Pith. sign in

REVIEW 3 major objections 6 minor 44 references

PhysFlood, a mask-based diffusion inpainting system, synthesizes realistic flood scenes from a single fisheye photo; its core claim is that a height-derived mask keeps generated water at the commanded level, where mask-free generators canno

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 23:02 UTC pith:5PYE5MFW

load-bearing objection PhysFlood is a sensible single-image fisheye flood synthesis system with a genuinely new combination of geometry-based masks and diffusion inpainting, but its central water-level-control claim rests on thin, low-agreement human evidence. the 3 major comments →

arxiv 2607.15527 v1 pith:5PYE5MFW submitted 2026-07-17 cs.CV

Physics-aware Masked Diffusion-based Flood Simulation for Urban Fisheye Disaster Detection

Sodtavilan Odonchimed , Tsogt Enkhbayar , Oyunzul Munkhtamga , Munkhjargal Gochoo This is my paper
classification cs.CV
keywords flood simulationfisheye imagediffusion modelimage inpaintingwater level control3D height estimationurban surveillancedata augmentation
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.

PhysFlood is built to solve a practical shortage: real flooded fisheye-surveillance images are rare, and existing image generators either need multiple views or cannot control the water depth. The paper's core claim is that if you first estimate per-pixel 3D heights and segment the ground, you can compute exactly which pixels would be submerged at a given water level, and use that mask to drive an inpainting diffusion model. This produces images whose water matches the commanded depth, whereas mask-free image-to-image generation makes realistic-looking floods at arbitrary, uncontrolled levels. The authors demonstrate this with four generators (two open diffusion backbones and two commercial editing models) and defend it with a three-rater human study across five water levels. If the claim holds, flood training data for fisheye surveillance can be manufactured on demand at specified severities.

Core claim

The central claim is that the mask, not the generator, carries the water-level control. Using monocular 3D estimation (Unik3D) and fisheye-trained ground segmentation (YOLO26) minus object masks (SAM3), the Masker builds a per-pixel height reference with the ground plane as zero and a locally fitted plane under each object to avoid occlusion errors. The mask is the set of pixels whose height falls below the commanded water level Hw. Two LoRA-fine-tuned diffusion backbones (SD3.5, FLUX.1) and two commercial editing models (NanoBanana Pro, GPT-Image-2) then inpaint only those pixels. The paper's human evaluation shows that masked generation matches commanded water levels across 0.2–4.0 m, whil

What carries the argument

The Masker's height estimation chain: monocular 3D point estimation yields per-pixel 3D candidate points; a fisheye-trained ground segmenter defines ground anchors; object masks from a general segmenter are subtracted; RANSAC fits the ground normal; a coordinate transform splits each point into ground-plane position and height; then a local planar fit per object footprint corrects the reference ground height under occluding objects. The output mask (pixels with height < Hw) is the single mechanism that couples the physically meaningful water level to the stochastic image generator.

Load-bearing premise

The physical meaning of the commanded water level rests entirely on the metric accuracy of the 3D heights estimated from a single fisheye image; if that estimate is biased, the mask still draws a plausible flood region but the water depth is not the number the user requested.

What would settle it

Generate a flood at a known shallow level (say 0.5 m) in a fisheye scene containing a reference object of known height, e.g., a traffic cone or a person, then measure where the waterline crosses that object in the output image; a deviation greater than roughly 10 cm would refute the metric-height assumption. A cleaner check is to compare Unik3D's per-pixel heights on fisheye frames against LiDAR ground truth.

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

If this is right

  • Because the mask is computed from a single image without multi-view reconstruction, any existing fisheye surveillance frame can be turned into a flood scene at a specified severity.
  • Since the mask is agnostic to the generator, the same mask can be applied to different diffusion backbones or commercial editors, and the paper shows level-consistent results with all four.
  • Varying the commanded water level Hw yields flood versions of the same scene at multiple severities, providing on-demand training data for flood detection models.
  • The human study's comparison of masked vs mask-free outputs indicates that mask-free generators produce visually realistic floods whose depth is not faithful to the command, while masked generation keeps depth aligned.

