoor scene datasets,” inIEEE/CVF Conf. Computer Vision and Pattern Recognition (CVPR), 2021, pp. 7190–7199

    2021

  47. [47]

    Lidar-cs dataset: Lidar point cloud dataset with cross-sensors for 3d object detection,

    J. Fang, D. Zhou, J. Zhao, C. Wu, C. Tang, C.-Z. Xu, and L. Zhang, “Lidar-cs dataset: Lidar point cloud dataset with cross-sensors for 3d object detection,” inIEEE Int’l Conf. Robotics and Automation (ICRA), 2024

    2024

  48. [48]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space,

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Pointnet++: Deep hierarchical feature learning on point sets in a metric space,”Advances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017

    2017

  49. [49]

    Quickfps: Architecture and algorithm co-design for farthest point sampling in large-scale point clouds,

    M. Han, L. Wang, L. Xiao, H. Zhang, C. Zhang, X. Xu, and J. Zhu, “Quickfps: Architecture and algorithm co-design for farthest point sampling in large-scale point clouds,”IEEE Trans. Computer-Aided Design, 2023

    2023

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.