Open-Vocabulary Indoor Object Grounding with 3D Hierarchical Scene Graph

Gleb Naumov; Sergey Linok

REVIEW 4 major objections 6 minor 1 cited by

Open-Vocabulary Indoor Object Grounding with 3D Hierarchical Scene Graph

T0 review · 4 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 Adding a 'location' level to a 3D hierarchical scene graph and letting an LLM reason over object-to-object links improves open-vocabulary object grounding for complex spatial queries in multi-floor indoor scenes.

desk verdict Reasonable system paper with a useful new location layer, but the key grounding result is only a delta against one baseline and not testable from the paper. read the letter →

arxiv 2507.12123 v1 pith:FLVMAFY4 submitted 2025-07-16 cs.CV

Sergey Linok , Gleb Naumov This is my paper

classification cs.CV

keywords hierarchicalscenegraphopen-vocabularygrounding3DobjectLLMreasoningRGB-Dspatialrelationslocationdetectionmulti-floor

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

OVIGo-3DHSG claims that open-vocabulary object grounding in large indoor scenes improves when the scene is organized as a hierarchy with a new 'location' level between rooms and objects, and when a large language model reasons over both the vertical links between levels and the horizontal links between specific objects. The paper's central assertion is that these two ideas let the system answer complex spatial queries — such as 'the vase closest to the window in the lounge area of the office on the first floor' — more accurately than prior scene-graph methods that lack object-to-object connections. To support this, the authors build the graph from posed RGB-D frames using open-vocabulary foundation models, train a detector to segment locations, and evaluate on multi-floor HM3DSem scenes, reporting an improvement over the closest baseline on a 50-question benchmark of queries that reference other objects. A sympathetic reader would care because the method targets a practical need: letting a robot or assistant find an object in a large building given language that points to it through context, not just through its name.

What carries the argument

The central machinery is the five-layer hierarchical scene graph $G = (N, E)$ with building, floor, room, location, and object nodes, together with the Deductive Hierarchical Reasoning Algorithm that navigates it. The new load-bearing piece is the location node: a contiguous 'meaningful area' inside a room, detected by a YOLOv11 model trained on the authors' DBSCAN-based heuristic labels, which reduces the search space before object matching. The reasoning algorithm then uses LLM calls to select only the objects relevant to the user's query, builds intra-layer edges between those targets and anchors, and grounds the answer by combining inter-layer and intra-layer context.

What would settle it

Have independent annotators, unfamiliar with the authors' labeling rules, draw location masks for the same eight HM3DSem scenes, then run the grounding benchmark with locations replaced by those human labels. If grounding accuracy does not drop when the detector's predictions conflict with the human-defined locations, the 'location' level is not the source of the reported gains.

Watch

Extended reading notes

Core claim

The central discovery claimed by the paper is that adding a 'location' abstraction inside rooms and using an LLM to exploit object-to-object relations produce a measurable gain in open-vocabulary 3D object grounding. In the authors' pipeline, floors are split using a height histogram, rooms by a wall-distance watershed on a bird's-eye view, locations by a trained detector that reproduces their heuristic clustering definition, and objects by RAM++ tagging, YOLO-World detection, and MobileSAMv2 segmentation aggregated across frames. For grounding, the Deductive Hierarchical Reasoning Algorithm first asks the LLM to pick relevant entities at each hierarchy level, then enriches only the query-related target and anchor objects with semantic edges (like 'left' and 'above' from ZSVG3D) and metric edges (like 'in 3 meters of'), and finally asks the LLM to return the object ID. On the authors' 50-question benchmark on eight HM3DSem scenes, the method reports an improvement over HOV-SG of 15.62 accuracy points at IoU 0.1 and 5.21 points at IoU 0.75.

Load-bearing premise

The load-bearing premise is that 'locations' are a real, useful category for grouping indoor objects; the ground-truth locations come from the authors' own clustering heuristic plus hand-designed filtering rules, so the hierarchy level may be an artifact of that labeling definition rather than a genuinely meaningful abstraction.

Editorial extensions

If this is right

A robot navigating a multi-floor building can narrow a free-form query level by level, from floor to room to location to object, instead of scanning all object candidates.
Object-to-object edges, built only for query-relevant anchors, enable grounding of relational queries like 'the pillow to the left of the sofa' while keeping the LLM context compact.
The location layer can be produced by a single binary instance-segmentation model, which the authors find to be more accurate and simpler than transformer-based floorplan prediction for this task.
Because the object layer relies on open-vocabulary tagging, detection, and segmentation, the resulting graph is not tied to a closed vocabulary of object classes.

