Pith. sign in

REVIEW 4 major objections 4 minor 26 references

Training-free optimal-transport reranker lifts 3D scene retrieval to 63.9% top-1 on hard subsets.

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

2026-08-01 13:23 UTC pith:A6SHSGPL

load-bearing objection The benchmark is the real deliverable, but it is built from the same predicates and LLM family the method uses, so the +36.7-point gain is an open question about external validity, not a closed result. the 4 major comments →

arxiv 2607.19115 v1 pith:A6SHSGPL submitted 2026-07-21 cs.CV

CR-Refiner: An Object-Centric Optimal Transport Reranker for Edit-Conditioned 3D Scene Retrieval

Hao Wu , Jinjing Zhu , Nanyu Wu , Qianyi Cai , Heyi Lin , Hao Wang , Hui Xiong This is my paper
classification cs.CV
keywords edit-conditioned 3D scene retrievalcomposed image retrievalrerankingunbalanced optimal transportstructural prior3D-CER benchmarktraining-freescene understanding
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.

This paper argues that edit-conditioned 3D scene retrieval can be substantially improved without any training by treating reranking as an asymmetric matching problem. The proposed CR-Refiner parses the edit into a structured query entity and scores each candidate room with an unbalanced optimal-transport problem over a cost matrix that couples category, style, material, and geometry, allowing the single edited object to 'ignore' irrelevant furniture. It adds an axis-conditional structural prior for size and spatial cues and an LLM verifier to refine near ties. On the released 3D-CER benchmark, the method improves hard-subset R@1 and mAP@10 across three qualitatively different base retrievers, on every edit axis. A sympathetic reader would care because it offers a plug-in, training-free upgrade path for existing retrievers, and the benchmark fills a gap in evaluating compositional matching over 3D object sets.

Core claim

CR-Refiner is a training-free reranker that takes any base retriever's top-K candidates and rescore them. The edit is parsed by a frozen LLM into a structured query entity (target category, style, material, optional spatial direction, and edit axis). Each candidate is scored by an unbalanced regularised optimal-transport problem over a 1×G cost matrix where G is the number of objects in the candidate; the four mismatch channels (category, style, material, direction) aggregate into a single cost per object, and the unbalanced solver lets the unit query mass be dropped rather than split across irrelevant objects. An axis-conditional structural prior adds bounding-box size cues for geometric ed

What carries the argument

The core mechanism is the unbalanced optimal-transport solver over an asymmetric 1×G cost matrix M, where the four attribute mismatch indicators are weighted and aggregated; the KL slack τ on the row marginal lets the single-entity query drop mass on objects that do not match, directly modelling the one-to-many asymmetry of an edit that names one object in a room full of furniture. The structural prior (size-class bonus for geometric/compound axes, subject-anchor direction bonus for spatial/compound axes) and the LLM verifier for the top three candidates are complementary. The entire pipeline has five scalar hyperparameters selected on validation.

Load-bearing premise

The benchmark is built from the same programmatic operator predicates and the same 3D-FUTURE attribute strings that CR-Refiner's cost matrix and structural prior use for matching, so the reported gains may largely measure how well the pipeline recovers annotations it helped generate.

What would settle it

Take a random sample of 3D-CER test queries, replace each paraphrase with a human-authored rewording of the same edit (keeping the operator fixed), and rerun CR-Refiner and the baselines. If the R@1 gain over Pic2Word drops by more than half, the method's advantage is tied to the benchmark's generation procedure rather than to generalizable edit reasoning.

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

If this is right

  • Any training-free or trained base retriever can be wrapped with CR-Refiner to gain 30-39 R@1 points on hard-subset reranking, with no retraining.
  • The unbalanced OT formulation is the key to asymmetric one-to-many matching; balanced Sinkhorn dilutes cost across all objects and loses 10.5 points on material edits.
  • The axis-conditional structural prior accounts for most of the gain: removing it costs 34.2 R@1 overall, especially on geometric (-61.2) and spatial (-35.8) edits.
  • The LLM verifier mainly helps style and compound edits, demoting candidates that match keyword-level style but fail contextually.
  • Full-corpus retrieval remains unsolved: even the strongest base retriever recalls only 88% of positives at K=1000, and reranking lift dies beyond K≈100.

