Pith. sign in

REVIEW 4 major objections 6 minor 30 references

QuASH: Using Natural-Language Heuristics to Query Visual-Language Robotic Maps

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that the region of embedding space matching a query is better estimated by a classifier trained on text synonyms and antonyms than by a cosine threshold or a single complement.

desk verdict Useful within-subfield method, but the abstract overclaims: the actual LLM variant fails on LSeg maps, and the OpenSeg gain is a precision-recall tradeoff shift that needs PR-curve evidence before it is declared a win. read the letter →

arxiv 2510.14546 v2 pith:FE44FS7Z submitted 2025-10-16 cs.RO

Matti Pekkanen , Francesco Verdoja , Ville Kyrki This is my paper
classification cs.RO
keywords latent-semanticmapsvision-languagemodelsopen-vocabularyqueryingembeddingclassificationsemanticsynonymsandantonymsroboticmappingsupportvectormachinequeryability
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

The paper argues that querying a vision-language map by comparing the query embedding to each map embedding with a cosine threshold is fundamentally limited: the region of embedding space corresponding to a query is neither isotropic nor connected. It proposes QuASH, which asks a language model for a set of semantic synonyms and antonyms of the query, embeds these text strings, and trains a support-vector machine to separate them. The classifier's decision boundary then serves as the estimate of the query region, and is applied directly to visual embeddings in maps or images. On image benchmarks this raises F1 by 9 to 15 percentage points over the single-complement baseline; on map voxels it raises F1 from 0.136 to 0.445 for one of the two tested encoders when ground-truth synonym sets are used, while language-model-generated synonyms help that encoder but hurt the other. A sympathetic reader should care because the method avoids threshold tuning and neural-network training, is encoder-agnostic, and opens the door to using encoders whose embeddings are otherwise hard to query.

What carries the argument

The load-bearing component is the semantic-augmentation heuristic h(q) = (S_q, A_q), which produces a set of semantically close strings and a set of antonyms, followed by an off-the-shelf SVM with a non-linear kernel trained on the text embeddings of these sets. The SVM's decision boundary is the estimate of R_Q. Unlike cosine-threshold cones or the hyperplane from a single complement, this boundary is anisotropic and can be non-connected, matching the expected shape of a query region whose dimensions carry unequal semantic weight. The method's other key component is the fixed set of generic negative strings added to every query, which anchors the boundary in regions of the space that are un

What would settle it

Measure, per query and per encoder, the gap between the centroid of the text-embedding samples (synonyms and antonyms) and the centroid of the visual embeddings of true matches in the map. If that gap is large while QuASH still outperforms the baseline, the transfer assumption is not the driver; if the gap is small and QuASH fails, the assumption is insufficient. A second falsifier: hand-craft a synonym set that is semantically close in language but visually far (for example, words describing the same category but with very different appearance), and show that QuASH's boundary degrades — confi

Watch

Extended reading notes

Core claim

The paper's central claim is that the set of embeddings matching a natural-language query — the region R_Q in the shared latent space — can be approximated by training a classifier on text-only samples. The query and its semantic synonyms serve as positive examples; antonyms and generic negatives serve as negative examples. Because the text and visual encoders map into a common space, the decision boundary learned from text transfers to visual embeddings, partitioning map cells or image patches into matches and non-matches. The paper shows this transfer works on image benchmarks and on one of the two tested map encoders, with F1 rising from 0.136 to 0.445 in map querying when good synonym se

Load-bearing premise

A classifier trained on text embeddings of synonyms and antonyms transfers to visual embeddings of the same concepts, because the text and visual encoders share a common embedding space — an assumption the paper itself notes is strained by the modality gap, and which is consistent with the method failing on one map encoder when language-model synonyms are used.

Editorial extensions

If this is right

  • Any existing map-querying pipeline that thresholds cosine similarity can swap in a text-trained classifier without retraining a network, and gain accuracy on encoders whose embeddings are hard to threshold.
  • The quality of the synonym/antonym heuristic becomes the main lever on query performance; improving the heuristic improves visual classification directly.
  • Encoders that are poor choices under threshold-based querying become usable for latent-semantic maps, broadening the set of backbones for robotic semantic mapping.
  • Prompt-engineering templates that work for one encoder do not generalize; the distribution of samples matters more than the prompt phrasing.
  • The formalization of querying as estimating the latent region R_Q unifies thresholding and single-complement methods as special cases of a classifier-based approach.

