Pith. sign in

REVIEW 3 major objections 5 minor 55 references

This paper claims that turning visual place recognition into an interactive dialogue, with a questioner that asks clarifying questions and a retriever that re-ranks candidates each round, markedly improves accuracy on ambiguous natural-lang

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 · deepseek-v4-flash

2026-08-02 14:34 UTC pith:YVQPA72V

load-bearing objection New task and dataset for dialogue-based place recognition, but the evaluation hinges on a synthetic oracle user; the framework is plausible, the numbers are unproven. the 3 major comments →

arxiv 2607.14115 v1 pith:YVQPA72V submitted 2026-05-08 cs.AI cs.CLcs.CV

DialogueVPR: Towards Conversational Visual Place Recognition

Yukun Song , Changwei Wang , Xingtian Pei , Shibiao Xu , Wenhao Xu , Shunpeng Chen , Yu Zhang , Ke Zhang
show 3 more authors
Rongtao Xu Xuxiang Feng Pengyang Wang
This is my paper
classification cs.AI cs.CLcs.CV
keywords dialogue place recognitionvisual geo-localizationinteractive retrievalmultimodal large language modelcurriculum learningreinforcement learningquestion generationbenchmark dataset
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.

Place recognition from natural language is usually done as a one-shot retrieval: a description is matched to a database once. This paper argues that this passive approach fails when descriptions are vague, and proposes Dialogue Place Recognition (DlgPR), in which an intelligent questioner analyzes candidate locations, asks the user a targeted clarifying question, and lets a retriever re-rank candidates with the new information. The authors build a large dialogue-based benchmark, DlgQuest-Cities, and train the questioner first by supervised fine-tuning on easy samples, then by reinforcement learning on hard samples using a reward that directly measures retrieval improvement. On the benchmark, the trained system raises top-1 recall from about 36% after the initial description to about 58% after five rounds, outperforming one-shot retrieval and a 72-billion-parameter vision-language model that does not ask questions. If correct, this makes interactive questioning a practical way to handle ambiguous spatial descriptions.

Core claim

The paper's central claim is that asking questions helps: casting place recognition as a dialogue, where the system reasons about what it cannot yet distinguish and asks the user for the missing evidence, leads to substantially better localization than any single static query. The framework pairs a cross-modal multi-level retriever (CMPL) with an intelligent questioner (DQ-pilot). DQ-pilot is trained in two stages—supervised fine-tuning on low-difficulty dialogues followed by reinforcement learning on high-difficulty ones—using a reward, the Positional Retrieval Gain (PRG), that scores how much a question improves the rank of the correct location. On DlgQuest-Cities, the full method reaches

What carries the argument

The central mechanism is the 'analysis–questioning–optimization' loop. At each round, DQ-pilot executes a four-step chain-of-thought: summarize what the dialogue has confirmed or ruled out, validate the candidate images against that summary, identify visual differentiators among remaining candidates, and phrase one question targeting the most decisive cue. The user's answer is appended to the dialogue history, and CMPL re-retrieves. Two task-aligned metrics carry the learning: the Discriminative Difficulty Index (DDI), a weighted combination of semantic ambiguity and retriever-informed difficulty used to sample easy then hard training dialogues, and the Positional Retrieval Gain (PRG), an nD

Load-bearing premise

The benchmark's scripted, machine-generated user answers are assumed to behave like real human clarifications during evaluation; if human users answer differently, the reported recall gains may not transfer to practice.

What would settle it

Run the trained system in a live study where human users, who have not seen the candidate images, answer DQ-pilot's questions about a held-out set of places, and compare top-1/top-5 recall against the synthetic-answer evaluation on the same queries.

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

If this is right

  • If the results hold, language-guided geo-localization can tolerate vague or incomplete initial descriptions by letting the system ask one or two targeted questions instead of requiring a complete description upfront.
  • The questioner–retriever loop is a general pattern: the same 'ask to disambiguate' mechanism could improve other retrieval tasks, such as person re-identification or image search from partial queries.
  • The PRG reward gives a concrete, measurable objective for training agents to ask useful questions, tied directly to retrieval performance rather than to surrogate heuristics.
  • A 7-billion-parameter model trained for interactive questioning can outperform a 72-billion-parameter model that is not, suggesting interaction can substitute for model scale in spatial-language reasoning.
  • The benchmark and curriculum (DDI-based sampling, then GRPO refinement) provide a reusable training recipe for dialogue-driven retrieval systems.