Reading between the lines

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

The location concept may transfer to other indoor scene tasks — spatial question answering, layout summarization, or navigation planning — wherever an intermediate grouping between room and object helps prune search.
The reported 15.6-point improvement at IoU 0.1 comes from a benchmark whose queries were written to require anchors; a natural extension is to measure the same pipeline on generic queries and on single-floor or open-plan layouts, where the location heuristic may be harder to learn.
An ablation that removes only the location level while keeping object-to-object edges would settle whether the gain comes from the new abstraction or from the edges alone, since the paper reports gains over a baseline without object edges but not this specific ablation.

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.

Desk Editor's Note

read the letter

Here's my take on OVIGo-3DHSG. The core idea—adding a 'location' level to a hierarchical scene graph and running LLM-based deductive reasoning over inter-layer and intra-layer edges—is a sensible integration, and the paper explains the pipeline clearly. The explicit adaptation of their own BBQ algorithm is acknowledged, which I appreciate. The location layer itself is the main novelty, and it's a reasonable thing to try for multi-floor search.

But the central claim that the method outperforms existing baselines is not actually supported by the numbers as reported. Table 2 lists only positive deltas (15.62, 12.54, 9.34, 5.21) against HOV-SG, with no absolute Acc@K values, no confidence intervals, no per-scene breakdown, and no description of the HOV-SG configuration. With only 50 hand-written queries on 8 scenes, a delta of 5.21 at Acc@0.75 could be noise, and without absolute values you can't tell whether the methods are near floor or ceiling. The stress-test note is right: the headline result is currently unfalsifiable from the paper.

The object-level semantic comparison in Table 1 is more transparent, but there OVIGo-3DHSG's improvement over HOV-SG is 0.25 AUC points (85.13 vs 84.88), which is essentially a tie. So the strong wording in the abstract and conclusion overstates what the data shows.

Other issues: the ground-truth locations are generated by the same clustering heuristics that define what a location is, so the detector is trained to reproduce the authors' definition. That's a real circularity, though it's not fatal if the location abstraction proves useful for grounding. The paper gives many hyperparameters but leaves some undisclosed (e.g., alpha_min/max, c, C for labeling). The location detector comparison is shown only as F1 curves without error bars or a table of values.

On the positive side, the writing is clear, the method is reproducible from the description in principle, and the authors are upfront about the dataset limitations. The idea of a location layer is worth exploring, and the 50-query benchmark, small as it is, targets a genuine bottleneck in multi-floor grounding.

Who would get value from this? Robotics and AR researchers working on language-driven object search, especially those building scene-graph systems. It's a solid system description, not a breakthrough.

My recommendation: send it to review, but require the authors to report absolute Acc@K with error bars, a per-scene breakdown, a description of the HOV-SG baseline they used, and a release of code and benchmark annotations. Without those, the paper should not be accepted as is.

Referee Report

4 major / 6 minor

Summary. The paper presents OVIGo-3DHSG, a method that builds a five-layer hierarchical scene graph (building, floors, rooms, locations, objects) from posed RGB-D sequences using open-vocabulary detectors and an LLM. The key novelty is a 'location' abstraction between rooms and objects, obtained via a trained YOLOv11 detector on BEV projections, and a deductive LLM reasoning algorithm that selects query-relevant subgraphs and uses inter- and intra-layer edges to ground objects from natural-language queries. Experiments on 8 HM3DSem scenes report object-level semantic metrics and grounding improvements over HOV-SG.

Significance. The location abstraction and the query-driven construction of spatial edges are plausible and potentially useful contributions for large multi-floor indoor grounding. The paper is clearly structured and the pipeline is described in sufficient architectural detail to be reproduced in principle. However, the quantitative evidence is currently too thin: the main grounding result is reported only as a delta against a single baseline, the benchmark is small and self-created, and the location ground truth is operationally defined by the same heuristics that the paper introduces. If the missing evidence and ablations are supplied, the method could be a solid contribution to open-vocabulary scene understanding.

major comments (4)