Reading between the lines

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

  • Inference: the same text-trained classifier could replace thresholding in downstream tasks such as object retrieval or grounding, where the identical bottleneck appears.
  • Inference: the failure on one encoder with language-model synonyms suggests a measurable precondition — overlap between text and visual embedding distributions — that could be checked before applying the method.
  • Inference: adaptive negative sampling per query, instead of the fixed generic set, may sharpen the boundary further, since the paper shows the choice of antonyms changes F1 substantially.
  • Inference: if the modality gap is the binding constraint, then aligning or fine-tuning the encoders on map data would likely amplify QuASH's gains.
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

4 major / 6 minor

Summary. The paper formalizes querying of vision-language-model (VLM) embedded maps as a binary classification problem in a shared latent space, generalizing existing thresholding and single-complement baselines. It proposes QuASH: for a text query q, use an LLM (GPT-4o) to generate semantic synonyms S_q and antonyms A_q, embed these text strings with the language encoder, train an off-the-shelf SVM classifier on those embeddings, and apply the classifier to visual embeddings of map voxels or image pixels. Experiments on COCO, Pascal Context 459, and Matterport images (with LSeg) and on Matterport3D maps (with LSeg and OpenSeg) compare against the VLMaps-style baseline that contrasts the query with the word "other." The image benchmarks show consistent F1 improvements. On maps, the method improves over baseline with OpenSeg, especially when "ground truth" synonyms sampled from the Matterport class hierarchy are used, but with LLM-generated synonyms it is worse than baseline on LSeg maps (F1 0.396 vs. 0.561). Ablations examine prompt engineering, choice of antonym, synonym-generation strategy, and voxel aggregation.

Significance. The formalization in Section IV-A is useful and provides a clean unifying view of common querying schemes. The central idea—using a set of text-side synonym/antonym embeddings to train a classifier instead of relying on one query point and a threshold—is timely and potentially valuable. The image-benchmark results are credible and consistent, and the paper provides a public code link and extensive ablations. If the OpenSeg map result survives a threshold-independent and significance-aware evaluation, the method could be practically important, since the baseline is very poor for OpenSeg. However, the headline claim of increased queryability is only partially supported: the practical LLM heuristic fails on LSeg maps, the OpenSeg improvement may arise from selecting a more conservative operating point, and the "ground truth" rows are an oracle based on the evaluation dataset's own class hierarchy rather than a deployable heuristic.

