pith. sign in

arxiv: 2607.01503 · v1 · pith:OUKH3PDYnew · submitted 2026-07-01 · 💻 cs.CV

Disentangling Pictorial Cue Understanding from Language Bias in VLMs via Depth Ordering Task

Yiqian Liu , Iuliia Kotseruba , John K. Tsotsos This is my paper

Pith reviewed 2026-07-03 20:49 UTC · model grok-4.3

classification 💻 cs.CV
keywords depth perceptionvision-language modelspictorial depth cueslanguage biasodd-one-out taskdepth orderingO3-D dataset
0
0 comments X

The pith

Vision-language models achieve only chance-level depth ordering because they under-use pictorial cues and rely on language biases instead.

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

The paper tests whether VLMs can perceive depth from images by isolating pictorial cues such as occlusion and perspective while varying the clarity of language instructions. It creates controlled stimuli from 3D scenes, builds the O3-D dataset of 37K images and 147K question pairs, and evaluates 12 models. Every model scores between 47 and 56 percent, none above chance, and a new metric shows the answers are driven by linguistic patterns rather than visual depth. The same low performance persists even when chain-of-thought or in-context examples are added.

Core claim

The paper claims that current VLMs do not extract depth order from pictorial cues. When one object is placed at a different depth among otherwise identical objects, models cannot reliably say whether the odd-one-out target is closer or farther; accuracies stay at or below chance. A metric that varies referring-expression clarity shows the decisions track language statistics far more than the controlled visual cues. The authors conclude that static image training alone leaves models without usable depth understanding.

What carries the argument

The Odd-One-Out Depth (O3-D) dataset and the vision-versus-language sensitivity metric, built by generating 2D views while toggling individual pictorial depth cues and referring-expression clarity.

If this is right

  • Pictorial depth cues remain unused by current VLMs even when the task is reduced to simple closer-farther judgments.
  • Chain-of-thought and in-context learning produce no meaningful gain on depth ordering once language bias is controlled.
  • Static image pre-training is insufficient to produce genuine depth ordering ability in VLMs.
  • Fine-grained cue control is required to separate visual understanding from language shortcuts in future evaluations.

Where Pith is reading between the lines

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

  • The same cue-isolation method could be applied to other perceptual judgments such as relative size or surface slant.
  • Models that pass this test might still fail when depth cues must be integrated across multiple objects or changing viewpoints.
  • Training regimes that add explicit depth supervision or 3D reconstruction losses could be tested directly against the O3-D baseline.

Load-bearing premise

The synthetic and real-scene stimuli isolate each pictorial depth cue without adding unintended correlations that models could solve by shortcut instead of depth perception.

What would settle it

Any model that reaches clearly above-chance accuracy on O3-D trials where language cues are removed or randomized would show that the under-utilization claim does not hold.

Figures

Figures reproduced from arXiv: 2607.01503 by Iuliia Kotseruba, John K. Tsotsos, Yiqian Liu.