[Sec. 4.2.3, Table 2] The central claim of outperforming existing methods in object grounding is supported only by the delta values 15.62, 12.54, 9.34, and 5.21 over HOV-SG. Absolute Acc@K numbers for either method are not reported, the HOV-SG configuration (LLM, prompt, edge construction) is not described, and there are no confidence intervals, per-scene breakdowns, or number of queries per scene. With a benchmark of 50 questions, a 5.21-point difference at Acc@0.75 is roughly 2.6 questions and may be within sampling noise. Please report absolute values with standard errors (e.g., bootstrap over the 50 queries and over repeated LLM samples) and specify the baseline setup in full.
[Sec. 4.1.1, Eqs. (4)-(7); Sec. 4.2.1] The ground-truth locations used for training and evaluation are generated by the same heuristic clustering and filtering rules that define what a location is, and the LocDetector is trained to reproduce that definition. The evaluation therefore measures internal consistency of the labeling heuristic rather than the usefulness of the location abstraction as an externally meaningful category. To support the claim that the location layer contributes to grounding, provide an ablation that replaces the location layer with room-level grouping alone, and ideally validate the location definition against human-annotated masks.
[Sec. 4.2.1] The grounding benchmark consists of 50 handwritten queries on 8 scenes, annotated within the same project, and the annotations and evaluation scripts are not released (the repository link in Sec. 1 is a bare URL). This makes the headline result unfalsifiable and unverifiable. Please release the benchmark annotations and the evaluation code, and include a per-scene results table so that readers can assess cross-scene variability.
[Sec. 3.2; Sec. 3.3, Eq. (4); Sec. 4.1.1, Eq. (7)] Several hyperparameters that are load-bearing for hierarchy construction are not reported: δ_wall in Sec. 3.2, α_min and α_max in Eq. (4), and c and C in Eq. (7). The floor-detection parameters are given, but these additional thresholds are needed to reproduce the room and location segmentation. Please provide the values used and a sensitivity analysis for at least the location clustering thresholds.

minor comments (6)

[Various] Section title 'Deductive Hierarchical Reasoning Alghorithm' contains a typo ('Alghorithm' should be 'Algorithm'); other typos include 'quires' (Abstract/Intro), 'labled' (Fig. 5 caption), 'abstaction' (Sec. 2.1), and 'existance' (Sec. 3).
[Sec. 4.2.2, Table 1] Table 1 mixes evaluation protocols: the baselines rank objects by CLIP similarity, while OVIGo-3DHSG uses direct label assignment. Please clarify the exact protocol and add standard deviations over runs or seeds.
[Sec. 4.1.4, Figs. 4 and 5] Figures 4 and 5 show F1 curves without numerical values or error bars; add a table of F1@δ at representative thresholds or provide the underlying data.
[Sec. 3.5.2] The LLM is used with chain-of-thought prompting, but the sampling temperature and number of runs per query are not reported; please specify these, as LLM outputs are stochastic.
[Sec. 4.1.2] Provide the batch size and other training hyperparameters for the YOLOv11 model, which are currently omitted.
[Sec. 1, Contributions] The claim of 'time-efficient' object layer construction is not supported by any runtime measurement; either add a runtime comparison or soften the claim.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the method is self-contained in its construction; the location level is an author-defined supervised target and the grounding benchmark is self-created, but no predicted quantity reduces to an input by construction.

full rationale

I find no circular step in the paper's derivation chain. The hierarchical scene graph is constructed from RGB-D data using external open-vocabulary models (RAM++, YOLO-World, MobileSAMv2), with floor, room, location, and object levels each specified algorithmically in Sec. 3. The location ground truth is generated by a DBSCAN-based heuristic with manual refinement (Sec. 4.1.1), and LocDetector is trained and evaluated against this author-defined target. This is supervised learning on an explicitly defined target, not a self-definitional prediction: the detector is tested on held-out HM3DSem scenes and could in principle fail to match the labeling heuristic. The object grounding benchmark (Sec. 4.2.1) consists of 50 author-written queries with ground-truth boxes, and Table 2 reports only deltas against HOV-SG; this is an evaluation-validity and reproducibility weakness rather than a circular reduction, because the reported improvement is an empirical measurement that is not forced by the benchmark definition alone. The deductive reasoning algorithm is attributed to the authors' own prior work BBQ (Ref. [25]) in Secs. 2.2 and 3.5, which is a self-citation; however, the adaptation is fully specified in the present paper and directly evaluated, so no load-bearing claim rests on an unverified citation. No equation equates a prediction to its input, and no fitted parameter is renamed as a prediction. Overall circularity is minimal or absent; the main concerns are benchmark self-authorship and missing absolute numbers, which belong to correctness risk rather than circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

