Pith. sign in

REVIEW 3 major objections 6 minor 70 references

The 3D Mirage: Probing and Taming 3D Hallucinations

T0 review · 3 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Monocular depth models systematically hallucinate 3D structure from flat optical illusions, and a 4M-parameter LoRA adaptation cuts those hallucinations by over 93%.

desk verdict The benchmark is a real contribution, but the '93% taming' number is not credible — the training objective is the evaluation metric, and there's no held-out split. read the letter →

arxiv 2512.15423 v2 pith:2KDEGOAC submitted 2025-12-17 cs.CV cs.RO

Hoang Nguyen , Xiaohao Xu , Xiaonan Huang This is my paper
classification cs.CVcs.RO
keywords 3DhallucinationmonoculardepthestimationfoundationmodelsopticalillusionscontextualrobustnessLaplacianmetricsLoRAadaptationself-distillation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that the 3D Mirage—spurious depth structure hallucinated by monocular depth foundation models on geometrically planar but perceptually ambiguous surfaces—is a systemic, context-dependent failure, and that it can be probed, measured, and tamed. It introduces a benchmark of real-world illusion images with planar-region masks and context-restricted crops; two Laplacian-based metrics, DCS and CCS, that separate hallucination magnitude from contextual instability; and a parameter-efficient Grounded Self-Distillation method. The method reduces DCS by 93.5% and CCS by 86.1% on a Depth-Anything-V2 baseline while keeping NYU-v2 accuracy at 89.73%, suggesting the failure is correctable without retraining the whole model.

What carries the argument

The load-bearing objects are (1) the 3D-Mirage benchmark, providing manually annotated planar ROI masks over real-world illusion images plus context-restricted crops; (2) the DCS/CCS metrics, which take Laplacian (second-order) responses inside the ROI and separate radial hallucination magnitude (DCS) from full-vs-crop contextual drift (CCS); and (3) Grounded Self-Distillation, which injects LoRA adapters into the frozen teacher's ViT encoder and optimizes two losses: Hallucination Knowledge Re-editing (HKR), which fits a gated mixture of local planes and flattens the ROI, and Non-hallucination Knowledge Preservation (NKP), which tethers the student to the teacher on background, seams, edges

What would settle it

Collect a test set of 'false illusions'—images whose annotated planar ROI actually contains a genuine protrusion (the cube-type cases the paper reports as failures). If the adapted model flattens those protrusions as aggressively as it flattens painted flat illusions, the planarity assumption fails and the mitigation is over-flattening true geometry.

Watch

Extended reading notes

Core claim

The paper claims that state-of-the-art monocular depth estimation foundation models, despite strong zero-shot generalization, exhibit a systemic failure: on planar regions whose appearance suggests 3D geometry (street art, forced-perspective murals), and especially when the surrounding scene context is cropped away, they output spurious non-planar depth structure. The paper introduces the 3D-Mirage benchmark to trigger this failure, Laplacian-based metrics DCS and CCS to measure it, and Grounded Self-Distillation—a LoRA-based adapter trained with a hallucination re-editing loss and a knowledge-preservation loss—that reduces DCS by 93.5% and CCS by 86.1% on DAv2-L while retaining NYU-v2 accur

Load-bearing premise

The benchmark's annotated illusion regions are assumed to be genuinely flat in the real world; if an illusion actually contains real raised geometry, the method will flatten that true geometry.

Editorial extensions

If this is right

  • If the 93% reduction holds broadly, depth foundation models can be made safe against deceptive flat surfaces without retraining from scratch.
  • The benchmark and metrics give a standard way to compare models on structural and contextual robustness rather than pixel averages.
  • The finding that all tested architectures—transformer, diffusion, generative, metric, relative—share the failure indicates the vulnerability is a property of the training paradigm, not a single architecture.
  • The ablation shows full finetuning destroys general depth knowledge while the LoRA adaptation preserves it, suggesting parameter-efficient adaptation is the viable route for fixing such failure modes.

Reading between the lines

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

  • One can test whether the same mirage appears when the deceptive texture is not an artistic illusion but natural high-contrast texture (e.g., a flat rock wall); if so, the failure is broader than street art.
  • Since the paper's own Section G reports failures on protruding cube illusions, a natural next benchmark would mix planar and genuinely non-planar illusions to bound where flattening should stop; the metric pair would need a third axis for correct non-planarity.
  • The gated plane mixture suggests the method is implicitly learning a planarity prior for local regions; an interesting extension would be to apply the same self-distillation to other ill-posed dense tasks, such as surface-normal or occlusion-edge estimation.
  • The DCS/CCS pair could be turned into an online detector: a depth model that reports a large DCS inside a region a cheap classifier flags as planar could trigger a safety veto in autonomous systems.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper identifies a failure mode in monocular depth foundation models, termed the "3D Mirage": planar or low-curvature surfaces with deceptive 2D patterns are predicted as having spurious 3D structure, especially when the input is cropped to remove global context. The authors introduce a benchmark (3D-Mirage) of 468 real-world illusion images, expanded to 1,872 images with context-restricted crops and manually annotated planar ROIs. They propose two Laplacian-based metrics, DCS (deviation/hallucination magnitude) and CCS (contextual instability), and a mitigation method, Grounded Self-Distillation, which injects LoRA adapters into Depth-Anything-V2's encoder and trains with a composite loss that flattens illusion ROIs while preserving background geometry via a frozen teacher. Experiments report a 93.5% DCS reduction and 86.1% CCS reduction over the DAv2-L teacher, together with NYU-v2 accuracy and background R² to show knowledge preservation.