major comments (4)
  1. [Abstract, §VI.B, Table II] The abstract and introduction claim that the method "demonstrates increased queryability" and "outperforms" the single-complement baseline, but Table II shows the proposed LLM-based heuristic underperforms baseline on LSeg maps (F1 0.396 vs. 0.561), while the ground-truth variant provides only a marginal, unreported-significance gain (0.574 vs. 0.561). The only map setting with substantial LLM improvement is OpenSeg, where the baseline itself is extremely weak (F1 0.136, precision 0.073, recall 0.926). The "ground truth" rows use 20 subclasses and 100 random negative subclasses from the Matterport3D label hierarchy—information not available in an open-vocabulary query. The central claim should be reframed to reflect that the method is beneficial on image benchmarks and on OpenSeg maps, but not on LSeg maps with the actual LLM heuristic.
  2. [§VI.B, Table II] No error bars, confidence intervals, or significance tests are reported across the 10 maps. More importantly, the OpenSeg comparison may reflect a precision/recall operating-point shift rather than better region estimation: baseline has precision 0.073/recall 0.926, QuASH-ground-truth has precision 0.622/recall 0.369. The baseline has no tunable threshold while the SVM decision boundary can implement arbitrary conservative cuts. Without precision-recall curves, PR-AUC, or a paired per-map comparison, the claim that QuASH "outperforms" the baseline is not established as more than moving along the same scoring trade-off. The paper itself notes in §VI.B that "the precision-recall curve of our method could be further optimized," reinforcing this concern. Please report threshold-independent metrics or per-map paired statistics.
  3. [§V-C] The classifier selection is not described with enough detail to rule out selection on the test data. The text says "we evaluated all of the applicable off-the-shelf classifiers from the scikit-learn library, and decided to use SVM with Euclidean and cosine similarity metrics as it had the best performance." There is no mention of a validation split or nested cross-validation. If the best classifier family was chosen using the same image/map benchmarks whose F1 numbers are later reported, the results are optimistically biased. This is load-bearing because the method is presented as using an off-the-shelf classifier; please specify the selection protocol and, if possible, report performance of the runner-up classifiers on held-out data.
  4. [§IV-C, Figure 3, §VI.B] The method's key assumption is that a decision boundary learned from language embeddings φ(S_q) and φ(A_q) transfers to visual embeddings ρ(v) because φ and ρ share a latent space. The paper itself cites the modality gap (§II, [11]) and later observes that maps and images behave differently (§VI.B). The LSeg LLM result—F1 0.396 vs. baseline 0.561—is consistent with this transfer assumption being violated for that encoder/domain. The paper treats this as a heuristic failure, but it is not analyzed quantitatively. To support the central idea, please provide direct evidence about the distance or alignment between the language-sample embeddings and the visual embeddings of matched/unmatched map cells (e.g., per-query boundary margins, or a comparison of text-side and visual-side distributions), or at least state clearly the conditions under which the transfer is expected to hold.
minor comments (6)
  1. [§III] The statement φ(l) = ϱ(v) for corresponding text and visual points is too strong as a formalization; VLM text and image embeddings are not identical in practice. Consider writing "φ(l) ≈ ϱ(v)" or "aligned" to avoid overclaiming, especially given the paper's own discussion of the modality gap.
  2. [§V-C.1, Table I] The entries "Our SVM" and "Our C-SVM" in Table I are not defined in the text. Please clarify which kernel (RBF? linear? cosine?) and which implementation is used for each row, and why both are reported.
  3. [§VI.B] The percentages "227% increase" and "78% improvement" are computed from very small baseline F1 values (0.136). These are misleading without the raw numbers and should be presented as absolute differences or with the baseline values repeated.
  4. [§VI.B, aggregation paragraph] The geometric median experiment is mentioned as giving only a 4% increase in F1, but no table or details are provided. Either include the full comparison or move this to a supplementary material reference.
  5. [General] The figure captions for Figure 5 refer to "(a), (e)" etc. but the subfigure letters are not always obvious in the rendered figure. Please label more clearly and refer to the exact subfigure names in the text.
  6. [§V-C] The number of maps (10) and the query set of 34 labels are stated, but per-class or per-map standard deviations are not reported anywhere. Even for the main table, adding standard deviations would substantially improve the statistical readability.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular steps found; the central result is not fitted to the target data. Score reflects the paper's minor self-citation of the authors' prior queryability metric [8], which is an evaluation protocol rather than a load-bearing derivation step.

full rationale

The derivation chain is self-contained. QuASH (Section IV-C) trains an off-the-shelf SVM on text embeddings φ(S_q ∪ {q}) and φ(A_q ∪ C) and applies the resulting classifier ψ_Q to visual embeddings ϱ(v); the training samples are generated from language heuristics (GPT-4o or lexicon/ground-truth label names), never from the target map cells or image labels. Consequently, the reported F1/IoU improvements over the single-complement baseline are genuine predictions, not re-fitted values. The paper reports honest failure cases: with LSeg, LLM-generated synonyms give F1 0.396 vs baseline 0.561 (Table II), and Section VI.B notes that 'the precision-recall curve of our method could be further optimized'; such negative/contingent results would not appear if the outcome were forced by construction. The 'ground truth' query-set ablation uses the dataset's subclass names as text examples, but it is explicitly an oracle that separates query-set generation from classifier use (Section V-C.2), not a claimed deployable prediction. The only self-citation is [8] (same authors, IROS 2025), used to define 'queryability' and the evaluation protocol: 'We employ the evaluation method described in [8] to assess queryability.' This frames the metrics but does not supply the content of the derivation; precision/recall/F1/IoU are also standard external-benchmark metrics. Therefore there is no circular reduction; the score of 2 reflects only this minor, non-load-bearing self-citation.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical or conceptual entities are introduced. The ledger captures the transfer assumption and the heuristic sampling assumptions that QuASH depends on; the free parameters are design choices whose effect on the central claim is not fully characterized.