Where Pith is reading between the lines

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

  • A quantitative validation of the commanded level — measuring the waterline against a reference object of known height in the generated image — would convert the reported human ratings into a metric check of the physics claim; the paper suggests this as future work.
  • The same height-mask abstraction applies to any level-based phenomenon (snow depth, smoke, lava), as long as per-pixel heights can be estimated, so the pipeline may transfer beyond flooding.
  • If Unik3D's metric depths are biased on fisheye frames, the mask still selects the low-lying scene regions, so the system likely degrades from 'physics-aware' to 'location-aware' rather than failing outright.
  • The low inter-rater agreement on physical plausibility at shallow levels suggests human judgment is a noisy ruler; future benchmarks should combine human scores with geometric waterline measurements.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. PhysFlood proposes a pipeline for synthesizing flood images from a single fisheye urban image. The Masker estimates per-pixel height above the ground using Unik3D and SAM3/YOLO26 segmentation, then binarizes a user-commanded water level Hw into a mask (Eq. 10). The Generator performs masked inpainting with diffusion models, optionally fine-tuned via LoRA, or with foundation models (NanoBanana Pro, GPT-Image-2). The paper claims that masked generation allows explicit control of water level, unlike mask-free image-to-image generation, and supports this with a three-rater human evaluation.

Significance. If the central claim were established, PhysFlood would be a useful tool for generating fisheye flood data for training disaster-detection models, addressing a real data-scarcity problem. The paper also introduces a non-trivial masking pipeline that combines monocular 3D estimation with object-ground interpolation, and it experiments with several modern diffusion backbones. However, the current evidence does not substantiate the water-level control claim: the mask definition is physically questionable on non-planar ground, and the only evaluation is a small human study with low inter-rater agreement. The comparative evaluation is also confounded because the fine-tuned LoRA models are trained on images generated by the same foundation models that serve as baselines.

major comments (3)
  1. [§3.1, Eq. (3) and Eq. (10)] The mask M = {(i,j) | h(P(i,j)) < Hw} is a per-pixel threshold on height above the local ground surface, not a mask of a level water plane. On sloped or curved ground, this defines a water layer of approximately uniform thickness draped over the terrain, whereas a real flood has a horizontal water surface whose waterline is the intersection of that plane with the scene. The paper never demonstrates that the generated masks are consistent with a horizontal water plane, nor does it validate the metric accuracy of Unik3D/YOLO26 height estimates on fisheye images. Since this mask is the core mechanism for 'physics-aware' water-level control, the physical meaning of Hw is unverified.
  2. [§3.3, §4.2, Table 2] The only evidence for the central claim that PhysFlood generates images matching commanded water levels is a human study with three raters. Krippendorff's alpha for Q2 (physical plausibility) is as low as 0.101 at 0.5 m and 0.135 at 0.2 m, indicating unreliable agreement. No objective measurement of the actual water height in the generated images is reported. Figure 5 shows qualitative mask-free failures, but there is no quantitative comparison of generated vs. commanded water levels. Section 4.3 explicitly concedes that objective metrics and ground-truth comparisons are missing. Thus the central claim is currently unsupported.
  3. [§3.2, §4.1, §4.2] The comparison between PhysFlood and the foundation models is circular. The LoRA-fine-tuned SD3.5 and FLUX.1 models are trained on images generated by NanoBanana Pro and GPT-Image-2 (Section 3.2, 'Data Augmentation'), and the same two foundation models are then used as baselines in the evaluation. Any performance gap between the fine-tuned models and the foundation models reflects this training/evaluation loop, not the intrinsic merit of the mask-based approach. This confound undermines the comparative conclusions in Figure 4 and the claim that the masked approach outperforms mask-free generation.
minor comments (6)
  1. [§3.2] Typo: 'Makser' should be 'Masker'.
  2. [§3.1] Typo: 'Noramlized' should be 'Normalized'.
  3. [§3.1, Eq. (2)] The notation P_i = (X_i,Y_i,Z_i) → (a_i,b_i,s_i) is unclear. Clarify how the 3D point cloud is transformed into the (a,b,s) coordinate system and how e1, e2, nhat are computed from the RANSAC plane.
  4. [§4.2, Figure 4] Relabeling the 0–3 ordinal scores into binary 'good/bad' collapses the scale and discards information. Also, no error bars or statistical significance tests are provided; with only three raters, the reported percentages are noisy.
  5. [§4.1] The training set construction is not fully specified: 32 source images × 3 levels × 2 foundation models = 192, not 224. Clarify how the extra images were obtained (e.g., multiple prompts or views).
  6. [§4.2, Table 2] The table reports alpha values but no confidence intervals. Given the low observed agreement, report per-question alpha with CIs and discuss whether the coding scheme is reliable enough to draw conclusions.