Where Pith is reading between the lines

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

  • The evaluation uses scripted, machine-generated user answers produced by the same teacher model that created the training dialogues; a human study where real users answer the system's questions on held-out places would test whether the reported gains transfer to genuine conversational behavior.
  • Part of the gain may come from the questioner exploiting simple, highly discriminative cues (like sign text) that the retriever is already good at matching; an adversarially constructed test set could bound this effect.
  • The retriever and questioner are trained separately here; co-training them so the questioner learns to ask about features the retriever is currently worst at could yield further gains.
  • A minimal control—asking the same number of generic or random questions instead of strategically chosen ones—would isolate how much of the improvement is due to question selection versus mere additional text.

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

3 major / 5 minor

Summary. The paper proposes Dialogue Place Recognition (DlgPR), a new task in which a system iteratively asks a user questions to refine visual place retrieval. The authors construct DQ-cities, a large benchmark built on GSV-Cities with synthetic initial captions, fine-grained descriptions, region-level bounding-box annotations, and five-round GPT-4o-generated dialogues. They also propose a framework composed of a cross-modal progressive retriever (CMPL) and a questioner (DQ-pilot) trained with SFT followed by GRPO, guided by a Discriminative Difficulty Index (DDI) and a Positional Retrieval Gain (PRG) reward. Experiments on DQ-cities report large R@1/R@5 improvements over zero-shot baselines and over one-shot retrieval.

Significance. If the results hold, this is a genuinely novel contribution to interactive language-driven geo-localization. The benchmark construction, the curriculum based on DDI, and the coupling of a retriever with a reinforcement-learned questioner are all reasonable and potentially valuable. The release of code and model is a strength that should aid reproducibility. However, the central claim that interactive questioning materially improves localization is currently supported only by evaluations against a synthetic, oracle-like user, and the training and evaluation metrics are partly entangled. The contribution is solid in conception but its empirical validation is not yet convincing.

major comments (3)
  1. [§3.2 Step 3 and §5.2] The evaluation protocol uses synthetic user answers generated by GPT-4o from the same evidence base (fine-grained descriptions and bounding boxes, E_i) that was used to generate DQ-pilot's training dialogues. Table 1's gains (e.g., DQ-pilot R@1 from 35.9 at round 0 to 58.4 at round 5) therefore assume that real users answer with the same region-grounded, oracle-like content. No human evaluation, no real-user simulation, and no sensitivity analysis to answer noise or incorrect answers are provided. This is the load-bearing assumption for the central claim. Please add a human evaluation subset or, at minimum, simulate answer noise, partial answers, and refusals to show the method degrades gracefully.
  2. [§4.3 Eq. (8)-(9), §3.2 Eq. (2)-(3), Table 1] PRG is used in three places: the DDI curriculum filter (Eq. 3), the GRPO retrieval reward (Eq. 8, combined in Eq. 9), and the evaluation metric reported in Table 1's PRG column. This means the reported PRG improvements are directly optimized for the same CMPL retriever and the same candidate distribution used at evaluation; they are not an independent measure of question quality. While R@1/R@5 are not directly the training objective, the reward shaping and data selection are aligned with the exact protocol later used to measure success. The paper should decouple this, e.g., evaluate with a retriever not used during reward computation, or report R@1/R@5 with fixed candidate sets independent of CMPL.
  3. [§5.1, Table 1] The main baselines (Qwen2.5-VL-7B/72B, PlugIR) are used zero-shot, while DQ-pilot is fine-tuned on 30k DQ-cities dialogues. The large gains in Table 1 thus conflate task adaptation with the benefit of interactive reasoning. A same-data trained baseline (e.g., fine-tuning Qwen2.5-VL-7B or PlugIR on the same curriculum splits) is necessary to isolate the contribution. In addition, no error bars or significance tests are reported; Table 1 appears to be a single run. The claim 'significantly outperforms baselines' (Abstract) is not fully supported without variance estimates across seeds or test splits.
minor comments (5)
  1. [Eq. (1)] The text says SA is computed from 'positive and negative textual embeddings ... and their corresponding visual embeddings,' but the equation uses only the text encoder φ_T. Please clarify whether visual embeddings are intended and, if so, correct the equation.
  2. [Table 1] The PRG column is not defined: is it PRG at round 5, average over rounds, or across regions? BRI is mentioned as from PlugIR but never defined. Please add definitions.
  3. [§5.2] The sentence 'Compared to its 7B backbone, our model improves R@1 by 9.2% and 13.4% after 3 and 5 dialogue rounds' does not match Table 1. For example, round-3 R@1 is 52.1 vs 42.4 (an absolute difference of 9.7 points), and round-5 is 58.4 vs 43.2 (15.2 points). Please correct the numbers or state what comparison is being reported.
  4. [Figure 2] Typo: 'Chain of Thought promot' should be 'prompt'.
  5. [General] The paper repeatedly refers to an Appendix, but the preprint text provided does not include one. Please ensure the appendix is present in the final version and that all referenced details are available.