free parameters (2)
  • synonym and antonym set sizes = N=20, M=20
    The GPT-4o prompt asks for 20 synonyms and 20 antonyms. No ablation varying N or M is reported, so sensitivity to training-set size is unknown (Section V-C).
  • SVM kernel and hyperparameters = unspecified; scikit-learn defaults or selection
    The paper says SVM with Euclidean and cosine metrics 'had the best performance' after evaluating classifiers, but no validation split or hyperparameter values are reported (Section V-C).
assumptions (4)
  • domain assumption The preimage sets Q and Qbar exist, q is in Q, and a natural-language heuristic h(q) samples from Q and Qbar.
    Section IV-A/IV-C. If the LLM or lexicon supplies synonyms that belong to other classes, the training set is mislabeled; the paper observes this with LLM synonyms in Section V-D.3.
  • domain assumption A classifier trained on text embeddings phi(S_q) and phi(A_q) transfers to visual embeddings rho(V_m).
    Core of QuASH, Section IV-C steps 2-4. The shared VLM latent space is assumed aligned enough that text-only decision boundaries match visual regions; the paper's own LSeg map failure and modality-gap discussion (Section VI.B) cast doubt on this for some encoders.
  • domain assumption VLM embeddings lie on a very narrow cone, effectively an (N-1)-dimensional hyperplane, so thresholding produces isotropic hypercones.
    Section II, citing [11]. This motivates the need for a non-linear classifier but is not proved in this paper.
  • domain assumption Embedding-space directions have semantic meaning and are not uniformly relevant to a query.
    Section IV-B: 'we expect the directions not to be uniformly important with respect to q'. This underpins the claim that a non-linear variance-aware classifier can outperform linear baselines.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QuASH: Using Natural-Language Heuristics to Query Visual-Language Robotic Maps." pith.science (2026). https://pith.science/paper/FE44FS7Z

@misc{pith2026251014546,
  author       = {Pith},
  title        = {Pith review of: QuASH: Using Natural-Language Heuristics to Query Visual-Language Robotic Maps},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FE44FS7Z}},
  note         = {Machine review of arXiv:2510.14546}
}
read the original abstract

Embeddings from Visual-Language Models are increasingly utilized to represent semantics in robotic maps, offering an open-vocabulary scene understanding that surpasses traditional, limited labels. Embeddings enable on-demand querying by comparing embedded user text prompts to map embeddings via a similarity metric. The key challenge in performing the task indicated in a query is that the robot must determine the parts of the environment relevant to the query. This paper proposes a solution to this challenge. We leverage natural-language synonyms and antonyms associated with the query within the embedding space, applying heuristics to estimate the language space relevant to the query, and use that to train a classifier to partition the environment into matches and non-matches. We evaluate our method through extensive experiments, querying both maps and standard image benchmarks. The results demonstrate increased queryability of maps and images. Our querying technique is agnostic to the representation and encoder used, and requires limited training.

Figures

Figures reproduced from arXiv: 2510.14546 by the authors.

Figure 1
Figure 1. Our method, QuASH, works by augmenting a sin [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The state-of-the-art solutions implicitly classify the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The intuition behind our method is that, as the textual [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Examples of predictions from the COCO data set. The original image is shown on the top row, our method using an [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Examples of map query experiment results. Our method with “ground truth” synonyms is shown in the top row, and [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 9 linked inside Pith

  1. [8]

    Do Visual- Language Grid Maps Capture Latent Semantics?

    M. Pekkanen, F. Verdoja, T. Mihaylova, and V . Kyrki, “Do Visual- Language Grid Maps Capture Latent Semantics?” inProc. IEEE/RSJ Int. Conf. on Intell. Robots and Syst. (IROS), Hangzhou, China, Oct. 2025

    2025

  2. [11]

    Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning,

    W. Liang, Y . Zhang, Y . Kwon, S. Yeung, and J. Zou, “Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning,” 2022, arXiv:2203.02053

    arXiv 2022

  3. [1]

    Attention Is All You Need,

    A. Vaswaniet al., “Attention Is All You Need,” inAdvances in Neural Inf. Process. Syst. 30 (NeurIPS), Long Beach, CA, USA, Dec. 2017, pp. 6000–6010

    2017

  4. [2]

    Learning Transferable Visual Models From Natural Language Supervision,

    A. Radfordet al., “Learning Transferable Visual Models From Natural Language Supervision,” inProc. Int. Conf. on Mach. Learn. (ICML), Virtual, Jul. 2021, pp. 8748–8763

    2021

  5. [3]

    LERF: Language Embedded Radiance Fields,

    J. Kerr, C. M. Kim, K. Goldberg, A. Kanazawa, and M. Tancik, “LERF: Language Embedded Radiance Fields,” inProc. IEEE/CVF Int. Conf. on Comput. Vis. (ICCV), Paris, France, Oct. 2023, pp. 19 672–19 682

    2023

  6. [4]

    Visual Language Maps for Robot Navigation,

    C. Huang, O. Mees, A. Zeng, and W. Burgard, “Visual Language Maps for Robot Navigation,” inProc. IEEE Int. Conf. on Robot. and Automat. (ICRA), London, United Kingdom, 5 2023, pp. 10 608– 10 615