Circularity Check

0 steps flagged

No significant circularity: the mask is definitionally tied to the commanded height, but the image-generation claim is an empirical result, not a reduction.

full rationale

The paper's derivation chain is: commanded water level Hw -> estimated per-pixel height h(P) (Eq. 3) -> edit mask M = {(i,j) | h(P(i,j)) < Hw} (Eq. 10) -> masked inpainting (Eq. 11) -> generated flood image. The mask is explicitly defined in terms of Hw, so any statement that the mask encodes the commanded height is true by construction. However, the paper's central empirical claim is not that the mask equals the commanded height; it is that mask-conditioned inpainting generates images that human raters judge as matching that height. That claim is stochastic and testable, even though the paper's validation is weak: the human study has low inter-rater agreement (e.g., Krippendorff's alpha = 0.101 for Q2 at 0.5m) and the Limitations section concedes the absence of objective water-level measurements. Lack of validation is a correctness concern, not circularity. The use of NanoBanana Pro and GPT-Image-2 both to synthesize training data and as comparison baselines creates a benchmarking limitation, but it does not reduce any derived quantity to an input by construction. The self-citation to Fisheye8K is a dataset citation, not a load-bearing uniqueness or ansatz theorem. No step in the paper's derivation is equivalent to its own inputs beyond the definitional mask equation, which the paper does not disguise as a prediction. Therefore no significant circularity is present.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities. The free parameters are mostly hand-set hyperparameters. The central assumptions are about the accuracy of pre-trained models (Unik3D, SAM3, YOLO26) and the simplification of flood geometry to a global height threshold, both load-bearing for the mask computation.

free parameters (5)
  • tau = 20th percentile of object's height components
    Threshold for selecting low-lying object points for footprint extraction; hand-set in Section 4.1.
  • delta = 1.5
    Margin around object footprint for local ground interpolation; hand-set in Section 4.1.
  • LoRA rank = 16
    Rank of low-rank adaptation; hand-set in Section 4.1.
  • training steps = 4000
    Number of fine-tuning steps; hand-set in Section 4.1.
  • inpainting strength = 0.8
    Strength of noise applied before denoising; hand-set in Section 4.1.
axioms (5)
  • domain assumption Unik3D provides accurate metric 3D coordinates for fisheye images
    Masker relies on this for all height estimates; if biased, the water level mask is wrong. Enters at Eq. (2)-(3).
  • domain assumption SAM3 and YOLO26 segment objects and ground correctly on fisheye images
    Ground region MG is computed by subtracting SAM3 object masks from YOLO26 segmentation; errors propagate to height estimation. Eq. (1).
  • domain assumption Ground areas can be approximated as locally planar for RANSAC and interpolation
    RANSAC fits a single normal vector to ground points; object-ground interpolation uses local planar fits over footprint margins. Eq. (2)-(9).
  • domain assumption A flood can be represented by a single scalar water height Hw over the whole scene
    Hydraulics, flow, and connectivity are ignored; the mask is defined by a global threshold on per-pixel height. Eq. (10).
  • ad hoc to paper NanoBanana Pro and GPT-Image-2 generate flood scenes that are realistic enough to serve as training data for fine-tuning
    The LoRA models are trained on images generated by these foundation models, so any bias in the foundation models is absorbed into the fine-tuned models. Section 3.2 and 4.1.

pith-pipeline@v1.3.0-alltime-deepseek · 10446 in / 10201 out tokens · 95372 ms · 2026-08-01T23:02:43.809266+00:00 · methodology

0 comments
read the original abstract

Physical simulations that predict the behavior of urban disasters, such as climate-related flooding, play a crucial role in disaster prevention and the development of anomaly detection models. However, the severe shortage of flood data in real-world environments, combined with the inherent distortions of fisheye lens images, which are used for urban surveillance, has made high-precision simulations challenging. To address this, we propose a new physical simulation system PhysFlood that leverages Diffusion Models to synthesize realistic floods from just a single image captured by a fisheye lens. Our system not only enables simulation from a single image, but also features the ability to freely control and generate diverse flood scenarios by manipulating physically meaningful variables, such as water levels. In our evaluation experiments, we conducted a qualitative human study and demonstrated that the simulation images generated by PhysFlood exhibit both acceptable realism and robustness.