Circularity Check

1 steps flagged

No load-bearing circularity; one minor self-referential metric (PRG as both GRPO reward and reported evaluation column) and a synthetic-user external-validity risk, but the R@1 comparison is an independent empirical result.

specific steps
  1. other [Sec. 4.3, Eq. (8); Table 1]
    "We reuse the Positional Retrieval Gain (PRG) from Sec 3.2 as a task-aligned measure ... the retrieval reward is defined as Rprg = PRGt, (8)"

    The same PRG scalar optimized by GRPO is also reported as an evaluation column in Table 1. Consequently the PRG improvement shown for DQ-pilot is the training objective itself, not an independent measurement of question quality. The headline R@1/R@5 gains are computed separately and are not forced by this identity, so the circularity is minor and non-load-bearing.

full rationale

The paper's derivation chain is otherwise self-contained. CMPL is trained from scratch with image-text alignment losses (Eq. 6); DQ-pilot is initialized by SFT on GPT-4o-generated dialogues and refined by GRPO with a verifiable reward. No central premise rests on a self-citation: the authors' earlier works [7,8,41] appear only in related-work lists, no uniqueness theorem is imported, and no ansatz is smuggled in by citation. The strongest concern is external validity rather than circularity: the DQ-cities dialogues and evaluation answers are synthesized by GPT-4o from the same CMPL candidate sets and evidence base (Sec. 3.2 Step 3), so Table 1 may measure fit to a synthetic oracle rather than human clarification behavior. That is an evaluation-protocol limitation, not an equation-level reduction of the claimed result. Hence score 2 for the minor PRG reward-metric overlap; the central R@1/R@5 claim has independent content.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 3 invented entities

The central claims rest on the fidelity of synthetic GPT-4o data, on the validity of the DDI/PRG metrics, and on many unreported hyperparameters. Since there is no external human-annotated benchmark or real-user test, the paper's empirical contribution is currently self-contained rather than independently grounded.

free parameters (6)
  • SA mixing weight α (Eq. 1) = not reported
    Balances semantic ambiguity terms in the difficulty score; value affects curriculum sampling.
  • DDI weights w_sa, w_rid (Eq. 3) = not reported
    Determine the curriculum split between easy (SFT) and hard (GRPO) data.
  • PRG filtering threshold τ1 = not reported
    Used to filter low-quality dialogues during dataset construction; a post-hoc data selection choice.
  • GRPO reward weights α, β (Eq. 9) = not reported
    Balance retrieval gain and format rewards during reinforcement learning.
  • CMPL loss weights λ_gs, λ_h and margin α (Eq. 6) = not reported
    Weights for global/local losses and triplet margin; values affect retrieval training but are not given.
  • SDM temperature τ (Eq. 5) = not reported
    Temperature in similarity distribution matching; not specified.
axioms (5)
  • domain assumption GPT-4o-generated long descriptions are factually accurate and sufficient for place-level retrieval.
    CMPL and the dialogues are trained on these descriptions; if they hallucinate or omit decisive details, the entire pipeline is unreliable (§3.2 Step 1).
  • domain assumption Scripted user answers in DQ-cities behave like real human clarification behavior.
    Evaluation assumes oracle answers from the teacher-generated dialogues; no real-user study is conducted (§5.2).
  • ad hoc to paper Hard-negative candidate sets produced by CMPL during data generation represent the difficulty distribution at test time.
    The questioner may learn only to distinguish CMPL's preferred negatives rather than general visual ambiguity (§3.2 Step 3).
  • ad hoc to paper DDI and PRG correctly quantify dialogue difficulty and retrieval gain.
    These metrics are introduced in the paper and used for training and evaluation, but are not validated against human judgments or external benchmarks (§3.2 Step 4, Eq. 2).
  • domain assumption GSV-Cities place labels are correct and a valid foundation for the benchmark.
    The dataset inherits all ground-truth labels from GSV-Cities without additional verification (§3.1).
invented entities (3)
  • DlgQuest-Cities benchmark no independent evidence
    purpose: Training and evaluation for dialogue-based place recognition.
    The benchmark is entirely synthesized by the paper's GPT-4o pipeline; no external human dialogue corpus or independent evaluation set exists yet.
  • Discriminative Difficulty Index (DDI) no independent evidence
    purpose: Curriculum sampling to split easy and hard training instances.
    A composite metric with unreported weights; no external validation that it measures true conversational difficulty.
  • Positional Retrieval Gain (PRG) no independent evidence
    purpose: Curriculum filter, GRPO reward, and reported evaluation metric.
    Defined, optimized, and evaluated within the same paper; the overlap between reward and metric limits independent confirmation.