    2023

  7. [5]

    Language-Embedded Gaussian Splats (LEGS): Incremen- tally Building Room-Scale Representations with a Mobile Robot,

    J. Yuet al., “Language-Embedded Gaussian Splats (LEGS): Incremen- tally Building Room-Scale Representations with a Mobile Robot,” in Proc. IEEE/RSJ Int. Conf. on Intell. Robots and Syst. (IROS), Abu Dhabi, United Arab Emirates, Oct. 2024, pp. 13 326–13 332

    2024

  8. [6]

    ConceptFusion: Open-set Multimodal 3D Mapping,

    K. M. Jatavallabhulaet al., “ConceptFusion: Open-set Multimodal 3D Mapping,” inProc. Robotics: Sci. and Syst. (RSS), Daegu, Republic of Korea, Jul. 2023

    2023

Show all 30 references
  1. [7]

    ConceptGraphs: Open-V ocabulary 3D Scene Graphs for Perception and Planning,

    Q. Guet al., “ConceptGraphs: Open-V ocabulary 3D Scene Graphs for Perception and Planning,” inProc. IEEE Int. Conf. on Robot. and Automat. (ICRA), Yokohama, Japan, 5 2024, pp. 5021–5028

    2024

  2. [9]

    Learning Generalizable Feature Fields for Mobile Manipulation,

    R.-Z. Qiuet al., “Learning Generalizable Feature Fields for Mobile Manipulation,” 2024, arXiv:2403.07563

    2024 arXiv

  3. [10]

    RAZER: Robust Accel- erated Zero-Shot 3D Open-V ocabulary Panoptic Reconstruction with Spatio-Temporal Aggregation,

    N. Patel, P. Krishnamurthy, and F. Khorrami, “RAZER: Robust Accel- erated Zero-Shot 3D Open-V ocabulary Panoptic Reconstruction with Spatio-Temporal Aggregation,” 2025, arXiv:2505.15373

    2025 arXiv

  4. [12]

    OVExp: Open V ocabulary Exploration for Object-Oriented Navigation,

    M. Wei, T. Wang, Y . Chen, H. Wang, J. Pang, and X. Liu, “OVExp: Open V ocabulary Exploration for Object-Oriented Navigation,” 2024, arXiv:2407.09016

    2024 arXiv

  5. [13]

    A Neural Representation Framework with LLM-Driven Spatial Reasoning for Open-V ocabulary 3D Visual Grounding,

    Z. Liuet al., “A Neural Representation Framework with LLM-Driven Spatial Reasoning for Open-V ocabulary 3D Visual Grounding,” 2025, arXiv:2507.06719

    2025 arXiv

  6. [14]

    SemanticSplat: Feed-Forward 3D Scene Understanding with Language-Aware Gaus- sian Fields,

    Q. Li, J. Sun, L. An, Z. Su, H. Zhang, and Y . Liu, “SemanticSplat: Feed-Forward 3D Scene Understanding with Language-Aware Gaus- sian Fields,” 2025, arXiv:2506.09565

    2025 arXiv

  7. [15]

    CRIS: CLIP-Driven Referring Image Segmentation,

    Z. Wanget al., “CRIS: CLIP-Driven Referring Image Segmentation,” inProc. IEEE/CVF Conf. on Comput. Vis. and Pattern Recognit. (CVPR), New Orleans, LA, USA, Jun. 2022, pp. 11 676–11 685

    2022

  8. [16]

    Image Segmentation Using Text and Image Prompts,

    T. Luddecke and A. Ecker, “Image Segmentation Using Text and Image Prompts,” inProc. IEEE/CVF Conf. on Comput. Vis. and Pattern Recognit. (CVPR), New Orleans, LA, USA, Jun. 2022, pp. 7076–7086

    2022

  9. [17]

    MAG-Nav: Language-Driven Object Navigation Leveraging Memory-Reserved Active Grounding,

    W. Zhang, T. Li, and Y . Liu, “MAG-Nav: Language-Driven Object Navigation Leveraging Memory-Reserved Active Grounding,” 2025, arXiv:2508.05021

    2025 arXiv

  10. [18]

    Open-V ocabulary Indoor Object Grounding with 3D Hierarchical Scene Graph,

    S. Linok and G. Naumov, “Open-V ocabulary Indoor Object Grounding with 3D Hierarchical Scene Graph,” 2025, arXiv:2507.12123