Significance. The problem is timely and practically important: depth foundation models are being deployed in safety-critical systems, and their sensitivity to contextual/illusory inputs is understudied. The paper's qualitative demonstrations of hallucination across several model families (DAv2, Depth Pro, Marigold, DepthFM, ZoeDepth, MiDaS) are compelling and support the existence of a real phenomenon. The proposed benchmark, if properly validated, could become a useful diagnostic tool, and the idea of a parameter-efficient, ROI-targeted correction is attractive. The authors also include thoughtful ablations showing that full finetuning causes catastrophic forgetting and that the knowledge-preservation loss is necessary to prevent indiscriminate flattening. However, the quantitative evidence for the central claim is currently compromised by the direct alignment between the training objective and the evaluation metric, the lack of a held-out evaluation, and the absence of statistical validation.

major comments (3)
  1. [Sec. 5.1.2–5.2, Eq. (9) vs Eq. (1)] The headline result is not interpretable as evidence of generalization. The model is trained on 3D-Mirage positives (Sec. 5.1.2: 'custom sampler with a 4:1 ratio of 3D-Mirage (positive) samples to regularizer (negative) samples') and evaluated on the same 3D-Mirage benchmark (Sec. 5.2), with no train/test split, cross-validation, or held-out illusion categories. Moreover, DCS (Eq. 1) aggregates Laplacian magnitude in the ROI, and the first loss term of L_HKR (Eq. 9) is alpha_1 |L(z)|_m over the same ROI. The training objective therefore directly minimizes the evaluation metric. The 93.5% DCS and 86.1% CCS reductions are in-sample and likely reflect optimization of the scoring function, not a learned ability to resist mirages. A held-out evaluation, ideally on illusion categories not seen in training, is mandatory; without it the central claim of 'taming' is unsupported.
  2. [Sec. 3.1 / Sec. G and Fig. O] The validity of DCS and L_HKR depends entirely on the assertion (Sec. 3.1) that each annotated ROI is 'planar in geometry.' This is a manual annotation with no reported reliability check, inter-annotator agreement, or independent geometric verification. The paper's own limitation section (Sec. G, Fig. O) acknowledges failures on 'Protrusion illusion' samples, and the benchmark scope is limited to planar illusions. Thus the benchmark cannot support the broad claim that 'all existing SOTA models are highly vulnerable to 3D mirages' beyond the curated planar subset. The authors should provide evidence for the planarity annotations (e.g., depth sensor measurements or multiple annotators), report metrics separately by illusion subtype, or temper the systemic-vulnerability claim accordingly.
  3. [Tables 1, 2, and A] All quantitative results are single point estimates without error bars, seeds, or significance tests. Training is only 1 epoch on 1,872 images and the paper selects the best checkpoint by epoch (Appendix E), which can inflate reported gains. Given the small dataset and the direct metric-matching objective, confidence intervals are necessary to support the magnitude of the claimed improvement. I also note that the No-L_NKP ablation reaches a lower DCS (46.8) than the full method (64.2); this reinforces that low DCS alone is not a reliable indicator of correct behavior, and the paper should report a perceptual or geometric accuracy measure inside the ROI, not only R² on the background.
minor comments (6)
  1. [Title] The arXiv title, 'The 3D Mirage: Probing and Taming 3D Hallucinations', differs from the title on the first page, 'Photorealistic Phantom Roads in Real Scenes: Disentangling 3D Hallucinations from Physical Geometry'. Please unify.
  2. [Sec. 5.1.1 / Sec. 5.3] The NYU-v2 evaluation reports 'mean pairwise accuracy' but no definition or citation is given. Specify the sampling procedure and the exact accuracy formula.
  3. [Sec. 2.3 / Intro] The claims 'first benchmark' and 'first end-to-end framework' are strong. Given prior work on depth hallucination and anomaly detection, please soften or clarify the novelty precisely, e.g., 'first benchmark centered on real-world optical-illusion scenes.'
  4. [Sec. 4.3, Eq. (8)] The gating network G is introduced but its architecture, input features, and training details are not specified. This is needed for reproducibility.
  5. [Table 1] Reporting R²(bg) = 100.00% for the baseline is tautological (teacher vs itself). Please state this explicitly or use a different reference point.
  6. [Fig. 3] The area statistics are useful but the dataset composition would be clearer with a breakdown of illusion types (chalk anamorphoses, forced-perspective murals, advertisements, etc.) and the number of ROIs with nested exclusions.

Circularity Check

2 steps flagged · score 7.0 of 10

Headline DCS/CCS reductions are in-sample: L_HKR (Eq. 9) penalizes the same ROI Laplacian energy that DCS (Eq. 1) measures, and 3D-Mirage is used for both training and evaluation without a held-out split.