Where Pith is reading between the lines

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

  • The reported hard-subset gains may partly reflect the benchmark's construction: paraphrases were generated by the same model family used as the parser and verifier, and the ground-truth predicates use the same attribute vocabulary as the cost matrix. A human-written rephrasing of the same edits could yield smaller gains.
  • The unbalanced OT formulation should transfer to other one-to-many retrieval tasks, e.g., referring-expression grounding or object-centric 2D retrieval, wherever a named entity must be matched against a set.
  • The spatial-branch reliance on a global frame convention in 3D-FRONT may not carry over to arbitrary room scans with inconsistent global orientation; a coordinate-frame alignment step would be needed.
  • The LLM verifier's continuous confidence could plausibly be replaced by a cheaper learned calibrator, but the paper does not test that; the verifier's two-second per-query cost is the main runtime bottleneck.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes CR-Refiner, a training-free reranker for edit-conditioned 3D scene retrieval. A frozen LLM parses a reference room and edit text into a structured query entity; each candidate is scored by an unbalanced optimal-transport problem over a 1×G cost matrix coupling category, style, material, and direction, then an axis-conditional structural prior adds size/direction bonuses, and an LLM verifier refines the top-3 candidates. The authors also release 3D-CER, a 4,963-query benchmark over 3D-FRONT/3D-FUTURE with multi-positive ground truth, CIRR-style hard subsets, and zero-target adversarials. On the hard subset (n=1,274), full CR-Refiner achieves R@1=0.639 and mAP@10=0.663, substantially above text-conditioned baselines, with bootstrap CIs and per-axis breakdowns.

Significance. If 3D-CER is a valid proxy for human edit satisfaction, the result is practically significant: a training-free reranker can produce large gains on curated pools, and the paper's full-corpus recall ceiling analysis is an unusually honest scope check. Strengths include clearly specified components, reproducible pseudocode, bootstrap CIs, and ablations showing per-axis contributions. The central risk is benchmark validity: the ground-truth predicate in Phase 5, the ℓ indicators in Eq. 1, and the GPT-5.1 paraphrase/verifier all share the same 3D-FUTURE attribute strings and model family. This does not make the internal experiments inconsistent, but it means the headline gain may partly reflect recovery of the annotation process rather than generalizable edit understanding.

major comments (4)
  1. [§4, Appendix B Phases 3–5; Eq. 1] Benchmark circularity: the multi-positive ground truth is mined by operator predicates requiring exact string equality on 3D-FUTURE style/material/category attributes, and Eq. 1's ℓ indicators check the same strings. The GPT-5.1 paraphrases (Phase 3) are written by the same model family later used as parser and verifier. Thus the +36.7 R@1 gain (Table 2) may measure how well CR-Refiner inverts the annotation process rather than human edit satisfaction. The Gemini cross-check and programmatic validation do not break this alignment because they share the same vocabulary and are also LLMs. Please add a human-annotated evaluation on a random sample (or an independently written query set) with inter-annotator agreement, and compare CR-Refiner's rankings against the operator-predicate oracle.
  2. [§5.1, Table 2, §4] Zero-target adversarials are never evaluated. §5.1 lists Abstention F1 as a metric and §4 states that about one fifth of test queries are zero-target (348 in test) and require abstention. However, no table reports results on these queries, and no abstention mechanism is defined for CR-Refiner: Eq. 3 always produces a ranking, so the method cannot abstain unless an external threshold is added. This is a missing evaluation of a stated benchmark component and of the model's behavior on impossible edits.
  3. [§5.6, Tables 8–9] Full-corpus versus hard-subset framing: end-to-end full-corpus R@1 is 0.010–0.055 for K=10–100 and declines at larger pools, and even the strongest base retriever recalls only 7.5% of positives at K=1. The title and opening sentence promise '3D scene retrieval,' but CR-Refiner is only effective on curated 20-distractor pools. The abstract hedges with 'hard-subset,' but introduction and conclusion state the task more broadly. Please reframe the contribution as fixed-pool reranking or provide a candidate generator that makes full-corpus retrieval viable.
  4. [§3.2, §3.3, Table 5] Spatial direction prior relies on an unquantified alignment assumption. §3.2 says ℓdir is unreliable because 3D-FRONT bounding-box centres live in scene-specific world frames; §3.3 then asserts 'most rooms align to consistent global frame conventions' and uses the direction bonus as the main source of the spatial-axis gain (Table 5: +37.1 R@1 on SPA with routing). No measurement of alignment rate or per-scene canonicalization is reported. If alignment is incomplete, the spatial gain may partly reflect the benchmark's own frame convention for spatial edits. Please report alignment statistics or evaluate after canonicalizing room orientations.
minor comments (4)
  1. [§5.1, Algorithm 1] The text says 'The hard-subset size is 20,' but Algorithm 1 defines K=|H ∪ T|, so the evaluation pool varies with the per-query positive count (mean 14.7, median 5). Clarify whether K denotes distractor count or total pool size.
  2. [Appendix B Phase 7] Phase 7 splits 8,834 validated queries into 4,963 released queries, but the criterion for discarding the remaining 3,871 is not described. State the filtering rule or note that only a stratified subsample is released.
  3. [Table 3] The WL-only baseline is not defined in the main text. Add a one-sentence description or a reference for the Weisfeiler–Lehman scene-graph kernel used.
  4. [Appendix E] The prompt templates use placeholders {spec}, {op}, and {paraphrase} without specifying the serialization of the room/operator. Release the exact serialization for reproducibility.