All free parameters are hand-chosen or inherited from trained models. The central reliance on a half-heuristic, half-learned location definition is the largest conceptual debt, since no external validation of 'location' exists.

free parameters (6)

bin_h = 0.01 m
Height histogram bin size for floor segmentation (Sec. 3.1), chosen by hand.
delta_f = 0.2 m
Neighborhood size for peak detection in floor histogram (Sec. 3.1).
p_h = 0.9
Relative height threshold for retaining histogram peaks (Sec. 3.1).
delta_wall = not specified
Threshold for generating the binary wall map in room segmentation (Sec. 3.2). Value not given in the paper.
alpha_min, alpha_max = not specified
Height-range hyperparameters for the location-layer point cloud filtering (Sec. 3.3). Values not given.
c, C = not specified
Minimum cluster object count and isoperimetric ratio threshold for the location labeling pipeline (Sec. 4.1.1). Values not given.

assumptions (5)

ad hoc to paper The DBSCAN clustering plus hand-designed filtering rules produce a valid definition of 'location'.
Sec. 4.1.1 defines the location ground truth with this heuristic; no external validation is provided.
domain assumption Height histogram with a two-peaks-per-cluster rule reliably separates floors in multi-floor buildings.
Sec. 3.1 assumes this reflects the dataset creation process where free space exists between floors.
domain assumption BEV wall distance field plus watershed produces correct room partitions.
Sec. 3.2 uses this heuristic without validation on diverse layouts.
domain assumption RAM++ tags are sufficient prompts for Yolo-World to detect all query-relevant objects.
Sec. 3.4 relies on this chain of open-vocabulary models; no recall analysis is provided.
domain assumption The 50-question benchmark is representative of complex spatial object queries.
Sec. 4.2.1 introduces the self-created benchmark; no external standard or inter-annotator agreement is reported.

invented entities (1)

"location" node layer in the hierarchical scene graph
purpose: Sub-room grouping of objects to narrow the LLM search space and add an abstraction level for queries.
No external validation of the location concept; its utility is only demonstrated on the authors' own benchmark.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Open-Vocabulary Indoor Object Grounding with 3D Hierarchical Scene Graph." pith.science (2026). https://pith.science/paper/FLVMAFY4

@misc{pith2026250712123,
  author       = {Pith},
  title        = {Pith review of: Open-Vocabulary Indoor Object Grounding with 3D Hierarchical Scene Graph},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FLVMAFY4}},
  note         = {Machine review of arXiv:2507.12123}
}

read the original abstract

We propose OVIGo-3DHSG method - Open-Vocabulary Indoor Grounding of objects using 3D Hierarchical Scene Graph. OVIGo-3DHSG represents an extensive indoor environment over a Hierarchical Scene Graph derived from sequences of RGB-D frames utilizing a set of open-vocabulary foundation models and sensor data processing. The hierarchical representation explicitly models spatial relations across floors, rooms, locations, and objects. To effectively address complex queries involving spatial reference to other objects, we integrate the hierarchical scene graph with a Large Language Model for multistep reasoning. This integration leverages inter-layer (e.g., room-to-object) and intra-layer (e.g., object-to-object) connections, enhancing spatial contextual understanding. We investigate the semantic and geometry accuracy of hierarchical representation on Habitat Matterport 3D Semantic multi-floor scenes. Our approach demonstrates efficient scene comprehension and robust object grounding compared to existing methods. Overall OVIGo-3DHSG demonstrates strong potential for applications requiring spatial reasoning and understanding of indoor environments. Related materials can be found at https://github.com/linukc/OVIGo-3DHSG.

Figures

Figures reproduced from arXiv: 2507.12123 by the authors.

**Figure 2.** Pipeline of OVIGo-3DHSG: Hierarchical Scene Graph construction based on Sensory Input [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Qualitative example of YOLOv11 [39] LocDetector predictions on HM3DSem [16]. The green [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Validation quantitative results of location [PITH_FULL_IMAGE:figures/full_fig_p009_4.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.

QuASH: Using Natural-Language Heuristics to Query Visual-Language Robotic Maps
cs.RO 2025-10 conditional novelty 6.0 of 10

Querying VLM robot maps with an SVM trained on LLM-generated synonym/antonym embeddings outperforms cosine-threshold and single-antonym baselines on images and OpenSeg maps, but not consistently on LSeg maps.