    2025 arXiv

  11. [19]

    A Language-Driven Navigation Strategy Integrating Semantic Maps and Large Language Models,

    Z. Zhong, Y . He, P. Li, F. Yu, and F. Ma, “A Language-Driven Navigation Strategy Integrating Semantic Maps and Large Language Models,” inProc. IEEE/RSJ Int. Conf. on Intell. Robots and Syst. (IROS), Abu Dhabi, United Arab Emirates, Oct. 2024, pp. 9753–9760

    2024

  12. [20]

    QueryAdapter: Rapid Adaptation of Vision-Language Models in Response to Natural Language Queries,

    N. H. Chapman, F. Dayoub, W. Browne, and C. Lehnert, “QueryAdapter: Rapid Adaptation of Vision-Language Models in Response to Natural Language Queries,” 2025, arXiv:2502.18735

    2025 arXiv

  13. [21]

    Uni-Fusion: Universal Continuous Map- ping,

    Y . Yuan and A. Nuechter, “Uni-Fusion: Universal Continuous Map- ping,”IEEE Trans. on Robot. (T-RO), vol. 40, pp. 1373–1392, Jan. 2024

    2024

  14. [22]

    Dynamic Open-V ocabulary 3D Scene Graphs for Long-term Language-Guided Mobile Manipulation,

    Z. Yanet al., “Dynamic Open-V ocabulary 3D Scene Graphs for Long-term Language-Guided Mobile Manipulation,”IEEE Robot. and Automat. Lett., vol. 10, no. 5, pp. 4252–4259, 5 2025

    2025

  15. [23]

    Semantic Mechanical Search with Large Vision and Language Models,

    S. Sharmaet al., “Semantic Mechanical Search with Large Vision and Language Models,” inProc. Conf. on Robot Learn. (CoRL), Atlanta, GA, USA, Nov. 2023, pp. 971–1005

    2023

  16. [24]

    OpenScene: 3D Scene Understanding with Open V ocabularies,

    S. Peng, K. Genova, C. M. Jiang, A. Tagliasacchi, M. Pollefeys, and T. Funkhouser, “OpenScene: 3D Scene Understanding with Open V ocabularies,” inProc. IEEE/CVF Conf. on Comput. Vis. and Pattern Recognit. (CVPR), Vancouver, BC, Canada, Jun. 2023, pp. 815–824

    2023

  17. [25]

    OV-MAP : Open- V ocabulary Zero-Shot 3D Instance Segmentation Map for Robots,

    J. Kim, Y . Park, H. J. Yoon, and B.-T. Zhang, “OV-MAP : Open- V ocabulary Zero-Shot 3D Instance Segmentation Map for Robots,” in Proc. IEEE/RSJ Int. Conf. on Intell. Robots and Syst. (IROS), Abu Dhabi, United Arab Emirates, Oct. 2024, pp. 13 780–13 786

    2024

  18. [26]

    Matterport3D: Learning From RGB-D Data in Indoor Environments,

    A. Changet al., “Matterport3D: Learning From RGB-D Data in Indoor Environments,” inProc. Int. Conf. on 3D Vis. (3DV), Qingdao, China, Oct. 2017, pp. 667–676

    2017

  19. [27]

    Huang, O

    C. Huang, O. Mees, A. Zeng, and W. Burgard. VLMaps. [Online]. Available: https://github.com/vlmaps/vlmaps

  20. [28]

    Language-driven Semantic Segmentation,

    B. Li, K. Q. Weinberger, S. Belongie, V . Koltun, and R. Ranftl, “Language-driven Semantic Segmentation,” inProc. Int. Conf. on Learn. Representations (ICLR), Virtual, Apr. 2022

    2022

  21. [29]

    Scaling Open-V ocabulary Image Segmentation with Image-Level Labels,

    G. Ghiasi, X. Gu, Y . Cui, and T.-Y . Lin, “Scaling Open-V ocabulary Image Segmentation with Image-Level Labels,” inProc. Eur . Conf. on Comput. Vis. (ECCV), Tel Aviv, Israel, Jul. 2022, pp. 540–557

    2022

  22. [30]

    Open-Set Recognition in the Age of Vision-Language Models,

    D. Miller, N. S ¨underhauf, A. Kenna, and K. Mason, “Open-Set Recognition in the Age of Vision-Language Models,” inProc. Eur . Conf. on Comput. Vis. (ECCV), Milano, Italy, Sep. 2024, pp. 1–18

    2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.