Circularity Check

2 steps flagged

Benchmark circularity: 3D-CER ground truth uses the same 3D-FUTURE attribute predicates and GPT-5.1 paraphrases that CR-Refiner's cost matrix and parser rely on; the +36.7 R@1 gain largely reflects inverting benchmark construction rather than human edit satisfaction.

specific steps
  1. self definitional [§4 / Appendix B Phase 5 vs §3.2 Eq. 1]
    "For each surviving query, we run the operator predicate over the same-room-type slice of the 23,381-room corpus to mine multi-positive ground truth. ... The category indicator ℓcat(q, oj)∈{0,1} is 0 iff cq and cj are string-equal. The style indicator ℓsty(q, oj)∈{0,0.5,1} takes value 0 when sq matches the candidate's style ..."

    The ground-truth positives are defined by an exact operator predicate over the 3D-FUTURE attribute vocabulary (category/style/material string equality on the named slot). The reranker's central score is built from mismatch indicators that are 0 exactly when the same fields are string-equal and 0.5 only for missing values. On the caption-NN hard subsets from the same room type, the candidate that flips the target attribute to the exact query string is almost always the label, so CR-Refiner's score is an approximate differentiable implementation of the benchmark's own label predicate. The reported +36.7 R@1 gain therefore measures how well the pipeline recovers the annotation rule, not whether a human judge would find the retrieved room satisfies the edit.

  2. other [§4 Phase 3 and §5.1]
    "Given each (source, operator) pair, we prompt GPT-5.1 to produce three short paraphrases per query ... The frozen LLM used for entity parsing, paraphrase rewriting in the CIReVL baseline, and the verifier is GPT-5.1-2025-11-13."

    The edit text is generated by GPT-5.1 from the same structured operator that defines ground truth, and the same GPT-5.1 model family is used to parse that text back into the query entity and to verify candidates. The parser therefore recovers the exact operator fields (category/style/material/axis) by construction, and the verifier re-checks candidates against wording produced by the same generator. The Gemini cross-check is also an LLM auditing the same operator-level predicate, so it does not break the alignment. This is not a mathematical tautology, but it removes the independent human-phrasing challenge that the task claims to test.

full rationale

The paper is internally well-supported: the ablation tables, bootstrap CIs, and honest full-corpus recall ceiling show that CR-Refiner genuinely re-ranks the curated 20-distractor pools better than the tested text baselines, and the full-corpus results are disclosed as a candidate-generation bottleneck rather than hidden. No self-citation chain or imported uniqueness theorem is load-bearing. However, the central headline gain is partially circular by construction: 3D-CER's multi-positive labels are mined by running the same programmatic operator predicates (exact string equality on 3D-FUTURE attributes) that CR-Refiner's Eq. 1 cost matrix encodes, and the paraphrases that define the edit text are written by GPT-5.1, the same model family used for parsing and verification. Thus the evaluation measures how well the pipeline inverts the benchmark's own label generator more than how well it satisfies independently expressed human edit intent. The paper's own limitation statement concedes that 'Both paraphrases and the multi-positive predicate operate at the operator level following the CIRR/CIRCO convention'; this is exactly the alignment that makes the benchmark an easy case for an attribute-string-matching reranker. The decisive missing evidence is human-annotated queries with independently phrased edits, which would break the generator-parser overlap. We therefore score 7: the result is not a formal tautology, but the main empirical claim partially reduces by construction to the benchmark's label-generating machinery.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The pipeline introduces no new physical entities. The 'query entity' q is a structured parse, not a postulated object. The 3D-CER dataset is a new artifact but not an invented entity in the physical sense. The listed free parameters and axioms are the load-bearing choices the central evaluation depends on.