  1. self definitional [Eq. (1) vs Eq. (9); Secs. 3.2 and 4.3]
    "d_cluster(i) = sqrt(t^2_full,i + t^2_crop,i), d_avg(i) = <sqrt((l^t_full(p))^2 + (l^t_crop(p))^2)>_{R_i} (1) and DCS_i = d_cluster(i) + d_avg(i). L_HKR = α1 |L(z)|_m + α2 (Σ_k w_k ℓ_k + w_null ℓ_null) (9)."

    DCS is a Laplacian-energy aggregate over the illusion ROI; the HKR training loss's primary term is the same Laplacian magnitude over the same ROI mask (α1|L(z)|_m). The paper's ablation (Table 2, No Hallucination Re-editing) confirms that this loss is the active ingredient: without it DCS stays at 971.1, with it DCS drops to 64.2. The reported 93.5% DCS reduction is therefore the direct effect of optimizing the evaluation quantity, not an independent measurement of hallucination taming.

  2. fitted input called prediction [Sec. 5.1.2, Sec. 5.1.3, Table 1]
    "Data. We use a custom sampler with a 4:1 ratio of 3D-Mirage (positive) samples to regularizer (negative) samples. ... We evaluate models on two fronts. First, we test for hallucination robustness using our 3D-Mirage benchmark with the proposed DCS (hallucination intensity) and CCS (contextual instability) metrics, where lower is better."

    LoRA adapters are trained on 3D-Mirage positive samples with the 4:1 sampler, and the headline DCS/CCS numbers are computed on the same '3D-Mirage benchmark.' No train/test split, cross-validation, or held-out illusion category is reported. Thus the 93.5% DCS and 86.1% CCS reductions are in-sample values: the parameters were fitted to the benchmark that is then used as the test, making the reported generalization a fitted input rather than a prediction.

full rationale

The framework is largely self-contained and does not rely on a load-bearing self-citation chain or an imported uniqueness theorem. The circularity is in the coupling between the metric and the training objective and in the absence of a held-out evaluation. DCS (Eq. 1) and the L_HKR term (Eq. 9) are built from the same ROI-Laplacian response; the paper's own No-L_HKR ablation shows that the loss is the active ingredient that produces the reported drop. Since the same 3D-Mirage benchmark is used for training (4:1 positive sampler) and evaluation, the headline reductions are in-sample fits. The independent NYU-v2 accuracy and background R² numbers show that knowledge is preserved, but they do not establish generalization to unseen illusions; Section G concedes failures on protrusion illusions (cubes) absent from training data. Hence a partial-circularity score of 7 is appropriate.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new physical entities, forces, or conserved quantities are introduced. DCS/CCS are derived metrics; LoRA adapters and the gating network are existing/architectural components. The central empirical claims rest on the planarity annotations, the Laplacian-as-hallucination proxy, the frozen teacher's background correctness, and a set of hand-set hyperparameters.

free parameters (5)
  • Loss weights α1-α7 = α1=1.0, α2=0.4, α3=1.0, α4=0.5, α5=0.3, α6=0.8, α7=0.3
    Hand-set in Sec. 5.1.2 to balance planarity, preservation, ring, and edge terms; directly defines the objective that produces the headline DCS/CCS numbers.
  • LoRA rank/alpha = r=16, α=32, dropout=0.05
    Chosen in Sec. 5.1.2; not swept or justified; capacity of the adaptation.
  • Positive:negative sampler ratio = 4:1
    Sec. 5.1.2 mixes 3D-Mirage positives with Penn-Fudan/CamVid negatives to avoid over-flattening; the balance is a hand-set knob.
  • Metric thresholds (top-10/mean-10, 1-99% percentile) = top 10%, discard lowest 10%, 1-99% per-view normalization
    Sec. 3.2; these thresholds define DCS/CCS and are chosen without sensitivity analysis.
  • Crop retention threshold = ≥40% of ROI diagonal
    Sec. 3.1; augmentation choice affects the CCS/context-restriction measurement.
assumptions (4)
  • domain assumption Illusion ROIs are geometrically planar.
    Sec. 3.1: 'These masks delineate regions that are planar in geometry yet suggest non-planarity in appearance.' All DCS/CCS and L_HKR planarity enforcement rest on this.
  • ad hoc to paper Laplacian magnitude in the ROI is a valid hallucination proxy.
    Sec. 3.2: 'A high response indicates a geometric hallucination.' No ground-truth depth is used; this is a modeling choice that can be gamed by over-flattening.
  • domain assumption Teacher background statistics are a safe normalization and the teacher's non-ROI predictions are trustworthy.
    Eqs. 6-7 and L_NKP in Eq. 10 tether the student to the frozen teacher everywhere outside ROIs; if the teacher errs there, the error is preserved.
  • domain assumption The human annotations and the 'all verified by human annotators' claim are accurate.
    Sec. 3.1; no inter-annotator agreement, no geometric verification, no source URLs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The 3D Mirage: Probing and Taming 3D Hallucinations." pith.science (2026). https://pith.science/paper/2KDEGOAC