Reference graph

Works this paper leans on

41 extracted references · 29 canonical work pages · cited by 1 Pith paper

[1]

STRL-Robotics: intelligent control for robotic platform in human-oriented environment

Konstantin V Mironov et al. “STRL-Robotics: intelligent control for robotic platform in human-oriented environment”. In:Artificial Intelligence and Decision Making2 (2023), pp. 45–63

work page 2023
[2]

Influence of neural network receptive field on monocular depth and ego-motion estimation

SA Linok and DA Yudin. “Influence of neural network receptive field on monocular depth and ego-motion estimation”. In:Optical Memory and Neural Networks32.Suppl 2 (2023), S206–S213

work page 2023
[3]

A comprehensive survey of visual slam algorithms

Andr´ ea Macario Barros et al. “A comprehensive survey of visual slam algorithms”. In: Robotics 11.1 (2022), p. 24

work page 2022
[4]

Nerf: Representing scenes as neural radiance fields for view syn- thesis

Ben Mildenhall et al. “Nerf: Representing scenes as neural radiance fields for view syn- thesis”. In:Communications of the ACM65.1 (2021), pp. 99–106

work page 2021
[5]

3d gaussian splatting for real-time radiance field rendering

Bernhard Kerbl et al. “3d gaussian splatting for real-time radiance field rendering.” In: ACM Trans. Graph.42.4 (2023), pp. 139–1

work page 2023
[6]

Comparative Analysis of YOLO Series (from V1 to V11) and Their Application in Computer Vision

Yong-Hwan Lee and Heung-Jun Kim. “Comparative Analysis of YOLO Series (from V1 to V11) and Their Application in Computer Vision”. In:Journal of the Semiconductor & Display Technology23.4 (2024), pp. 190–198

work page 2024
[7]

Image segmentation using deep learning: A survey

Shervin Minaee et al. “Image segmentation using deep learning: A survey”. In: IEEE transactions on pattern analysis and machine intelligence44.7 (2021), pp. 3523–3542

work page 2021
[8]

Segment everything everywhere all at once

Xueyan Zou et al. “Segment everything everywhere all at once”. In:Advances in neural information processing systems36 (2023), pp. 19769–19782

work page 2023

Show all 41 references

[9]

Groundingdino: Marryingdinowithgroundedpre-trainingforopen-set object detection

ShilongLiuetal. “Groundingdino: Marryingdinowithgroundedpre-trainingforopen-set object detection”. In:European Conference on Computer Vision. Springer. 2024, pp. 38– 55

2024
[10]

Yolo-world: Real-time open-vocabulary object detection

Tianheng Cheng et al. “Yolo-world: Real-time open-vocabulary object detection”. In:Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024, pp. 16901–16911

2024
[11]

Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Rendering

Cheng Sun et al. “Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Rendering”. In:arXiv preprint arXiv:2412.04459(2024)

2024 arXiv
[12]

3D-SPATIAL MULTIMODAL MEMORY

Xueyan Zou et al. “3D-SPATIAL MULTIMODAL MEMORY”. In:The Thirteenth Inter- national Conference on Learning Representations. 2025

2025
[13]

Rgb-only gaussian splatting slam for unbounded outdoor scenes

Sicheng Yu et al. “Rgb-only gaussian splatting slam for unbounded outdoor scenes”. In: arXiv preprint arXiv:2502.15633(2025)

2025
[14]

Multimodal 3D map reconstruction for intelligent robotcs usingneuralnetwork-basedmethods

Dmitrii Aleksandrovich Yudin. “Multimodal 3D map reconstruction for intelligent robotcs usingneuralnetwork-basedmethods”.In: Doklady Mathematics.Vol.110.Suppl1.Springer. 2024, S117–S125

2024
[15]

ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes

Chandan Yeshwanth et al. “ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes”. In: Proceedings of the International Conference on Computer Vision (ICCV). 2023

2023
[16]

Habitat-matterport 3d semantics dataset

Karmesh Yadav et al. “Habitat-matterport 3d semantics dataset”. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 4927– 4936. 11

2023
[17]

3dscenegraph:Astructureforunifiedsemantics,3dspace,andcamera

IroArmenietal.“3dscenegraph:Astructureforunifiedsemantics,3dspace,andcamera”. In: Proceedings of the IEEE/CVF international conference on computer vision. 2019, pp. 5664–5673