pith-pipeline@v1.3.0-alltime-deepseek · 13777 in / 11972 out tokens · 125579 ms · 2026-08-02T14:34:27.384526+00:00 · methodology

0 comments
read the original abstract

Inspired by how humans communicate spatial information, language-guided geo-localization has gained significant traction for its intuitive and practical value. Despite this progress, most methods still rely on a static, one-shot retrieval paradigm, which fails to handle the ambiguity and incompleteness inherent in real-world natural language descriptions. We propose a paradigm shift to reasoning retrieval and introduce Dialogue Place Recognition (DlgPR), which casts localization as an interactive, dialogue-driven reasoning process. To support this new task, we present DlgQuest-Cities, the first large-scale dialogue-based benchmark for place recognition, and a unified reasoning framework that couples a cross-modal multi-level retriever with an intelligent questioner, DQ-pilot. DQ-pilot is trained in a curriculum: supervised fine-tuning on a curated DQ-cities-20k subset followed by reinforcement refinement on a harder DQ-cities-10k split via GRPO. Two task-aligned metrics guide learning: a Discriminative Difficulty Index (DDI) for curriculum sampling and a Positional Retrieval Gain (PRG) reward that directly measures retrieval improvement induced by a question. Experiments show this reasoning-based approach significantly outperforms baselines. The code and model are available at https://github.com/Graysonggg/DlgPR.

Figures

Figures reproduced from arXiv: 2607.14115 by Changwei Wang, Ke Zhang, Pengyang Wang, Rongtao Xu, Shibiao Xu, Shunpeng Chen, Wenhao Xu, Xingtian Pei, Xuxiang Feng, Yukun Song, Yu Zhang.

Figure 1
Figure 1. Figure 1: An Illustration of the Dialogue Place Recognition (DlgPR) Framework. The key component, DQ-Pilot, functions as a reasoning agent that transforms geolocalization from a simple one-shot ”retrieval” into a sophisticated ”reasoning-based retrieval” process: the user provides an ”initial description”, and the retriever CMPL performs a preliminary retrieval (Round 0) to generate multiple visu￾ally similar candid… view at source ↗
Figure 2
Figure 2. Figure 2: Dataset Construction Flowchart. It is mainly divided into 4 parts: Text modality expansion; Region-level visual evidence [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Flowchart of the proposed cross-modal progressive learning retriever. The core is the cross-modal progressive learning (CMPL) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] 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

55 extracted references · 15 linked inside Pith

  1. [1]

    Gsv-cities: Toward appropriate supervised visual place recognition.Neurocomputing, 513:194–203, 2022

    Amar Ali-bey, Brahim Chaib-draa, and Philippe Gigu `ere. Gsv-cities: Toward appropriate supervised visual place recognition.Neurocomputing, 513:194–203, 2022. 3

    2022

  2. [2]

    Global proxy-based hard mining for visual place recognition

    Amar Ali-Bey, Brahim Chaib-draa, and Philippe Gigu `ere. Global proxy-based hard mining for visual place recognition. arXiv preprint arXiv:2302.14217, 2023. 2

    Pith/arXiv arXiv 2023

  3. [3]

    BoQ: A place is worth a bag of learnable queries

    Amar Ali-bey, Brahim Chaib-draa, and Philippe Gigu `ere. BoQ: A place is worth a bag of learnable queries. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17794–17803, 2024

    2024

  4. [4]