@misc{pith2026251215423,
  author       = {Pith},
  title        = {Pith review of: The 3D Mirage: Probing and Taming 3D Hallucinations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2KDEGOAC}},
  note         = {Machine review of arXiv:2512.15423}
}
read the original abstract

Monocular depth foundation models achieve remarkable generalization by learning large-scale semantic priors, but this creates a critical vulnerability: they hallucinate illusory 3D structures from planar/low-curvature but perceptually ambiguous inputs. We term this failure the 3D Mirage. This paper introduces a novel end-to-end framework to probe, score, and tame this under-quantified safety risk in monocular depth under context variation. To probe, we present 3D-Mirage, the first benchmark to combine context variation and precise annotation for real-world illusions with real object exclusions, multi-surface support; purpose-built to stress-test monocular depth on real-world illusions. To score, we propose a second-order magnitude-based evaluation with two metrics: the Deviation Composite Score (DCS) for high second-order 3D structure and the Confusion Composite Score (CCS) for contextual instability. To tame this failure, we introduce Grounded Self-Distillation, a parameter-efficient strategy on Depth-Anything-V2 baseline that surgically targets and resolves hallucination on illusion ROIs while preserving background knowledge, avoiding catastrophic forgetting. Our work provides an innovative pipeline for diagnosing and addressing this phenomenon, urging a necessary shift in the evaluation of MDE from pixel-wise accuracy to structural and contextual robustness.

Figures

Figures reproduced from arXiv: 2512.15423 by the authors.

Figure 1
Figure 1. The 3D Mirage: Hallucinations induced by Illusive Phantom Road Patterns. (a) A driving scene featuring a decep￾tive phantom road pattern (3D illusion). (c) With full global con￾text, the depth foundation model [63] correctly identifies the road as planar. (d-f) However, when the view is restricted to the local region, the model fails to disambiguate the texture from geometry. It hallucinates significant non-existent… view at source ↗
Figure 2
Figure 2. , we subjected a wide range of architectures—from transformer-based (Depth-Anything V2 [63]) and diffusion￾based (Marigold [28]) to generative (DepthFM [20]) and commercially-developed (Depth Pro [6])—to these 3D mi￾rage inputs. All models exhibited similar failures, unstably predicting spurious 3D structures from planar surfaces. This collective failure exposes a critical gap in how we evaluate these models. Standa… view at source ↗
Figure 3
Figure 3. Statistics of illusion regions in the 3D-Mirage dataset. Area distributions for illusion regions (left) and their correspond￾ing random crops (right), as a percentage of the original image area. The dotted vertical line denotes the average value. create a strong perceptual suggestion of 3D geometry on a 2D plane. Planar ROI Annotation. After filtering, we manually an￾notated precise polygonal Region of Interest (ROI… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Overview of our Grounded Self-Distillation Pipeline. Our pipeline trains an Student model (fθ′ ) by injecting trainable LoRA adapters into the encoder of a frozen Teacher model (fθ). The system uses three streams to process an image containing a 3D illusion: (1) The Te…
Figure 5
Figure 5. Figure 5: Qualitative results of our Grounded Self-Distillation. Each row compares our model to the baseline on a 3D-Mirage sample. (1) Input RGB. (2) Error heatmap (Ours vs. Baseline), showing changes are confined to the ROI. (3) Baseline (DAv2-L) depth, which hallucinates 3D s…
Figure 6
Figure 6. Figure 6: Ablation: Effect of Knowledge Preservation (LNKP). (Left) Input RGB. (Center) Our full model’s output. (Right) The output without the LNKP loss. While the hallucination on the road is removed, the flattening effect leaks into the background, blur￾ring and distorting re…
Figure 7
Figure 7. Figure 7: Ablation: Effect of Hallucination Re-editing (LHKR). (Left) Input RGB. (Center) Our full model’s output. (Right) The output without the LHKR loss. The background is preserved, but the model completely fails to tame the 3D mirage, leaving the spu￾rious 3D structure on t…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 11 linked inside Pith

  1. [1]

    Intrinsic dimensional- ity explains the effectiveness of language model fine-tuning

    Armen Aghajanyan, Akshat Gupta, Akshat Shrivastava, Luke Zettlemoyer, and Sonal Gupta. Intrinsic dimensional- ity explains the effectiveness of language model fine-tuning. arXiv preprint arXiv:2012.13255, 2021. 5

    arXiv 2012

  2. [2]

    Stereo anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail.arXiv Table C.Performance over epochs: Ablation for No Knowledge Preservation (L NKP)

    Luca Bartolomei, Fabio Tosi, Matteo Poggi, and Stefano Mattoccia. Stereo anywhere: Robust zero-shot deep stereo matching even where either stereo or mono fail.arXiv Table C.Performance over epochs: Ablation for No Knowledge Preservation (L NKP). Epochd cluster↓d avg↓DCS↓D cluster↓D avg↓CCS↓R 2 [%] NYUv2 acc [%] DA-2k [%] 1 33.71 35.46 69.171.104×10 −4 1.1...

    arXiv 2024

  3. [3]