Figures

Figures reproduced from arXiv: 2607.15527 by Munkhjargal Gochoo, Oyunzul Munkhtamga, Sodtavilan Odonchimed, Tsogt Enkhbayar.

Figure 1
Figure 1. Figure 1: Overview of Generating Fisheye Flood Images. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of Object Footprints and The Margin Areas. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of object heights estimated from [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Percentage of "good" Samples. The figure shows the scores for each model under both masked and mask-free conditions. For each question (Q1, Q2, and Q3), the scores have been relabeled into binary classes, mapping 0–1 to "bad" and 2–3 to "good". The x-axis of each plot represents the input water levels (0.2 m, 0.5 m, 1.0 m, 2.0 m, and 4.0 m), while the y-axis indicates the percentage of images classified as… view at source ↗
Figure 5
Figure 5. Figure 5: No Mask Outputs. Example plots of the top four samples with the lowest Q2 scores for Image-to-Image generation using NanoBanana Pro and GPT-Image-2. The height indicated above each image represents the commanded water level. These examples reveal that while the generated images are realistic, the actual water levels are noticeably lower or higher than the specified input. sues with the training data distri… view at source ↗
Figure 6
Figure 6. Figure 6: Failure Cases. Results generated using GPT-Image-2 and NanoBanana Pro at commanded water levels of 1.0 m and 2.0 m. Left: Masker failure case where excessive fisheye lens tilt leads to inaccurate mask generation. Right: Generator failure case where the flooded regions are colored despite the grayscale input [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Masker Outputs. Plotting results of the masks generated when varying the water height to 0.2m, 0.5m, 1.0m, and 2.0m. The light blue regions represent the estimated flood levels [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Generated Samples. Plotting generated images under both masked and mask-free conditions for all water levels [PITH_FULL_IMAGE:figures/full_fig_p014_8.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

44 extracted references · 11 linked inside Pith

  1. [1]

    Floodnet: A high resolution aerial imagery dataset for post flood scene understanding.IEEE Access, 9:89644–89654, 2021

    Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and Robin Roberson Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding.IEEE Access, 9:89644–89654, 2021

    2021

  2. [2]

    Deep sensing of urban waterlogging.IEEE Access, 9:127185–127203, 2021

    Shi-Wei Lo, Jyh-Horng Wu, Jo-Yu Chang, Chien-Hao Tseng, Meng-Wei Lin, and Fang-Pang Lin. Deep sensing of urban waterlogging.IEEE Access, 9:127185–127203, 2021

    2021

  3. [3]

    Deep-learning-based aerial image classification for emergency response applications using unmanned aerial vehicles

    Christos Kyrkou and Theocharis Theocharides. Deep-learning-based aerial image classification for emergency response applications using unmanned aerial vehicles. InCVPR workshops, pages 517–525, 2019

    2019

  4. [4]

    Climategan: Raising cli- mate change awareness by generating images of floods.arXiv preprint arXiv:2110.02871, 2021

    Victor Schmidt, Alexandra Sasha Luccioni, Mélisande Teng, Tianyu Zhang, Alexia Reynaud, Sunand Raghupathi, Gautier Cosne, Adrien Juraver, Vahe Vardanyan, Alex Hernandez-Garcia, et al. Climategan: Raising cli- mate change awareness by generating images of floods.arXiv preprint arXiv:2110.02871, 2021

    Pith/arXiv arXiv 2021

  5. [5]

    Climatenerf: Extreme weather synthesis in neural radiance field

    Yuan Li, Zhi-Hao Lin, David Forsyth, Jia-Bin Huang, and Shenlong Wang. Climatenerf: Extreme weather synthesis in neural radiance field. InPro- ceedings of the ieee/cvf international conference on computer vision, pages 3227–3238, 2023

    2023

  6. [6]

    Diffusion models: A comprehensive survey of methods and applications.ACM computing surveys, 56(4):1–39, 2023

    Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications.ACM computing surveys, 56(4):1–39, 2023

    2023

  7. [7]