    Netvlad: Cnn architecture for weakly supervised place recognition

    Relja Arandjelovic, Petr Gronat, Akihiko Torii, Tomas Pa- jdla, and Josef Sivic. Netvlad: Cnn architecture for weakly supervised place recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5297–5307, 2016. 2

    2016

  5. [5]

    Ask&confirm: active detail enriching for cross-modal retrieval with partial query

    Guanyu Cai, Jun Zhang, Xinyang Jiang, Yifei Gong, Lianghua He, Fufu Yu, Pai Peng, Xiaowei Guo, Feiyue Huang, and Xing Sun. Ask&confirm: active detail enriching for cross-modal retrieval with partial query. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 1835–1844, 2021. 3

    2021

  6. [6]

    where am i?

    Jiaqi Chen, Daniel Barath, Iro Armeni, Marc Pollefeys, and Hermann Blum. “where am i?” scene retrieval with lan- guage. InEuropean Conference on Computer Vision, pages 201–220. Springer, 2024. 2

    2024

  7. [7]

    Region matters: Efficient and re- liable region-aware visual place recognition.arXiv preprint arXiv:2604.22390, 2026

    Shunpeng Chen, Yukun Song, Changwei Wang, Rong- tao Xu, Kexue Fu, Longxiang Gao, Li Guo, Ruisheng Wang, and Shibiao Xu. Region matters: Efficient and re- liable region-aware visual place recognition.arXiv preprint arXiv:2604.22390, 2026. 2

    Pith/arXiv arXiv 2026

  8. [8]

    SAGE: Spatial-visual adaptive graph exploration for efficient visual place recognition

    Shunpeng Chen, Changwei Wang, Rongtao Xu, Xingtian Pei, Yukun Song, Jinzhou Lin, Wenhao Xu, Jingyi Zhang, Li Guo, and Shibiao Xu. SAGE: Spatial-visual adaptive graph exploration for efficient visual place recognition. In The Fourteenth International Conference on Learning Rep- resentations, 2026. 2

    2026

  9. [9]

    Towards natural language-guided drones: Geotext-1652 benchmark with spatial relation matching

    Meng Chu, Zhedong Zheng, Wei Ji, Tingyu Wang, and Tat-Seng Chua. Towards natural language-guided drones: Geotext-1652 benchmark with spatial relation matching. In European Conference on Computer Vision, pages 213–231. Springer, 2024. 3

    2024

  10. [10]

    Sam- ple4geo: Hard negative sampling for cross-view geo- localisation

    Fabian Deuser, Konrad Habel, and Norbert Oswald. Sam- ple4geo: Hard negative sampling for cross-view geo- localisation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 16847–16856, 2023. 2

    2023

  11. [11]

    Seq- matchnet: Contrastive learning with sequence matching for place recognition & relocalization

    Sourav Garg, Madhu Vankadari, and Michael Milford. Seq- matchnet: Contrastive learning with sequence matching for place recognition & relocalization. InConference on Robot Learning, pages 429–443. PMLR, 2022. 2

    2022

  12. [12]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025. 3

    Pith/arXiv arXiv 2025

  13. [13]

    Textplace: Visual place recognition and topolog- ical localization through reading scene texts

    Ziyang Hong, Yvan Petillot, David Lane, Yishu Miao, and Sen Wang. Textplace: Visual place recognition and topolog- ical localization through reading scene texts. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 2861–2870, 2019. 3

    2019

  14. [14]

    Progeo: Generating prompts through image- text contrastive learning for visual geo-localization

    Jingqi Hu, Chen Mao, Chong Tan, Hui Li, Hong Liu, and Min Zheng. Progeo: Generating prompts through image- text contrastive learning for visual geo-localization. InIn- ternational Conference on Artificial Neural Networks, pages 448–462. Springer, 2024. 2, 3

    2024

  15. [15]

    Optimal transport ag- gregation for visual place recognition

    Sergio Izquierdo and Javier Civera. Optimal transport ag- gregation for visual place recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2

    2024

  16. [16]

    Close, but not there: Boosting geographic distance sensitivity in visual place recognition

    Sergio Izquierdo and Javier Civera. Close, but not there: Boosting geographic distance sensitivity in visual place recognition. InComputer Vision – ECCV 2024, pages 240– 257, Cham, 2025. Springer Nature Switzerland. 2

    2024

  17. [17]

    Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richard- son, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024. 3