Figure 1
Figure 1. Figure 1: O3-D probes VLM depth and language understanding. We start by construct￾ing synthetic and real 3D scenes with diverse backgrounds and objects. Each scene contains 5 objects of the same class, one of which (the target) is of different size and placed at a different depth plane. We then generate a number of 2D views with one or two depth cues by controlling the camera, light position, etc. For each image, we… view at source ↗
Figure 2
Figure 2. Figure 2: Each 3D scene in O3-D contains 1 target and 4 distractors, where the target is larger (or smaller) and located on a different depth plane (a). When a camera is placed at a certain position, (b) the target appears at the same depth as the distractors. (c) Gray arrows between the images indicate how disambiguated 2D views are generated from the base view (in the center) by adding one or two pictorial depth c… view at source ↗
Figure 3
Figure 3. Figure 3: Image post-processing. (a) Resize a real image to 1024 × 683. (b) Optionally add markers for easier referring (Tab. 3 bottom row), to gauge the effects of referring expression comprehension (REC) on depth ordering responses [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance summary of VLMs on depth ordering VQA. Both depth under￾standing (y-axis, see Sec. 5.1) and language consistency (x-axis, see Sec. 5.2) can be probed using our O3-D dataset. Depth ordering accuracies of VLMs are close to ran￾dom guess and inferior to DepthAnythingV2 [57] baseline. VLMs’ language consistency has a wide spread. understanding, not the referring comprehension. As an example, a demo… view at source ↗
Figure 5
Figure 5. Figure 5: A combined heatmap of 1-cue and 2-cue mean accuracies of all tested VLMs (bottom-left), vs. baseline (top-right). Red- and blue-tinted cells indicate performance above and below chance level (0.5). The two main diagonal cells (within green dotted rectangle) show accuracies for 1-cue depth ordering, whereas the other cells report 2- cue interactions. The depth ordering performance is better whenever HP or R… view at source ↗
Figure 6
Figure 6. Figure 6: Near-far bias. Both plots share the y-axis. Left: Boxes show variations across cues. VLMs exhibit different extents of biases toward answering near vs. far, compared with human reference [10]. Right: None of the VLMs has consistent bias reduction as cue strength increases, compared with DepthAnythingV2. not significant across cues. The ICL and CoT prompting is most effective for the RS cue (δ = 0.069) and … view at source ↗
Figure 7
Figure 7. Figure 7: Vision vs. language influence on depth ordering VQA. Bars show the standard deviations of mean accuracy (SDGM, see text), as a measure of influence. Language influence is uniformly larger than vision. and language. The high language sensitivity of the VLMs is consistent with results in [9]. Inconsistencies are mainly caused by varying depth order vocabu￾lary and yes-no/multi-choice query. Depth order vocab… view at source ↗
Figure 1
Figure 1. Figure 1: illustrates sampled images for the base view and 9 controlled pictorial cues: Occlusion (OC), Relative Size (RS), Light-and-Shadow (LS), Texture Gra￾dient (TG), Linear Perspective (LP), Height-in-Plane (HP), Familiar Size (FS), Saturation (SA), and Focusness (FO). Images with no cues (leftmost in the 1st and 4th row) and two cue combinations are also included. None OC RS LS TG LP* Real HP FS FO HP, LS HP, … view at source ↗
Figure 2
Figure 2. Figure 2: shows the cropping of real-world images specifically for depth ordering task. Although after cropping the image is no longer odd-one-out, the cropping generates additional images for depth ordering VQA. One such cropped image is used to test both near and far cases, depending which one of the two objects is referred to. (a) Original (b) Cropped [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Word cloud generated from the collection of regular prompts, without ICL and CoT prompting. 5 VLM Versions Tab. 5 summarizes the Hugging Face model versions for the evaluated models in the paper, except GPT4.1-mini and Gemini2.5 which are accessed via OpenAI and Google API, respectively [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: shows the performance gap between evaluated VLMs and the DepthAny￾thingV2 baseline, for single-cue cases and two-cue interactions. Qwen2-VL GPT4.1-mini Gemini2.5 InternVL2.5 PaliGemma2 Phi3.5 DeepSeek-VL LLaVA1.5 Cambrian VILA1.5 BLIP2 Kosmos2 0.0 0.2 0.4 0.6 0.8 Depth ordering accuracy One cue Qwen2-VL GPT4.1-mini Gemini2.5 InternVL2.5 PaliGemma2 Phi3.5 DeepSeek-VL LLaVA1.5 Cambrian VILA1.5 BLIP2 Kosmos2 … view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of depth ordering accuracies as cue strengths double, for each cue and each VLM. Grey dashed lines are chance level (0.5). The double cue strength is not applicable to the Familiar Size (FS) cue [PITH_FULL_IMAGE:figures/full_fig_p026_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: shows Spearman’s rank correlations among the 8 single-cue depth ordering accuracies. The Texture Gradient/Familiar Size (0.71) correlates the most among all pairs, while the Relative Size/Focusness pair (-0.38) has the lowest correlation. The correlations are consistently lower than those in DepthCues [15]. FO FS HP LS OC RS SA Familiar Size (FS) Height-in-Plane (HP) Light-and-Shadow (LS) Occlusion (OC) Re… view at source ↗
read the original abstract