    Anomaly detection in 3d point clouds using deep geometric descriptors

    Paul Bergmann and David Sattlegger. Anomaly detection in 3d point clouds using deep geometric descriptors. InPro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, pages 2613–2623, 2023. 11, 12

    2023

  4. [4]

    Zoedepth: Zero-shot trans- fer by combining relative and metric depth.arXiv preprint arXiv:2302.12288, 2023

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias M ¨uller. Zoedepth: Zero-shot trans- fer by combining relative and metric depth.arXiv preprint arXiv:2302.12288, 2023. 1, 2, 6, 7

    arXiv 2023

  5. [5]

    Unsuper- vised scale-consistent depth and ego-motion learning from monocular video

    Jia-Wang Bian, Zhichao Li, Naiyan Wang, Huangying Zhan, Chunhua Shen, Ming-Ming Cheng, and Ian Reid. Unsuper- vised scale-consistent depth and ego-motion learning from monocular video. InAdvances in Neural Information Pro- cessing Systems (NeurIPS), 2019. 2

    2019

  6. [6]

    Richter, and Vladlen Koltun

    Aleksei Bochkovskii, Ama ¨el Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. InProceedings of the International Conference on Learning Representations (ICLR), 2025. 2, 6, 7

    2025

  7. [7]

    Brostow, Jamie Shotton, Julien Fauqueur, and Roberto Cipolla

    Gabriel J. Brostow, Jamie Shotton, Julien Fauqueur, and Roberto Cipolla. Semantic object classes in video: A high- definition ground truth database.Pattern Recognition Let- ters, 2009. 6

    2009

  8. [8]

    Comple- mentary pseudo multimodal feature for point cloud anomaly detection.Pattern Recognition, 2024

    Yunkang Cao, Xiaohao Xu, and Weiming Shen. Comple- mentary pseudo multimodal feature for point cloud anomaly detection.Pattern Recognition, 2024. 12

    2024

Show all 70 references
  1. [9]

    Kritika Chawla, Arjun Majumdar, Shubham Raman, Chetan Arora, and C.V . Jawahar. Error diagnosis of deep monocular depth estimation models. In2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 8400–8407. IEEE, 2021. 2

    2021

  2. [10]

    Fpr: False-positive rec- tification for weakly-supervised semantic segmentation

    Liang Chen, Jia Wu, and Ping Luo. Fpr: False-positive rec- tification for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16282–16291, 2023. 2

    2023

  3. [11]

    Easynet: An easy net- work for 3d industrial anomaly detection.arXiv preprint arXiv:2307.13925, 2023

    Ruitao Chen, Guoyang Xie, Jiaqi Liu, Jinbao Wang, Ziqi Luo, Jinfan Wang, and Feng Zheng. Easynet: An easy net- work for 3d industrial anomaly detection.arXiv preprint arXiv:2307.13925, 2023. 13

    2023 arXiv

  4. [12]

    Single- image depth perception in the wild

    Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single- image depth perception in the wild. InAdvances in Neural Information Processing Systems (NeurIPS), 2016. 2 Figure O.Limitations and Failure Cases.Examples where our model fails to fully suppress the 3D mirage. Left to ri...

    2016

  5. [13]

    Learning depth estimation for transparent and mirror surfaces

    Walter Costanzino, Yixin Zhou, Hao Jiang, and Jiwen Lu. Learning depth estimation for transparent and mirror surfaces. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3788–3797. IEEE, 2023. 3

    2023

  6. [14]

    Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2432–2443, 2017. 3

    2017

  7. [15]

    Depth map prediction from a single image using a multi-scale deep net- work

    David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep net- work. InAdvances in Neural Information Processing Sys- tems (NeurIPS), 2014. 2

    2014

  8. [16]

    Vision meets robotics: The kitti dataset.Interna- tional Journal of Robotics Research (IJRR), 2013

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset.Interna- tional Journal of Robotics Research (IJRR), 2013. 3

    2013

  9. [17]

    Cl ´ement Godard, Oisin Mac Aodha, and Gabriel J. Brostow. Unsupervised monocular depth estimation with left–right consistency. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6602–6611, 2017. 2

    2017

  10. [18]

    Cl ´ement Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J. Brostow. Digging into self-supervised monocular depth estimation. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision (ICCV), pages 3827– 3837, 2019. 2

    2019

  11. [19]

    Rethinking reverse distillation for multi-modal anomaly detection

    Zhihao Gu, Jiangning Zhang, Liang Liu, Xu Chen, Jin- long Peng, Zhenye Gan, Guannan Jiang, Annan Shu, Yabiao Wang, and Lizhuang Ma. Rethinking reverse distillation for multi-modal anomaly detection. InProceedings of the AAAI Conference on Artificial Intelligence, pages 8445– 8...

    2024

  12. [20]

    Depthfm: Fast monocular depth estimation with flow matching

    Ming Gui, Johannes Schusterbauer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Bj ¨orn Ommer. Depthfm: Fast monocular depth estimation with flow matching. InProceedings of the AAAI Conference on Artificial Intelli...