free parameters (7)
  • Cost-matrix channel weights w_cat, w_sty, w_mat, w_dir = 0.4, 0.3, 0.2, 0.1
    Hand-set weights in Eq. 1; not listed among the 'five scalars' tuned on validation (§3.5).
  • Entropic regularization ε = 0.01
    Eq. 2; selected on validation split.
  • Unbalanced KL slack τ = 1.0
    Eq. 2; selected on validation split.
  • Structural-prior weight β_B = 0.3
    Eq. 3; selected on validation split.
  • LLM-verifier weight β_V = 0.5
    Eq. 3; selected on validation split.
  • Near/far distance thresholds in direction bonus = 1.5 m / 3.0 m
    Appendix F; hand-chosen thresholds not mentioned in the 'five scalars' count.
  • Size-class keyword vocabulary = hand-built map (small/medium/large)
    Appendix F; ad hoc vocabulary that the GEO bonus relies on.
axioms (4)
  • domain assumption Exact string equality on 3D-FUTURE attribute vocabulary is a faithful match for category/style/material and direction.
    Eq. 1 and Appendix F compute ℓ_cat, ℓ_sty, ℓ_mat, ℓ_dir with string equality/null rules; the benchmark's operator predicate uses the same vocabulary, so real-world synonym/paraphrase mismatch is not tested.
  • domain assumption 3D-FRONT room frames are sufficiently axis-aligned that sign of bounding-box-centre differences encodes 'left/right/front/behind'.
    §3.3: 'in practice this signal is sufficient on 3D-FRONT, where most rooms align to consistent global frame conventions'; also ℓ_dir weight is only 0.1 because frames are scene-specific.
  • ad hoc to paper The first object whose category equals c_q is the edit subject, and the first non-subject object whose keyword appears in the paraphrase is the anchor.
    Appendix F: subject/anchor selection is a first-match heuristic; no evidence it generalizes to rooms with multiple same-category objects or paraphrase rephrasings.
  • domain assumption GPT-5.1-generated paraphrases are representative of natural-language edit instructions.
    Benchmark construction (Phase 3) uses GPT-5.1 paraphrases and audits by the same/cross model; no human-annotation study is reported.

pith-pipeline@v1.3.0-alltime-deepseek · 16211 in / 14154 out tokens · 118158 ms · 2026-08-01T13:23:27.968063+00:00 · methodology

0 comments
read the original abstract

Edit-conditioned 3D scene retrieval pairs a reference 3D room with a natural-language modification and retrieves rooms from a corpus that satisfy the edit. Three lines of prior work each fall short on this task. 2D composed image retrieval reasons over pixel-level edits and has no primitive for 3D object sets. 3D foundation encoders embed individual objects but cannot compose at the scene level. 3D scene-grounding methods localize references inside a static scene rather than rank modified rooms across a corpus. We present CR-Refiner, a training-free reranker that wraps any base retriever's top-K candidates with three components. A frozen LLM parses the edit into a structured query entity, and each candidate is scored by an unbalanced optimal-transport problem over a 1xG cost matrix coupling category, style, material, and geometry. The unbalanced solver lets the single-entity query drop mass on irrelevant objects, modelling the asymmetry directly. An axis-conditional structural prior adds size-keyword cues for geometric edits and subject-anchor direction cues for spatial edits. An LLM verifier refines the top three candidates with continuous confidence. Because no benchmark evaluates compositional matching over 3D object sets, we additionally release 3D-CER, 4,963 edit-conditioned queries over a 23,381-room indoor corpus across five edit axes, with multi-positive ground truth, CIRR-style hard subsets, and zero-target adversarials. Across three qualitatively distinct base retrievers, CR-Refiner consistently improves hard-subset R@1 and mAP@10 on every edit axis.