In this paper, we study depth perception of vision-language models (VLMs) to isolate the effects of pictorial depth cues and disentangle vision and language influences on model performance. To this end, we combine depth-ordering and odd-one-out psychophysical tasks: the VLMs are presented with images where one object is at different depth relative to other, otherwise identical, objects, and must determine whether the odd-one-out target is closer or farther to the observer. To create stimuli, we generate 2D views from simulated and real 3D scenes while controlling the presence of individual pictorial depth cues, enabling a fine-grained analysis of cue-level contributions. Language effects are examined by varying referring expression clarity. We also introduce a novel metric to quantify vision-vs-language sensitivities. Applying this methodology, we create the Odd-One-Out Depth (O3-D) dataset with 37K real and synthetic images and 147K image-question pairs. Evaluation of 12 open-source and commercial models on O3-D shows under-utilization of depth cues and depth-ordering accuracies between 47% and 56%, with no model above chance level. At the same time, our metric reveals strong linguistic bias in the answers. Neither chain-of-thought (CoT) nor in-context learning (ICL) significantly improves performance, suggesting that static image data alone may be insufficient for depth understanding. All code, the image generation pipeline, and the O3-D dataset are publicly released at https://github.com/lyiqian/o3-d.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces the Odd-One-Out Depth (O3-D) dataset of 37K real and synthetic images and 147K image-question pairs, generated by rendering 2D views from 3D scenes while controlling the presence of individual pictorial depth cues. It combines depth-ordering and odd-one-out tasks to evaluate 12 VLMs, reporting accuracies of 47-56% (no model above chance) and introducing a novel vision-vs-language sensitivity metric that reveals strong linguistic bias. Neither CoT nor ICL improves results, and the authors publicly release the dataset, generation pipeline, and code.

Significance. If the cue controls are shown to isolate depth perception without confounds, the results would provide concrete evidence that current VLMs fail to leverage pictorial depth cues and default to language biases, with direct implications for multimodal model design. The public release of the full pipeline and dataset is a clear strength that supports reproducibility.

major comments (3)
  1. [Abstract and §3] Abstract and §3 (Dataset): The headline finding that models exhibit under-utilization of depth cues (accuracies 47-56%, none above chance) requires that the 37K images cannot be solved via non-depth regularities. The description of controlling individual cues from simulated and real 3D scenes provides no quantitative verification (e.g., correlation analysis or control ablations) that rendering artifacts, texture, or lighting do not introduce exploitable patterns.
  2. [§4] §4 (Evaluation) and metric definition: The vision-vs-language sensitivity metric is used to attribute failures to linguistic bias, yet the manuscript does not report how the metric is computed from the varying referring-expression conditions or demonstrate that it is robust to question phrasing variations that could independently affect performance.
  3. [§4] §4 (Results): The reported depth-ordering accuracies are given as ranges (47-56%) without per-model breakdowns, error bars, or statistical comparison to chance level, making it impossible to assess whether the 'no model above chance' claim holds after accounting for variance across the 147K pairs.
minor comments (2)
  1. [Abstract] The abstract states that 'static image data alone may be insufficient,' but this conclusion would be strengthened by explicit reference to the specific CoT/ICL results rather than appearing as a general claim.
  2. [Figures/Tables] Figure captions and table headers could more clearly distinguish real vs. synthetic image subsets when reporting cue-level contributions.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment point by point below, indicating planned revisions where appropriate to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (Dataset): The headline finding that models exhibit under-utilization of depth cues (accuracies 47-56%, none above chance) requires that the 37K images cannot be solved via non-depth regularities. The description of controlling individual cues from simulated and real 3D scenes provides no quantitative verification (e.g., correlation analysis or control ablations) that rendering artifacts, texture, or lighting do not introduce exploitable patterns.