2019
[18]

3Ddynamicscenegraphs:Actionablespatialperceptionwithplaces, objects, and humans

AntoniRosinoletal.“3Ddynamicscenegraphs:Actionablespatialperceptionwithplaces, objects, and humans”. In:arXiv preprint arXiv:2002.06289(2020)

2020 arXiv
[19]

Hydra: A real-time spatial perception systemfor3Dscenegraphconstructionandoptimization

Nathan Hughes, Yun Chang, and Luca Carlone. “Hydra: A real-time spatial perception systemfor3Dscenegraphconstructionandoptimization”.In: arXiv preprint arXiv:2201.13360 (2022)

2022 arXiv
[20]

S-graphs+: Real-time localization and mapping leveraging hierar- chical representations

Hriday Bavle et al. “S-graphs+: Real-time localization and mapping leveraging hierar- chical representations”. In:IEEE Robotics and Automation Letters8.8 (2023), pp. 4927– 4934

2023
[21]

Foundations of spatial perception for robotics: Hierarchical rep- resentations and real-time systems

Nathan Hughes et al. “Foundations of spatial perception for robotics: Hierarchical rep- resentations and real-time systems”. In:The International Journal of Robotics Research 43.10 (2024), pp. 1457–1505

2024
[22]

Hierarchical open-vocabulary 3d scene graphs for language- grounded robot navigation

Abdelrhman Werby et al. “Hierarchical open-vocabulary 3d scene graphs for language- grounded robot navigation”. In:First Workshop on Vision-Language Models for Naviga- tion and Manipulation at ICRA 2024. 2024

2024
[23]

Sceneverse: Scaling 3d vision-language learning for grounded scene understanding

Baoxiong Jia et al. “Sceneverse: Scaling 3d vision-language learning for grounded scene understanding”. In:European Conference on Computer Vision. Springer. 2024, pp. 289– 310

2024
[24]

Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning

Qiao Gu et al. “Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning”. In:2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE. 2024, pp. 5021–5028

2024
[25]

Beyond Bare Queries: Open-Vocabulary Object Grounding with 3D Scene Graph

Sergey Linok et al. “Beyond Bare Queries: Open-Vocabulary Object Grounding with 3D Scene Graph”. In:arXiv preprint arXiv:2406.07113(2024)

2024 arXiv
[26]

Visual programming for zero-shot open-vocabulary 3d visual ground- ing

Zhihao Yuan et al. “Visual programming for zero-shot open-vocabulary 3d visual ground- ing”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024, pp. 20623–20633

2024
[27]

3DGraphLLM: Combining Semantic Graphs and LargeLanguageModelsfor3DSceneUnderstanding

Tatiana Zemskova and Dmitry Yudin. “3DGraphLLM: Combining Semantic Graphs and LargeLanguageModelsfor3DSceneUnderstanding”.In: arXiv preprint arXiv:2412.18450 (2024)

2024 arXiv
[28]

OpenObject-NAV:Open-VocabularyObject-OrientedNavigationBased on Dynamic Carrier-Relationship Scene Graph

YujieTangetal.“OpenObject-NAV:Open-VocabularyObject-OrientedNavigationBased on Dynamic Carrier-Relationship Scene Graph”. In: arXiv preprint arXiv:2409.18743 (2024)

2024 arXiv
[29]

Clio: Real-time task-driven open-set 3d scene graphs

Dominic Maggio et al. “Clio: Real-time task-driven open-set 3d scene graphs”. In:IEEE Robotics and Automation Letters(2024)

2024
[30]

A density-based algorithm for discovering clusters in large spa- tial databases with noise

Martin Ester et al. “A density-based algorithm for discovering clusters in large spa- tial databases with noise”. In: Proceedings of the Second International Conference on Knowledge Discovery and Data Mining. KDD’96. Portland, Oregon: AAAI Press, 1996, 226тАУ231

1996
[31]

A Threshold Selection Method from Gray-Level Histograms

Nobuyuki Otsu. “A Threshold Selection Method from Gray-Level Histograms”. In:IEEE Transactions on Systems, Man, and Cybernetics9.1 (1979), pp. 62–66.doi: 10.1109/ TSMC.1979.4310076. 12

1979
[32]