    Jonathan Ho, Ajay Jain, and P. Abbeel. Denoising diffusion probabilistic models.ArXiv, abs/2006.11239, 2020

    Pith/arXiv arXiv 2006

  8. [8]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  9. [9]

    Gener- ative adversarial nets.Advances in neural information processing systems, 27, 2014

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Gener- ative adversarial nets.Advances in neural information processing systems, 27, 2014

    2014

  10. [10]

    Conditional generative adversarial nets

    Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014

    Pith/arXiv arXiv 2014

  11. [11]

    Astyle-basedgeneratorarchitec- ture for generative adversarial networks

    TeroKarras,SamuliLaine,andTimoAila. Astyle-basedgeneratorarchitec- ture for generative adversarial networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401–4410, 2019

    2019

  12. [12]

    Image- to-image translation with conditional adversarial networks

    Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image- to-image translation with conditional adversarial networks. InProceedings 16 S. Odonchimed et al. of the IEEE conference on computer vision and pattern recognition, pages 1125–1134, 2017

    2017

  13. [13]

    Sdedit: Guided image synthesis and editing with stochastic differential equations.arXiv preprint arXiv:2108.01073, 2021

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations.arXiv preprint arXiv:2108.01073, 2021

    Pith/arXiv arXiv 2021

  14. [14]

    Repaint: Inpainting using denoising diffu- sion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffu- sion probabilistic models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461–11471, 2022

    2022

  15. [15]

    Rad:Region-awarediffusionmodels for image inpainting

    SoraKim,SunghoSuh,andMinsikLee. Rad:Region-awarediffusionmodels for image inpainting. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 2439–2448, 2025

    2025

  16. [16]

    Progressive artwork outpainting via latent diffusion models

    Dae-Young Song, Jung-Jae Yu, and Donghyeon Cho. Progressive artwork outpainting via latent diffusion models. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15405–15415, October 2025

    2025

  17. [17]

    Using simulated data to generate images of climate change.arXiv preprint arXiv:2001.09531, 2020

    Gautier Cosne, Adrien Juraver, Mélisande Teng, Victor Schmidt, Vahe Var- danyan, Alexandra Luccioni, and Yoshua Bengio. Using simulated data to generate images of climate change.arXiv preprint arXiv:2001.09531, 2020

    Pith/arXiv arXiv 2001

  18. [18]

    Flame diffuser: Wildfire image synthesis using mask guided diffusion

    Hao Wang, Sayed Pedram Haeri Boroujeni, Xiwen Chen, Ashish Bastola, Huayu Li, Wenhui Zhu, and Abolfazl Razi. Flame diffuser: Wildfire image synthesis using mask guided diffusion. In2024 IEEE International Confer- ence on Big Data (BigData), pages 6171–6179. IEEE, 2024

    2024

  19. [19]

    FloodscapeDiffuser: Low- rankconditioningfordiffusion-basedpost-floodsatelliteimagerysimulation,

    Khadijah Anwar, Martin Pollack, and Xinqi Yu. FloodscapeDiffuser: Low- rankconditioningfordiffusion-basedpost-floodsatelliteimagerysimulation,

  20. [20]

    Nerf: Representing scenes as neural radi- ance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radi- ance fields for view synthesis.Communications of the ACM, 65(1):99–106, 2021

    2021

  21. [21]

    Climategs: Real-time climate simulationwith3dgaussianstyletransfer.arXiv preprint arXiv:2503.14845, 2025

    Yuezhen Xie, Meiying Zhang, and Qi Hao. Climategs: Real-time climate simulationwith3dgaussianstyletransfer.arXiv preprint arXiv:2503.14845, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Weath- eredit: Controllable weather editing with 4d gaussian field

    Chenghao Qian, Wenjing Li, Yuhu Guo, and Gustav Markkula. Weath- eredit: Controllable weather editing with 4d gaussian field. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 8511– 8519, 2026

    2026

  23. [23]

    Fisheye8k: A benchmark and dataset for fisheye camera object detection

    Munkhjargal Gochoo, Munkh-Erdene Otgonbold, Erkhembayar Ganbold, Jun-Wei Hsieh, Ming-Ching Chang, Ping-Yang Chen, Byambaa Dorj, Hamad Al Jassmi, Ganzorig Batnasan, Fady Alnajjar, et al. Fisheye8k: A benchmark and dataset for fisheye camera object detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5305–5313, 2023

    2023

  24. [24]

    Fishdreamer: PhysFlood 17 Towards fisheye semantic completion via unified image outpainting and seg- mentation

    Hao Shi, Yu Li, Kailun Yang, Jiaming Zhang, Kunyu Peng, Alina Roitberg, Yaozu Ye, Huajian Ni, Kaiwei Wang, and Rainer Stiefelhagen. Fishdreamer: PhysFlood 17 Towards fisheye semantic completion via unified image outpainting and seg- mentation. InProceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR) Workshops, pages 643...