    Authors: We agree that explicit quantitative verification would further support the isolation of depth cues. The generation pipeline randomizes lighting, textures, camera poses, and object properties across the 3D scenes to minimize systematic artifacts, and real-world captures use varied environments. In the revision we will add correlation analyses between non-depth image statistics (e.g., texture histograms, lighting gradients) and ground-truth labels, plus control ablations on subsets with fixed cues, to demonstrate absence of exploitable regularities. revision: yes

  2. Referee: [§4] §4 (Evaluation) and metric definition: The vision-vs-language sensitivity metric is used to attribute failures to linguistic bias, yet the manuscript does not report how the metric is computed from the varying referring-expression conditions or demonstrate that it is robust to question phrasing variations that could independently affect performance.

    Authors: The metric is defined in §4 as the differential performance between clear and ambiguous referring-expression conditions on the same images. We will expand the description with the exact formula and add a robustness check across two additional phrasings per condition to confirm the bias attribution is not phrasing-dependent. revision: partial

  3. Referee: [§4] §4 (Results): The reported depth-ordering accuracies are given as ranges (47-56%) without per-model breakdowns, error bars, or statistical comparison to chance level, making it impossible to assess whether the 'no model above chance' claim holds after accounting for variance across the 147K pairs.

    Authors: We will replace the aggregate range with a table listing per-model accuracies, standard errors computed over the 147K pairs (or per-image subsets), and one-sample t-test p-values against 50% chance level. This will be included in the revised §4. revision: yes

Circularity Check

0 steps flagged

Empirical evaluation study; no derivations or predictions reduce to fitted parameters or self-citations.

full rationale

This is an empirical evaluation paper that generates a controlled dataset (O3-D) from 3D scenes, tests 12 VLMs on depth-ordering tasks, and introduces a metric for vision-vs-language bias. No equations, fitted parameters, or predictions are claimed; the central results (47-56% accuracies, linguistic bias) follow directly from model outputs on the released stimuli. No self-citation load-bearing steps, uniqueness theorems, or ansatzes are invoked. The analysis is self-contained against external model benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claims rest on the assumption that depth cues can be independently controlled in image generation and that the odd-one-out task isolates depth understanding from other factors.

axioms (1)
  • domain assumption Pictorial depth cues can be selectively enabled or disabled in rendered images without introducing new unintended correlations
    Invoked in the stimulus creation process described in the abstract
invented entities (1)
  • vision-vs-language sensitivity metric no independent evidence
    purpose: Quantify the relative influence of visual cues versus language on model answers
    Newly introduced in the abstract to analyze results

pith-pipeline@v0.9.1-grok · 5819 in / 1206 out tokens · 18277 ms · 2026-07-03T20:49:49.873715+00:00 · methodology

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

60 extracted references · 6 canonical work pages · 5 internal anchors

  1. [1]

    Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone

    Abdin, M., et al.: Phi-3 technical report: A highly capable language model locally on your phone. arXiv:2404.14219 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  2. [2]

    IEEE TPAMI46(4), 2396–2414 (2023)

    Arampatzakis, V., Pavlidis, G., Mitianoudis, N., Papamarkos, N.: Monocular depth estimation: A thorough review. IEEE TPAMI46(4), 2396–2414 (2023)

    2023

  3. [3]

    In: CVPRW (2025)

    Azad, S., Jain, Y., Garg, R., Rawat, Y., Vineet, V.: Understanding depth and height perception in large visual-language models. In: CVPRW (2025)

    2025

  4. [4]

    In: CVPR (2022)

    Azuma, D., Miyanishi, T., Kurita, S., Kawanabe, M.: ScanQA: 3D Question An- swering for Spatial Scene Understanding. In: CVPR (2022)

    2022

  5. [5]