    Pith/arXiv arXiv 2024

  18. [18]

    Cross-modal implicit relation rea- soning and aligning for text-to-image person retrieval

    Ding Jiang and Mang Ye. Cross-modal implicit relation rea- soning and aligning for text-to-image person retrieval. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 2787–2797, 2023. 5

    2023

  19. [19]

    Text2pos: Text-to-point-cloud cross-modal localiza- tion

    Manuel Kolmet, Qunjie Zhou, Aljo ˇsa Oˇsep, and Laura Leal- Taix´e. Text2pos: Text-to-point-cloud cross-modal localiza- tion. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 6687–6696,

  20. [20]

    Whittlesearch: Interactive image search with relative at- tribute feedback.International Journal of Computer Vision, 115(2):185–210, 2015

    Adriana Kovashka, Devi Parikh, and Kristen Grauman. Whittlesearch: Interactive image search with relative at- tribute feedback.International Journal of Computer Vision, 115(2):185–210, 2015. 3

    2015

  21. [21]

    Cosmo: Content-style modulation for image retrieval with text feed- back

    Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feed- back. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 802–812, 2021. 3

    2021

  22. [22]

    Interactive text-to-image retrieval with large language models: A plug-and-play approach.arXiv preprint arXiv:2406.03411, 2024

    Saehyung Lee, Sangwon Yu, Junsung Park, Jihun Yi, and Sungroh Yoon. Interactive text-to-image retrieval with large language models: A plug-and-play approach.arXiv preprint arXiv:2406.03411, 2024. 3, 7, 8

    Pith/arXiv arXiv 2024

  23. [23]

    Chatting makes perfect: Chat-based image retrieval

    Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischin- ski. Chatting makes perfect: Chat-based image retrieval. Advances in Neural Information Processing Systems, 36: 61437–61449, 2023. 3

    2023

  24. [24]

    Omnicity: Omnipotent city understanding with multi-level and multi- view images

    Weijia Li, Yawen Lai, Linning Xu, Yuanbo Xiangli, Jinhua Yu, Conghui He, Gui-Song Xia, and Dahua Lin. Omnicity: Omnipotent city understanding with multi-level and multi- view images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17397– 17407, 2023. 2

    2023

  25. [25]

    Simple baselines for interactive video retrieval with questions and answers

    Kaiqu Liang and Samuel Albanie. Simple baselines for interactive video retrieval with questions and answers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11091–11101, 2023. 3

    2023

  26. [26]

    Visual- rft: Visual reinforcement fine-tuning.arXiv preprint arXiv:2503.01785, 2025

    Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual- rft: Visual reinforcement fine-tuning.arXiv preprint arXiv:2503.01785, 2025. 3

    Pith/arXiv arXiv 2025

  27. [27]

    Towards seamless adapta- tion of pre-trained models for visual place recognition

    Feng Lu, Lijun Zhang, Xiangyuan Lan, Shuting Dong, Yaowei Wang, and Chun Yuan. Towards seamless adapta- tion of pre-trained models for visual place recognition. In The Twelfth International Conference on Learning Represen- tations, 2024. 2

    2024

  28. [28]

    Supervlad: Com- pact and robust image descriptors for visual place recogni- tion.Advances in Neural Information Processing Systems, 37:5789–5816, 2024

    Feng Lu, Xinyao Zhang, Canming Ye, Shuting Dong, Lijun Zhang, Xiangyuan Lan, and Chun Yuan. Supervlad: Com- pact and robust image descriptors for visual place recogni- tion.Advances in Neural Information Processing Systems, 37:5789–5816, 2024. 2

    2024

  29. [29]

    Llava-reid: Selective multi-image ques- tioner for interactive person re-identification.arXiv preprint arXiv:2504.10174, 2025

    Yiding Lu, Mouxing Yang, Dezhong Peng, Peng Hu, Yijie Lin, and Xi Peng. Llava-reid: Selective multi-image ques- tioner for interactive person re-identification.arXiv preprint arXiv:2504.10174, 2025. 3

    Pith/arXiv arXiv 2025

  30. [30]

    Tell me where you are: Multimodal llms meet place recognition.arXiv preprint arXiv:2406.17520, 2024

    Zonglin Lyu, Juexiao Zhang, Mingxuan Lu, Yiming Li, and Chen Feng. Tell me where you are: Multimodal llms meet place recognition.arXiv preprint arXiv:2406.17520, 2024. 2, 3