An overview of watershed algorithm implementa- tions in open source libraries

Anton S Kornilov and Ilia V Safonov. “An overview of watershed algorithm implementa- tions in open source libraries”. In:Journal of Imaging4.10 (2018), p. 123

2018
[33]

Recognize Anything: A Strong Image Tagging Model

Youcai Zhang et al. “Recognize Anything: A Strong Image Tagging Model”. In:arXiv preprint arXiv:2306.03514 (2023)

2023 arXiv
[34]

Mobilesamv2: Faster segment anything to everything

Chaoning Zhang et al. “Mobilesamv2: Faster segment anything to everything”. In:arXiv preprint arXiv:2312.09579 (2023)

2023 arXiv
[35]

Chain-of-thought prompting elicits reasoning in large language models

Jason Wei et al. “Chain-of-thought prompting elicits reasoning in large language models”. In: Advances in neural information processing systems35 (2022), pp. 24824–24837

2022
[36]

Connecting the Dots: Floorplan Reconstruction Using Two-Level Queries

Yuanwen Yue et al. “Connecting the Dots: Floorplan Reconstruction Using Two-Level Queries”.In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2023

2023
[37]

Yolov11:Anoverviewofthekeyarchitectural enhancements

RahimaKhanamandMuhammadHussain.“Yolov11:Anoverviewofthekeyarchitectural enhancements”. In:arXiv preprint arXiv:2410.17725(2024)

2024 arXiv
[38]

Structured3D: A Large Photo-realistic Dataset for Structured 3D Mod- eling

Jia Zheng et al. “Structured3D: A Large Photo-realistic Dataset for Structured 3D Mod- eling”. In:Proceedings of The European Conference on Computer Vision (ECCV). 2020

2020
[39]

Ultralytics YOLO11

Glenn Jocher and Jing Qiu. Ultralytics YOLO11. Version 11.0.0. 2024. url: https : //github.com/ultralytics/ultralytics

2024
[40]

Visual language maps for robot navigation

Chenguang Huang et al. “Visual language maps for robot navigation”. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE. 2023, pp. 10608– 10615

2023
[41]

Gpt-4o system card

Aaron Hurst et al. “Gpt-4o system card”. In:arXiv preprint arXiv:2410.21276(2024). 13

2024 arXiv

Pith tools

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

[1] [1]

STRL-Robotics: intelligent control for robotic platform in human-oriented environment

Konstantin V Mironov et al. “STRL-Robotics: intelligent control for robotic platform in human-oriented environment”. In:Artificial Intelligence and Decision Making2 (2023), pp. 45–63

work page 2023

[2] [2]

Influence of neural network receptive field on monocular depth and ego-motion estimation

SA Linok and DA Yudin. “Influence of neural network receptive field on monocular depth and ego-motion estimation”. In:Optical Memory and Neural Networks32.Suppl 2 (2023), S206–S213

work page 2023

[3] [3]

A comprehensive survey of visual slam algorithms

Andr´ ea Macario Barros et al. “A comprehensive survey of visual slam algorithms”. In: Robotics 11.1 (2022), p. 24

work page 2022

[4] [4]

Nerf: Representing scenes as neural radiance fields for view syn- thesis

Ben Mildenhall et al. “Nerf: Representing scenes as neural radiance fields for view syn- thesis”. In:Communications of the ACM65.1 (2021), pp. 99–106

work page 2021

[5] [5]

3d gaussian splatting for real-time radiance field rendering

Bernhard Kerbl et al. “3d gaussian splatting for real-time radiance field rendering.” In: ACM Trans. Graph.42.4 (2023), pp. 139–1

work page 2023

[6] [6]

Comparative Analysis of YOLO Series (from V1 to V11) and Their Application in Computer Vision

Yong-Hwan Lee and Heung-Jun Kim. “Comparative Analysis of YOLO Series (from V1 to V11) and Their Application in Computer Vision”. In:Journal of the Semiconductor & Display Technology23.4 (2024), pp. 190–198

work page 2024

[7] [7]

Image segmentation using deep learning: A survey

Shervin Minaee et al. “Image segmentation using deep learning: A survey”. In: IEEE transactions on pattern analysis and machine intelligence44.7 (2021), pp. 3523–3542

work page 2021

[8] [8]

Segment everything everywhere all at once

Xueyan Zou et al. “Segment everything everywhere all at once”. In:Advances in neural information processing systems36 (2023), pp. 19769–19782

work page 2023