Figures

Figures reproduced from arXiv: 2607.19115 by Hao Wang, Hao Wu, Heyi Lin, Hui Xiong, Jinjing Zhu, Nanyu Wu, Qianyi Cai.

Figure 1
Figure 1. Figure 1: Overview of CR-Refiner. Given a reference 3D room R and a natural-language edit t, a base retriever returns a top-K candidate pool P while a frozen LLM parses (R, t) into a structured query entity q. CR-Refiner rescores each candidate via a Coupled-OT Backbone, an axis-conditional Structural Prior, and an LLM Verifier applied to the top-3 candidates after the backbone and prior; the three signals are combi… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison on 3D-CER hard subsets. Source room (column 1) and top-1 retrievals from CR-Refiner (column 2, green border), Caption-MPNet (column 3, red border), and Pic2Word (column 4, red border). The edit axis is labelled at the left of each row; the edit instruction appears in italics below. Refiner a top-K from this corpus produces a modest lift at small K that disappears at large K as the st… view at source ↗
Figure 3
Figure 3. Figure 3: Full-corpus recall ceiling on the 23,381-room corpus (n=200). Recall@K for the three base retrievers across K ∈ {1, 5, 10, 50, 100, 500, 1000}, with 95% bootstrap confidence intervals shaded. The dotted line marks the hard-subset R@1 regime (≈ 0.27) reported in Section 5.2, illustrating how far full-corpus recall lags the curated-pool numbers at small K. Visualisation of [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 4
Figure 4. Figure 4: Pipeline R@1 as a function of reranking pool size K (n=200). Solid lines are the end￾to-end pipeline R@1 when CR-Refiner reranks the base retriever’s top-K from the full 23K-room corpus; dashed lines are the corresponding base retriever’s full-corpus top-1 for reference. The lift over the base retriever survives at K≤100 and dies beyond K≈100 as the structural prior dilutes against an increasingly noisy po… view at source ↗
Figure 5
Figure 5. Figure 5: visualises the drop-one ablation in [PITH_FULL_IMAGE:figures/full_fig_p017_5.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

26 extracted references · 12 linked inside Pith

  1. [1]

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

    Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. InEuropean conference on computer vision, pages 202–221. Springer, 2020

    2020

  2. [2]

    Multi3drefer: Grounding text description to multiple 3d objects

    Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225–15236, 2023

    2023

  3. [3]

    Image retrieval on real-life images with pre-trained vision-and-language models

    Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. InProceedings of the IEEE/CVF international conference on computer vision, pages 2125–2134, 2021

    2021

  4. [4]

    Zero-shot composed image retrieval with textual inversion

    Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. InProceedings of the IEEE/CVF international conference on computer vision, pages 15338–15347, 2023

    2023

  5. [5]

    Composing text and image for image retrieval - an empirical odyssey, 2018

    Nam V o, Lu Jiang, Chen Sun, Kevin Murphy, Li-Jia Li, Li Fei-Fei, and James Hays. Composing text and image for image retrieval - an empirical odyssey, 2018. URL https://arxiv.org/ abs/1812.07119

    Pith/arXiv arXiv 2018

  6. [6]

    Effective conditioned and composed image retrieval combining clip-based features

    Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21466–21474, 2022

    2022

  7. [7]

    Vision-by- language for training-free compositional image retrieval.arXiv preprint arXiv:2310.09291, 2023

    Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by- language for training-free compositional image retrieval.arXiv preprint arXiv:2310.09291, 2023

    Pith/arXiv arXiv 2023

  8. [8]

    Pic2word: Mapping pictures to words for zero-shot composed image retrieval

    Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19305–19314, 2023

    2023

  9. [9]

    Compodiff: Versatile composed image retrieval with latent diffusion, 2024

    Geonmo Gu, Sanghyuk Chun, Wonjae Kim, HeeJae Jun, Yoohoon Kang, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion, 2024. URL https: //openreview.net/forum?id=0NruoU6s5Z

    2024

  10. [10]

    Language-only efficient training of zero-shot composed image retrieval, 2024

    Geonmo Gu, Sanghyuk Chun, Wonjae Kim, Yoohoon Kang, and Sangdoo Yun. Language-only efficient training of zero-shot composed image retrieval, 2024. URL https://arxiv.org/ abs/2312.01998

    Pith/arXiv arXiv 2024

  11. [11]