    Pith/arXiv arXiv 2024

  31. [31]

    Learn- ing to retrieve videos by asking questions

    Avinash Madasu, Junier Oliva, and Gedas Bertasius. Learn- ing to retrieve videos by asking questions. InProceedings of the 30th ACM International Conference on Multimedia, pages 356–365, 2022. 3

    2022

  32. [32]

    Emvp: Em- bracing visual foundation model for visual place recognition with centroid-free probing.Advances in Neural Information Processing Systems, 37:120928–120950, 2024

    Qibo Qiu, Shun Zhang, Haiming Gao, Honghui Yang, Haochao Ying, Wenxiao Wang, and Xiaofei He. Emvp: Em- bracing visual foundation model for visual place recognition with centroid-free probing.Advances in Neural Information Processing Systems, 37:120928–120950, 2024. 2

    2024

  33. [33]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 8

    2021

  34. [34]

    From coarse to fine: Robust hierarchical localization at large scale

    Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12716–12725, 2019. 2

    2019

  35. [35]

    Bridging text and vision: A multi-view text-vision registration approach for cross- modal place recognition.arXiv preprint arXiv:2502.14195,

    Tianyi Shang, Zhenyu Li, Pengjie Xu, Jinwei Qiao, Gang Chen, Zihan Ruan, and Weijun Hu. Bridging text and vision: A multi-view text-vision registration approach for cross- modal place recognition.arXiv preprint arXiv:2502.14195,

    Pith/arXiv arXiv

  36. [36]

    Deepseekmath: Pushing the limits of math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024. 3

    Pith/arXiv arXiv 2024

  37. [37]

    Vlm-r1: A stable and generaliz- able r1-style large vision-language model.arXiv preprint arXiv:2504.07615, 2025

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, et al. Vlm-r1: A stable and generaliz- able r1-style large vision-language model.arXiv preprint arXiv:2504.07615, 2025. 3

    Pith/arXiv arXiv 2025

  38. [38]

    Where am i looking at? joint location and orientation es- timation by cross-view matching

    Yujiao Shi, Xin Yu, Dylan Campbell, and Hongdong Li. Where am i looking at? joint location and orientation es- timation by cross-view matching. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4064–4072, 2020. 2

    2020

  39. [39]

    Textinplace: Indoor vi- sual place recognition in repetitive structures with scene text spotting and verification.arXiv preprint arXiv:2503.06501,

    Huaqi Tao, Bingxi Liu, Calvin Chen, Tingjun Huang, He Li, Jinqiang Cui, and Hong Zhang. Textinplace: Indoor vi- sual place recognition in repetitive structures with scene text spotting and verification.arXiv preprint arXiv:2503.06501,

    Pith/arXiv arXiv

  40. [40]

    Loc4plan: Locating be- fore planning for outdoor vision and language navigation

    Huilin Tian, Jingke Meng, Wei-Shi Zheng, Yuan-Ming Li, Junkai Yan, and Yunong Zhang. Loc4plan: Locating be- fore planning for outdoor vision and language navigation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4073–4081, 2024. 2

    2024

  41. [41]

    Focus on local: Finding reliable discriminative regions for visual place recognition

    Changwei Wang, Shunpeng Chen, Yukun Song, Rongtao Xu, Zherui Zhang, Jiguang Zhang, Haoran Yang, Yu Zhang, Kexue Fu, Shide Du, et al. Focus on local: Finding reliable discriminative regions for visual place recognition. InPro- ceedings of the AAAI Conference on Artificial Intelligence, pages 7536–7544, 2025. 2

    2025

  42. [42]

    Lvlm-empowered multi-modal representation learning for visual place recognition.arXiv preprint arXiv:2407.06730,

    Teng Wang, Lingquan Meng, Lei Cheng, and Changyin Sun. Lvlm-empowered multi-modal representation learning for visual place recognition.arXiv preprint arXiv:2407.06730,

    Pith/arXiv arXiv

  43. [43]

    Multi-similarity loss with general pair weighting for deep metric learning

    Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R Scott. Multi-similarity loss with general pair weighting for deep metric learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5022–5030, 2019. 5

    2019

  44. [44]

    Fine-grained cross-view geo-localization using a correlation-aware homography estimator.Advances in Neu- ral Information Processing Systems, 36:5301–5319, 2023