    Baldassini, F.B., Shukor, M., Cord, M., Soulier, L., Piwowarski, B.: What makes multimodal in-context learning work? In: CVPRW. pp. 1539–1550 (2024)

    2024

  6. [6]

    In: Stevens’ Handbook of Experi- mental Psychology and Cognitive Neuroscience, pp

    Brenner, E., Smeets, J.B.J.: Depth perception. In: Stevens’ Handbook of Experi- mental Psychology and Cognitive Neuroscience, pp. 1–30. John Wiley & Sons, Ltd (2018)

    2018

  7. [7]

    In: CVPR (2024)

    Chen, B., Xu, Z., Kirmani, S., Ichter, B., Sadigh, D., Guibas, L., Xia, F.: Spa- tialVLM: Endowing Vision-Language Models with Spatial Reasoning Capabilities. In: CVPR (2024)

    2024

  8. [8]

    In: ECCV (2020) 16 Y

    Chen, D.Z., Chang, A.X., Nießner, M.: ScanRefer: 3D Object Localization in RGB- D Scans Using Natural Language. In: ECCV (2020) 16 Y. Liu et al

    2020

  9. [9]

    In: NeurIPS (2023)

    Chen, S., Gu, J., Han, Z., Ma, Y., Torr, P., Tresp, V.: Benchmarking Robustness of Adaptation Methods on Pre-trained Vision-Language Models. In: NeurIPS (2023)

    2023

  10. [10]

    In: NeurIPS (2016)

    Chen, W., Fu, Z., Yang, D., Deng, J.: Single-image depth perception in the wild. In: NeurIPS (2016)

    2016

  11. [11]

    Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling

    Chen, Z., et al.: Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling. arXiv:2412.05271 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    In: NeurIPS (2024)

    Cheng, A.C., Yin, H., Fu, Y., Guo, Q., Yang, R., Kautz, J., Wang, X., Liu, S.: Spa- tialRGPT: Grounded Spatial Reasoning in Vision-Language Models. In: NeurIPS (2024)

    2024

  13. [13]

    In: ICLR (2025)

    Chow, W., Mao, J., Li, B., Seita, D., Guizilini, V., Wang, Y.: Physbench: Bench- marking and enhancing vision-language models for physical world understanding. In: ICLR (2025)

    2025

  14. [14]

    In: ICCV (2025)

    Dahou, Y., Huynh, N.D., Le-Khac, P.H., Para, W.R., Singh, A., Narayan, S.: Sal- Bench: Vision-language models can’t see the obvious. In: ICCV (2025)

    2025

  15. [15]

    In: CVPR (2025)

    Danier, D., Aygün, M., Li, C., Bilen, H., Mac Aodha, O.: DepthCues: Evaluating Monocular Depth Perception in Large Vision Models. In: CVPR (2025)

    2025

  16. [16]

    Deng, A., Cao, T., Chen, Z., Hooi, B.: Words or vision: Do vision-language models have blind faith in text? In: CVPR (2025)

    2025

  17. [17]

    In: CVPR (2024)

    El Banani, M., Raj, A., Maninis, K.K., Kar, A., Li, Y., Rubinstein, M., Sun, D., Guibas,L.,Johnson,J.,Jampani,V.:Probingthe3dawarenessofvisualfoundation models. In: CVPR (2024)

    2024

  18. [18]

    In: Yarowsky, D., Baldwin, T., Korho- nen, A., Livescu, K., Bethard, S

    FitzGerald, N., Artzi, Y., Zettlemoyer, L.: Learning Distributions over Logical Forms for Referring Expression Generation. In: Yarowsky, D., Baldwin, T., Korho- nen, A., Livescu, K., Bethard, S. (eds.) Proceedings of the Conference on Empirical Methods in Natural Language Processing (2013)

    2013

  19. [19]

    In: ECCV (2025)

    Fu, X., Hu, Y., Li, B., Feng, Y., Wang, H., Lin, X., Roth, D., Smith, N.A., Ma, W.C., Krishna, R.: BLINK: Multimodal Large Language Models Can See but Not Perceive. In: ECCV (2025)