    2025

  13. [21]

    3d packing for self-supervised monocular depth estimation

    Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raven- tos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2485–2494, 2020. 3

    2020

  14. [22]

    Monocular depth estimation using whole strip masking and reliability-based refinement

    Minhyeok Heo, Jaehan Lee, Kyung-Rae Kim, Han-Ul Kim, and Chang-Su Kim. Monocular depth estimation using whole strip masking and reliability-based refinement. InEu- ropean Conference on Computer Vision (ECCV), pages 219– 234, 2018. 3

    2018

  15. [23]

    Back to the feature: clas- sical 3d features are (almost) all you need for 3d anomaly detection

    Eliahu Horwitz and Yedid Hoshen. Back to the feature: clas- sical 3d features are (almost) all you need for 3d anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2967– 2976, 2023. 12

    2023

  16. [24]

    Parameter-efficient transfer learn- ing for nlp.arXiv preprint arXiv:1902.00751, 2019

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attia, and Sylvain Gelly. Parameter-efficient transfer learn- ing for nlp.arXiv preprint arXiv:1902.00751, 2019. 3, 6

    1902 arXiv

  17. [25]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. InIn- ternational Conference on Learning Representations (ICLR). OpenReview.net, 2022. 3, 5

    2022

  18. [26]

    OPERA: Alleviating hallucination in multi- modal large language models via over-trust penalty and retrospection-allocation

    Qidong Huang, Xiaoyi Dong, Pan Zhang, Bin Wang, Con- ghui He, Jiaqi Wang, Dahua Lin, Weiming Zhang, and Nenghai Yu. OPERA: Alleviating hallucination in multi- modal large language models via over-trust penalty and retrospection-allocation. InProceedings of the IEEE/CVF Confere...

    2024

  19. [27]

    van Gemert

    O ˘guzhan Kayhan and Jan C. van Gemert. Hallucination in object detection: A study in visual part verification.arXiv preprint, arXiv:2106.02523, 2021. 2

    2021 arXiv

  20. [28]

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Met- zger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos- Figure P.Qualitative Results: Relative Models.Our model (right cols) successfully resolves the top hallucination cases of the Depth Anything series (left cols) under ...

    2024

  21. [29]

    Automated audit and self- correction algorithm for seg-hallucination using meshcnn- based on-demand generative ai.Bioengineering (Basel), 12 (1):81, 2025

    Jiwon Kim and Soo Hyun Lee. Automated audit and self- correction algorithm for seg-hallucination using meshcnn- based on-demand generative ai.Bioengineering (Basel), 12 (1):81, 2025. 2

    2025

  22. [30]

    V olcano: Mitigating multimodal hallucina- tion through self-feedback guided revision.arXiv preprint arXiv:2311.07362, 2024

    Seongyun Lee, Sue Hyun Park, Yongrae Jo, and Min- joon Seo. V olcano: Mitigating multimodal hallucina- tion through self-feedback guided revision.arXiv preprint arXiv:2311.07362, 2024. 2, 3

    2024 arXiv

  23. [31]

    Mitigat- ing object hallucinations in large vision-language models through visual contrastive decoding

    Sicong Leng, Hang Zhang, Guanzheng Chen, Xinting Li, Shuai Lu, Chunyan Miao, and Lidong Bing. Mitigat- ing object hallucinations in large vision-language models through visual contrastive decoding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni...

    2024

  24. [32]

    Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network

    Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. Figure Q.Qualitative Results: Metric Models.Our model (right cols) c...

    2024

  25. [33]

    Multi-sensor object anomaly detection: Unifying appearance, geometry, and in- ternal properties

    Wenqiao Li, Bozhong Zheng, Xiaohao Xu, Jinye Gan, Fad- ing Lu, Xiang Li, Na Ni, Zheng Tian, Xiaonan Huang, Shenghua Gao, and Yingna Wu. Multi-sensor object anomaly detection: Unifying appearance, geometry, and in- ternal properties. InProceedings of the IEEE/CVF Confer- ence o...

    2025

  26. [34]

    Online video instance segmentation via robust context fusion.arXiv preprint arXiv:2207.05580, 2022

    Xiang Li, Jinglu Wang, Xiaohao Xu, Bhiksha Raj, and Yan Lu. Online video instance segmentation via robust context fusion.arXiv preprint arXiv:2207.05580, 2022. 1

    2022 arXiv

  27. [35]

    Robust referring video object segmentation with cyclic structural consensus

    Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust referring video object segmentation with cyclic structural consensus. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22236–22245, 2023

    2023

  28. [36]

    Towards noise-tolerant speech-referring video object segmentation: Bridging speech and text

    Xiang Li, Jinglu Wang, Xiaohao Xu, Muqiao Yang, Fan Yang, Yizhou Zhao, Rita Singh, and Bhiksha Raj. Towards noise-tolerant speech-referring video object segmentation: Bridging speech and text. InProceedings of the 2023 Confer- ence on Empirical Methods in Natural Language Proc...