    Pointclip: Point cloud understanding by clip, 2021

    Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip, 2021. URL https: //arxiv.org/abs/2112.02413

    Pith/arXiv arXiv 2021

  12. [12]

    Openshape: Scaling up 3d shape representation towards open-world understanding.Advances in neural information processing systems, 36:44860–44879, 2023

    Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding.Advances in neural information processing systems, 36:44860–44879, 2023

    2023

  13. [13]

    Uni3d: Exploring unified 3d representation at scale.arXiv preprint arXiv:2310.06773, 2023

    Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale.arXiv preprint arXiv:2310.06773, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Ulip-2: Towards scalable multimodal pre-training for 3d understanding

    Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27091–27101, 2024

    2024

  15. [15]

    3d-vista: Pre-trained transformer for 3d vision and text alignment, 2023

    Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment, 2023. URL https://arxiv.org/ abs/2308.04352. 10

    Pith/arXiv arXiv 2023

  16. [16]

    3d-llm: Injecting the 3d world into large language models, 2023

    Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models, 2023. URL https: //arxiv.org/abs/2307.12981

    Pith/arXiv arXiv 2023

  17. [17]

    Pointllm: Empowering large language models to understand point clouds, 2024

    Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds, 2024. URL https://arxiv. org/abs/2308.16911

    Pith/arXiv arXiv 2024

  18. [18]

    Scenegraphloc: Cross-modal coarse visual localization on 3d scene graphs, 2024

    Yang Miao, Francis Engelmann, Olga Vysotska, Federico Tombari, Marc Pollefeys, and Dániel Béla Baráth. Scenegraphloc: Cross-modal coarse visual localization on 3d scene graphs, 2024. URLhttps://arxiv.org/abs/2404.00469

    Pith/arXiv arXiv 2024

  19. [19]

    Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes

    Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors,Computer Vision – ECCV 2020, pages 422–440, Cham, 2020. Springer International Publishing. ISBN 978-3...

    2020

  20. [20]

    Scanqa: 3d question answering for spatial scene understanding, 2022

    Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding, 2022. URL https://arxiv.org/abs/2112. 10482

    2022

  21. [21]

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

    Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. InEuropean Conference on Computer Vision, pages 289–310. Springer, 2024

    2024

  22. [22]

    Shapetalk: A language dataset and framework for 3d shape edits and deformations

    Panos Achlioptas, Ian Huang, Minhyuk Sung, Sergey Tulyakov, and Leonidas Guibas. Shapetalk: A language dataset and framework for 3d shape edits and deformations. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12685–12694, 2023

    2023

  23. [23]

    Diffuscene: Denoising diffusion models for generative indoor scene synthesis, 2024

    Jiapeng Tang, Yinyu Nie, Lev Markhasin, Angela Dai, Justus Thies, and Matthias Nießner. Diffuscene: Denoising diffusion models for generative indoor scene synthesis, 2024. URL https://arxiv.org/abs/2303.14207

    Pith/arXiv arXiv 2024

  24. [24]

    Instructscene: Instruction-driven 3d indoor scene synthesis with semantic graph prior, 2024

    Chenguo Lin and Yadong Mu. Instructscene: Instruction-driven 3d indoor scene synthesis with semantic graph prior, 2024. URLhttps://arxiv.org/abs/2402.04717

    Pith/arXiv arXiv 2024

  25. [25]

    Editroom: Llm-parameterized graph diffusion for composable 3d room layout editing, 2025

    Kaizhi Zheng, Xiaotong Chen, Xuehai He, Jing Gu, Linjie Li, Zhengyuan Yang, Kevin Lin, Jianfeng Wang, Lijuan Wang, and Xin Eric Wang. Editroom: Llm-parameterized graph diffusion for composable 3d room layout editing, 2025. URL https://arxiv.org/abs/2410.12836

    Pith/arXiv arXiv 2025

  26. [26]

    in roughly the same layout

    Lenaic Chizat, Gabriel Peyré, Bernhard Schmitzer, and François-Xavier Vialard. Scaling algorithms for unbalanced optimal transport problems.Mathematics of computation, 87(314): 2563–2609, 2018. A Pseudocode of CR-Refiner Algorithm 1 summarises the full inference procedure of CR-Refiner. Each step references the corresponding component of the main text. Li...

    arXiv 2018