    2025

  20. [20]

    In: CVPR (2017)

    Goyal, Y., Khot, T., Summers-Stay, D., Batra, D., Parikh, D.: Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question An- swering. In: CVPR (2017)

    2017

  21. [21]

    In: CVPR (2022)

    Greff, K., Belletti, F., Beyer, L., Doersch, C., Du, Y., Duckworth, D., Fleet, D.J., Gnanapragasam, D., Golemo, F., Herrmann, C., Kipf, T., Kundu, A., Lagun, D., Laradji, I., Liu, H.T.D., Meyer, H., Miao, Y., Nowrouzezahrai, D., Oztireli, C., Pot, E., Radwan, N., Rebain, D., Sabour, S., Sajjadi, M.S.M., Sela, M., Sitzmann, V., Stone, A., Sun, D., Vora, S....

    2022

  22. [22]

    In: CVPR (2009)

    He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. In: CVPR (2009)

    2009

  23. [23]

    Infant Behavior and Development 35(1), 109–128 (2012)

    Kavšek, M., Yonas, A., Granrud, C.E.: Infants’ sensitivity to pictorial depth cues: A review and meta-analysis of looking studies. Infant Behavior and Development 35(1), 109–128 (2012)

    2012

  24. [24]

    ACM Computing Surveys57(10), 1–35 (2025)

    Kim, B.S., Kim, J., Lee, D., Jang, B.: Visual Question Answering: A Survey of Methods, Datasets, Evaluation, and Challenges. ACM Computing Surveys57(10), 1–35 (2025)

    2025

  25. [25]

    In: BMVC (2019)

    Kotseruba, I., Wloka, C., Rasouli, A., Tsotsos, J.K.: Do Saliency Models Detect Odd-One-Out Targets? New Datasets and Evaluations. In: BMVC (2019)

    2019

  26. [26]

    In: CVPR (2024)

    Li, B., Ge, Y., Ge, Y., Wang, G., Wang, R., Zhang, R., Shan, Y.: SEED-Bench: Benchmarking Multimodal Large Language Models. In: CVPR (2024)

    2024

  27. [27]

    Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping Language-Image Pre- trainingwithFrozenImageEncodersandLargeLanguageModels.In:ICML(2023) Disentangling Pictorial Cue Understanding from Language Bias 17

    2023

  28. [28]

    In: CVPR (2024)

    Lin, J., Yin, H., Ping, W., Molchanov, P., Shoeybi, M., Han, S.: VILA: On Pre- training for Visual Language Models. In: CVPR (2024)

    2024

  29. [29]

    In: ICLR (2025)

    Linsley, D., Zhou, P., Ashok, A.K., Nagaraj, A., Gaonkar, G., Lewis, F.E., Pizlo, Z., Serre, T.: The 3D-PC: A benchmark for visual perspective taking in humans and machines. In: ICLR (2025)

    2025

  30. [30]

    In: CVPR (2023)

    Liu, C., Ding, H., Jiang, X.: GRES: Generalized Referring Expression Segmenta- tion. In: CVPR (2023)

    2023

  31. [31]

    In: CVPR (2024)

    Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved Baselines with Visual Instruction Tun- ing. In: CVPR (2024)

    2024

  32. [32]

    Liu, Y., Duan, H., Zhang, Y., Li, B., Zhang, S., Zhao, W., Yuan, Y., Wang, J., He, C., Liu, Z., Chen, K., Lin, D.: MMBench: Is Your Multi-modal Model an All- Around Player? In: ECCV (2025)

    2025

  33. [33]

    DeepSeek-VL: Towards Real-World Vision-Language Understanding

    Lu, H., Liu, W., Zhang, B., Wang, B., Dong, K., Liu, B., Sun, J., Ren, T., Li, Z., Yang, H., Sun, Y., Deng, C., Xu, H., Xie, Z., Ruan, C.: DeepSeek-VL: Towards real-world vision-language understanding. arXiv:2403.05525 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  34. [34]