    2023

  29. [37]

    R 2-bench: Benchmarking the robustness of referring perception models under perturbations

    Xiang Li, Kai Qiu, Jinglu Wang, Xiaohao Xu, Rita Singh, Kashu Yamazaki, Hao Chen, Xiaonan Huang, and Bhiksha Raj. R 2-bench: Benchmarking the robustness of referring perception models under perturbations. InEuropean Confer- ence on Computer Vision, pages 211–230. Springer, 2024

    2024

  30. [38]

    Qdformer: Towards ro- bust audiovisual segmentation in complex environments with quantization-based semantic decomposition

    Xiang Li, Jinglu Wang, Xiaohao Xu, Xiulian Peng, Rita Singh, Yan Lu, and Bhiksha Raj. Qdformer: Towards ro- bust audiovisual segmentation in complex environments with quantization-based semantic decomposition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat...

    2024

  31. [39]

    Optimizing lidar placements for robust driving perception in adverse conditions.CoRR, 2024

    Ye Li, Lingdong Kong, Hanjiang Hu, Xiaohao Xu, and Xiao- nan Huang. Optimizing lidar placements for robust driving perception in adverse conditions.CoRR, 2024. 1

    2024

  32. [40]

    Real3d- ad: A dataset of point cloud anomaly detection.Advances in Neural Information Processing Systems, 36:30402–30415,

    Jiaqi Liu, Guoyang Xie, Ruitao Chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3d- ad: A dataset of point cloud anomaly detection.Advances in Neural Information Processing Systems, 36:30402–30415,

  33. [41]

    Negative object presence evaluation (nope) to measure object hallucination in vision-language models

    Holy Lovenia, Wenliang Dai, Samuel Cahyawijaya, Ziwei Ji, and Pascale Fung. Negative object presence evaluation (nope) to measure object hallucination in vision-language models. InProceedings of the ACL 2024 Workshop on Advancing Language and Vision Research (ALVR), 2024. arXi...

    2024 arXiv

  34. [42]

    Universal adversarial perturba- tions

    Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturba- tions. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 86–94, 2017. 2

    2017

  35. [43]

    Hallucinat- ing hidden obstacles for unmanned surface vehicles using a compositional model

    Jon Muhovi ˇc, Gregor Koporec, and Janez Per ˇs. Hallucinat- ing hidden obstacles for unmanned surface vehicles using a compositional model. InProceedings of the 26th Computer Vision Winter Workshop (CVWW). University of Ljubljana,

  36. [44]

    Deep neural networks are easily fooled: High confidence predictions for unrecognizable images

    Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. InProceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 427–436. IEEE, 2015. 2

    2015

  37. [45]

    Penn- fudan database for pedestrian detection and segmentation,

    University of Pennsylvania and Fudan University. Penn- fudan database for pedestrian detection and segmentation,

  38. [46]

    Image tokenizer needs post-training.arXiv preprint arXiv:2509.12474, 2025

    Kai Qiu, Xiang Li, Hao Chen, Jason Kuen, Xiaohao Xu, Jiuxiang Gu, Yinyi Luo, Bhiksha Raj, Zhe Lin, and Mar- ios Savvides. Image tokenizer needs post-training.arXiv preprint arXiv:2509.12474, 2025. 1

    2025

  39. [47]

    Robust latent matters: Boosting image generation with sampling error synthesis.arXiv preprint arXiv:2503.08354, 2025

    Kai Qiu, Xiang Li, Jason Kuen, Hao Chen, Xiaohao Xu, Jiuxiang Gu, Yinyi Luo, Bhiksha Raj, Zhe Lin, and Mar- ios Savvides. Robust latent matters: Boosting image generation with sampling error synthesis.arXiv preprint arXiv:2503.08354, 2025. 1

    2025 arXiv

  40. [48]

    Vi- sion transformers for dense prediction

    Ren ´e Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vi- sion transformers for dense prediction. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12179–12188, 2021. 1

    2021

  41. [49]

    Ren ´e Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 44(3):1623–1637, 2022. 1, 2, 6, 7

    2022

  42. [50]

    Asymmetric student-teacher networks for indus- trial anomaly detection

    Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bas- tian Wandt. Asymmetric student-teacher networks for indus- trial anomaly detection. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2592–2602, 2023. 12

    2023

  43. [51]

    Chung, and Andrew Y

    Ashutosh Saxena, Sung H. Chung, and Andrew Y . Ng. Learning depth from single monocular images. InAdvances in Neural Information Processing Systems (NIPS), 2005. 2

    2005

  44. [52]

    Indoor segmentation and support inference from RGBD images

    Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from RGBD images. InProceedings of the European Conference on Computer Vision (ECCV), pages 746–760, 2012. 3, 6

    2012

  45. [53]

    Don’t judge an object by its context: Learning to overcome con- textual bias

    Krishna Kumar Singh, Dhruv Mahajan, Kristen Grauman, Yong Jae Lee, Matt Feiszli, and Deepti Ghadiyaram. Don’t judge an object by its context: Learning to overcome con- textual bias. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pag...

    2020

  46. [54]

    Antonio Torralba and Alexei A. Efros. Unbiased look at dataset bias. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1521–1528, 2011. 2

    2011

  47. [55]

    Self-supervised feature adaptation for 3d industrial anomaly detection

    Yuanpeng Tu, Boshen Zhang, Liang Liu, Yuxi Li, Jiangning Zhang, Yabiao Wang, Chengjie Wang, and Cairong Zhao. Self-supervised feature adaptation for 3d industrial anomaly detection. InEuropean Conference on Computer Vision, pages 75–91. Springer, 2024. 11

    2024

  48. [56]

    Tartanair: A dataset to push the limits of visual SLAM

    Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Se- bastian Scherer. Tartanair: A dataset to push the limits of visual SLAM. InProceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), ...