    Xiaolong Wang, Runsen Xu, Zhuofan Cui, Zeyu Wan, and Yu Zhang. Fine-grained cross-view geo-localization using a correlation-aware homography estimator.Advances in Neu- ral Information Processing Systems, 36:5301–5319, 2023. 2

    2023

  45. [45]

    A theoretical analysis of ndcg type ranking measures

    Yining Wang, Liwei Wang, Yuanzhi Li, Di He, and Tie-Yan Liu. A theoretical analysis of ndcg type ranking measures. In Conference on learning theory, pages 25–54. PMLR, 2013. 5

    2013

  46. [46]

    Text2loc: 3d point cloud localization from natural language

    Yan Xia, Letian Shi, Zifeng Ding, Joao F Henriques, and Daniel Cremers. Text2loc: 3d point cloud localization from natural language. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 14958–14967, 2024. 2, 3

    2024

  47. [47]

    Adapting fine-grained cross-view localization to areas with- out fine ground truth

    Zimin Xia, Yujiao Shi, Hongdong Li, and Julian FP Kooij. Adapting fine-grained cross-view localization to areas with- out fine ground truth. InEuropean Conference on Computer Vision, pages 397–415. Springer, 2024. 2

    2024

  48. [48]

    Flair: Vlm with fine- grained language-informed image representations

    Rui Xiao, Sanghwan Kim, Mariana-Iuliana Georgescu, Zeynep Akata, and Stephan Alaniz. Flair: Vlm with fine- grained language-informed image representations. InPro- ceedings of the Computer Vision and Pattern Recognition Conference, pages 24884–24894, 2025. 8

    2025

  49. [49]

    Fg- clip: Fine-grained visual and textual alignment.arXiv preprint arXiv:2505.05071, 2025

    Chunyu Xie, Bin Wang, Fanjing Kong, Jincheng Li, Dawei Liang, Gengshen Zhang, Dawei Leng, and Yuhui Yin. Fg- clip: Fine-grained visual and textual alignment.arXiv preprint arXiv:2505.05071, 2025. 4, 8

    Pith/arXiv arXiv 2025

  50. [50]

    Skydiffusion: Street-to-satellite im- age synthesis with diffusion models and bev paradigm.arXiv e-prints, pages arXiv–2408, 2024

    Junyan Ye, Jun He, Weijia Li, Zhutao Lv, Jinhua Yu, Haote Yang, and Conghui He. Skydiffusion: Street-to-satellite im- age synthesis with diffusion models and bev paradigm.arXiv e-prints, pages arXiv–2408, 2024. 2

    2024

  51. [51]

    Cross-view image geo- localization with panorama-bev co-retrieval network

    Junyan Ye, Zhutao Lv, Weijia Li, Jinhua Yu, Haote Yang, Huaping Zhong, and Conghui He. Cross-view image geo- localization with panorama-bev co-retrieval network. In European Conference on Computer Vision, pages 74–90. Springer, 2024. 2

    2024

  52. [52]

    Where am i? cross-view geo-localization with natural language descrip- tions

    Junyan Ye, Honglin Lin, Leyan Ou, Dairong Chen, Zihao Wang, Qi Zhu, Conghui He, and Weijia Li. Where am i? cross-view geo-localization with natural language descrip- tions. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 5890–5900, 2025. 2, 3

    2025

  53. [53]

    Long-clip: Unlocking the long-text capability of clip

    Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. InEuropean conference on computer vision, pages 310–325. Springer, 2024. 8

    2024

  54. [54]

    Navid: Video-based vlm plans the next step for vision-and-language navigation.arXiv preprint arXiv:2402.15852, 2024

    Jiazhao Zhang, Kunyu Wang, Rongtao Xu, Gengze Zhou, Yicong Hong, Xiaomeng Fang, Qi Wu, Zhizheng Zhang, and He Wang. Navid: Video-based vlm plans the next step for vision-and-language navigation.arXiv preprint arXiv:2402.15852, 2024. 2

    Pith/arXiv arXiv 2024

  55. [55]

    Enhancing interactive image retrieval with query rewriting using large language models and vision lan- guage models

    Hongyi Zhu, Jia-Hong Huang, Stevan Rudinac, and Evan- gelos Kanoulas. Enhancing interactive image retrieval with query rewriting using large language models and vision lan- guage models. InProceedings of the 2024 International Conference on Multimedia Retrieval, pages 978–987, 2024. 3

    2024