    In: ICLR (2023)

    Ma, X., Yong, S., Zheng, Z., Li, Q., Liang, Y., Zhu, S.C., Huang, S.: SQA3D: Situated Question Answering in 3D Scenes. In: ICLR (2023)

    2023

  35. [35]

    In: ICCV (2024)

    Man, Y., Zheng, S., Bao, Z., Hebert, M., Gui, L.Y., Wang, Y.X.: Lexicon3D: Prob- ing Visual Foundation Models for Complex 3D Scene Understanding. In: ICCV (2024)

    2024

  36. [36]

    In: CVPR (2016)

    Mao, J., Huang, J., Toshev, A., Camburu, O., Yuille, A.L., Murphy, K.: Generation and Comprehension of Unambiguous Object Descriptions. In: CVPR (2016)

    2016

  37. [37]

    Marshall, J.A., Burbeck, C.A., Ariely, D., Rolland, J.P., Martin, K.E.: Occlusion edge blur: a cue to relative visual depth. J. Opt. Soc. Am. A13(4), 681–688 (Apr 1996)

    1996

  38. [38]

    IEEE TPAMI44(7), 3523–3542 (2021)

    Minaee, S., Boykov, Y., Porikli, F., Plaza, A., Kehtarnavaz, N., Terzopoulos, D.: Image segmentation using deep learning: A survey. IEEE TPAMI44(7), 3523–3542 (2021)

    2021

  39. [39]

    In: CVPR (2021)

    Niu, Y., Tang, K., Zhang, H., Lu, Z., Hua, X.S., Wen, J.R.: Counterfactual VQA: A Cause-Effect Look at Language Bias. In: CVPR (2021)

    2021

  40. [40]

    In: ICLR (2024)

    Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos-2: Grounding Multimodal Large Language Models to the World. In: ICLR (2024)

    2024

  41. [41]

    In: Findings of the Association for Computa- tional Linguistics: NAACL

    Pezeshkpour, P., Hruschka, E.: Large language models sensitivity to the order of options in multiple-choice questions. In: Findings of the Association for Computa- tional Linguistics: NAACL. pp. 2006–2017 (2024)

    2006

  42. [42]

    IEEE TMM23, 4426–4440 (2021)

    Qiao, Y., Deng, C., Wu, Q.: Referring expression comprehension: A survey of meth- ods and datasets. IEEE TMM23, 4426–4440 (2021)

    2021

  43. [43]

    In: NeurIPS (2024)

    Qin, L., Chen, Q., Fei, H., Chen, Z., Li, M., Che, W.: What Factors Affect Multi- Modal In-Context Learning? An In-Depth Exploration. In: NeurIPS (2024)

    2024

  44. [44]

    In: Three-Dimensional Imaging, Visualization, and Display

    Reichelt, S., Häussler, R., Fütterer, G., Leister, N.: Depth cues in human visual perception and their realization in 3D displays. In: Three-Dimensional Imaging, Visualization, and Display. vol. 7690, pp. 92–103. SPIE (2010)

    2010

  45. [45]

    Arti- ficial Intelligence Review56(9), 9175–9219 (2023)

    Samavati, T., Soryani, M.: Deep learning-based 3d reconstruction: A survey. Arti- ficial Intelligence Review56(9), 9175–9219 (2023)

    2023

  46. [46]

    In: CVPR (June 2019)

    Shah, M., Chen, X., Rohrbach, M., Parikh, D.: Cycle-consistency for robust visual question answering. In: CVPR (June 2019)

    2019

  47. [47]

    In: International Conference on Development and Learning

    Sinapov, J., Stoytchev, A.: The odd one out task: Toward an intelligence test for robots. In: International Conference on Development and Learning. pp. 126–131 (2010) 18 Y. Liu et al

    2010

  48. [48]