    2020

  49. [57]

    Multimodal industrial anomaly detection via hybrid fusion

    Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. Multimodal industrial anomaly detection via hybrid fusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8032–8041, 2023. 12

    2023

  50. [58]

    C. J. Willmott and K. Matsuura. Root mean square error (rmse) or mean absolute error (mae)? arguments against avoiding rmse in the literature. 2015. 2

    2015

  51. [59]

    Targeted adver- sarial perturbations for monocular depth prediction

    Alex Wong, Safa Cicek, and Stefano Soatto. Targeted adver- sarial perturbations for monocular depth prediction. InAd- vances in Neural Information Processing Systems (NeurIPS),

  52. [60]

    Customizing visual-language foundation models for multi-modal anomaly detection and reasoning

    Xiaohao Xu, Yunkang Cao, Huaxin Zhang, Nong Sang, and Xiaonan Huang. Customizing visual-language foundation models for multi-modal anomaly detection and reasoning. In 2025 28th International Conference on Computer Supported Cooperative Work in Design (CSCWD), pages 1443–1448,

    2025

  53. [61]

    Towards ambiguity-free spatial foundation model: Rethinking and decoupling depth ambiguity.arXiv preprint arXiv:2503.06014, 2025

    Xiaohao Xu, Feng Xue, Xiang Li, Haowei Li, Shusheng Yang, Tianyi Zhang, Matthew Johnson-Roberson, and Xi- aonan Huang. Towards ambiguity-free spatial foundation model: Rethinking and decoupling depth ambiguity.arXiv preprint arXiv:2503.06014, 2025. 3

    2025 arXiv

  54. [62]

    Scal- able benchmarking and robust learning for noise-free ego- motion and 3d reconstruction from noisy video

    Xiaohao Xu, Tianyi Zhang, Shibo Zhao, Xiang Li, Sibo Wang, Yongqi Chen, Ye Li, Bhiksha Raj, Matthew Johnson- Roberson, Sebastian Scherer, and Xiaonan Huang. Scal- able benchmarking and robust learning for noise-free ego- motion and 3d reconstruction from noisy video. InThe Thi...

    2025

  55. [63]

    Depth any- thing v2

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2. InAdvances in Neural Information Processing Sys- tems (NeurIPS), 2024. 1, 2, 6, 7

    2024

  56. [64]

    Depth anything: Unleashing the power of large-scale unlabeled data

    Lihe Yin, Pengfei Liu, Qian Zhu, Zikang Lin, Linrui Kong, and Fan Yang. Depth anything: Unleashing the power of large-scale unlabeled data. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024. 2, 6, 7

    2024

  57. [65]

    Cheat- ing depth: Enhancing 3d surface anomaly detection via depth simulation

    Vitjan Zavrtanik, Matej Kristan, and Danijel Sko ˇcaj. Cheat- ing depth: Enhancing 3d surface anomaly detection via depth simulation. InProceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision, pages 2164–2172,

  58. [66]

    Keep dræming: discriminative 3d anomaly detection through anomaly simulation.Pattern Recognition Letters, 181:113– 119, 2024

    Vitjan Zavrtanik, Matej Kristan, and Danijel Sko ˇcaj. Keep dræming: discriminative 3d anomaly detection through anomaly simulation.Pattern Recognition Letters, 181:113– 119, 2024. 11

    2024

  59. [67]

    Bridging 3d anomaly localization and repair via high-quality continuous geometric representation

    Bozhong Zheng, Jinye Gan, Xiaohao Xu, Xintao Chen, Wen- qiao Li, Xiaonan Huang, Na Ni, and Yingna Wu. Bridging 3d anomaly localization and repair via high-quality continuous geometric representation. InProceedings of the IEEE/CVF International Conference on Computer Vision (IC...

    2025

  60. [68]

    Physical 3d adversarial attacks against monocular depth estimation in autonomous driving

    Junhao Zheng, Chenhao Lin, Jiahao Sun, Zhengyu Zhao, Qian Li, and Chao Shen. Physical 3d adversarial attacks against monocular depth estimation in autonomous driving. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3

    2024

  61. [69]

    Tinghui Zhou, Matthew Brown, Noah Snavely, and David G. Lowe. Unsupervised learning of depth and ego-motion from video. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1851–1858,

  62. [70]

    R3d-ad: Reconstruction via diffu- sion for 3d anomaly detection

    Zheyuan Zhou, Le Wang, Naiyu Fang, Zili Wang, Lemiao Qiu, and Shuyou Zhang. R3d-ad: Reconstruction via diffu- sion for 3d anomaly detection. InEuropean Conference on Computer Vision, pages 91–107. Springer, 2024. 13 Figure R.Performance Comparison 1.Visual comparison of our Gr...

    2024

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.