    PaliGemma 2: A Family of Versatile VLMs for Transfer

    Steiner, A., Pinto, A.S., Tschannen, M., Keysers, D., Wang, X., Bitton, Y., Grit- senko, A., Minderer, M., Sherbondy, A., Long, S., Qin, S., Ingle, R., Bugliarello, E., Kazemzadeh, S., Mesnard, T., Alabdulmohsin, I., Beyer, L., Zhai, X.: PaliGemma 2: A Family of Versatile VLMs for Transfer. arXiv:2412.03555 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  49. [49]

    Springer Nature (2022)

    Szeliski, R.: Computer Vision: Algorithms and Applications. Springer Nature (2022)

    2022

  50. [50]

    In: NeurIPS (2024)

    Tong, S., Brown, E., Wu, P., Woo, S., Middepogu, M., Akula, S.C., Yang, J., Yang, S., Iyer, A., Pan, X., Wang, A., Fergus, R., LeCun, Y., Xie, S.: Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs. In: NeurIPS (2024)

    2024

  51. [51]

    In: CVPR (2024)

    Tong, S., Liu, Z., Zhai, Y., Ma, Y., LeCun, Y., Xie, S.: Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs. In: CVPR (2024)

    2024

  52. [52]

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution

    Wang, P., Bai, S., Tan, S., Wang, S., Fan, Z., Bai, J., Chen, K., Liu, X., Wang, J., Ge, W., Fan, Y., Dang, K., Du, M., Ren, X., Men, R., Liu, D., Zhou, C., Zhou, J., Lin, J.: Qwen2-VL: Enhancing Vision-Language Model’s Perception of the World at Any Resolution. arXiv:2409.12191 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  53. [53]

    Journal of Vision5(10), 7 (2005)

    Watt,S.J.,Akeley,K.,Ernst,M.O.,Banks,M.S.:Focuscuesaffectperceiveddepth. Journal of Vision5(10), 7 (2005)

    2005

  54. [54]

    Journal of vision5(10), 7–7 (2005)

    Watt,S.J.,Akeley,K.,Ernst,M.O.,Banks,M.S.:Focuscuesaffectperceiveddepth. Journal of vision5(10), 7–7 (2005)

    2005

  55. [55]

    Acta Psychologica99(3), 293–310 (1998)

    Westerman, S.J., Cribbin, T.: Individual differences in the use of depth cues: Im- plications for computer- and video-based tasks. Acta Psychologica99(3), 293–310 (1998)

    1998

  56. [56]

    Wickens, C.D., Todd, S., Seidler, K.: Three-dimensional displays: Perception, im- plementation, and applications. Tech. Rep. AD-A259 937, University of Illinois (1989)

    1989

  57. [57]

    In: NeurIPS (2024)

    Yang,L.,Kang,B.,Huang,Z.,Zhao,Z.,Xu,X.,Feng,J.,Zhao,H.:DepthAnything V2. In: NeurIPS (2024)

    2024

  58. [58]

    In: ICLR (2024)

    You, H., Zhang, H., Gan, Z., Du, X., Zhang, B., Wang, Z., Cao, L., Chang, S.F., Yang, Y.: Ferret: Refer and Ground Anything Anywhere at Any Granularity. In: ICLR (2024)

    2024

  59. [59]

    In: NeurIPS (2024)

    Zhan, G., Zheng, C., Xie, W., Zisserman, A.: A General Protocol to Probe Large Vision Models for 3D Physical Understanding. In: NeurIPS (2024)

    2024

  60. [60]

    Answer:”instead of“Answer A or B

    Zou, Z., Chen, K., Shi, Z., Guo, Y., Ye, J.: Object detection in 20 years: A survey. Proceedings of the IEEE111(3), 257–276 (2023) Disentangling Pictorial Cue Understanding from Language Bias in VLMs via Depth Ordering Task (Supplementary Material) 1 Controlled Pictorial Cues Fig. 1 illustrates sampled images for the base view and 9 controlled pictorial c...

    work page arXiv 2023