    2023

  25. [25]

    Edge-casesynthesisforfisheyeobject detection: A data-centric perspective.arXiv preprint arXiv:2507.16254, 2025

    SeunghyeonKimandKyeongryeolGo. Edge-casesynthesisforfisheyeobject detection: A data-centric perspective.arXiv preprint arXiv:2507.16254, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Generalized object detection on fisheye cameras for autonomous driving: Dataset, representa- tions and baseline

    Hazem Rashed, Eslam Mohamed, Ganesh Sistu, Varun Ravi Kumar, Cia- ran Eising, Ahmad El-Sallab, and Senthil Yogamani. Generalized object detection on fisheye cameras for autonomous driving: Dataset, representa- tions and baseline. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2272–2280, January 2021

    2021

  27. [27]

    Progressively complementary network for fisheye image rectification using appearance flow

    Shangrong Yang, Chunyu Lin, Kang Liao, Chunjie Zhang, and Yao Zhao. Progressively complementary network for fisheye image rectification using appearance flow. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6348–6357, 2021

    2021

  28. [28]

    Unidepth: Universal monocular metric depth estimation

    Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 10106–10116, 2024

    2024

  29. [29]

    Unidepthv2:Universalmonocularmetric depth estimation made simpler.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    LuigiPiccinelli,ChristosSakaridis,Yung-HsuYang,MattiaSegu,SiyuanLi, WimAbbeloos,andLucVanGool. Unidepthv2:Universalmonocularmetric depth estimation made simpler.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  30. [30]

    Unik3d: Universal camera monocular 3d estimation

    LuigiPiccinelli,ChristosSakaridis,MattiaSegu,Yung-HsuYang,SiyuanLi, Wim Abbeloos, and Luc Van Gool. Unik3d: Universal camera monocular 3d estimation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 1028–1039, 2025

    2025

  31. [31]

    Sam 3: Segment anything with concepts

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Rong- hang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, et al. Sam 3: Segment anything with concepts. arXiv preprint arXiv:2511.16719, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Yolo26: key architectural enhancements and performance benchmarking for real-time object detection.arXiv preprint arXiv:2509.25164, 2025

    Ranjan Sapkota, Rahul Harsha Cheppally, Ajay Sharda, and Manoj Karkee. Yolo26: key architectural enhancements and performance benchmarking for real-time object detection.arXiv preprint arXiv:2509.25164, 2025

    arXiv 2025

  33. [33]

    Efficient RANSAC for Point-Cloud Shape Detection.Computer Graphics Forum, 26(2):214–226, 6 2007

    Ruwen Schnabel, Roland Wahl, and Reinhard Klein. Efficient RANSAC for Point-Cloud Shape Detection.Computer Graphics Forum, 26(2):214–226, 6 2007

    2007

  34. [34]

    Scaling rectified flow transformers for high-resolution image syn- thesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boe- sel, et al. Scaling rectified flow transformers for high-resolution image syn- thesis. InForty-first international conference on machine learning, 2024

    2024

  35. [35]

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, et al. Flux. 1 kontext: Flow matching for in-context image 18 S. Odonchimed et al. generation and editing in latent space.arXiv preprint arXiv:2506.15742, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.Iclr, 1(2):3, 2022

    2022

  37. [37]

    Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Pith/arXiv arXiv 2011

  38. [38]

    Elucidating the design space of diffusion-based generative models.Advances in neural in- formation processing systems, 35:26565–26577, 2022

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models.Advances in neural in- formation processing systems, 35:26565–26577, 2022

    2022

  39. [39]

    Nano Banana 2: Gemini 3.1 Flash Image Model

    Google DeepMind. Nano Banana 2: Gemini 3.1 Flash Image Model. Avail- able via Gemini API and Google DeepMind, 2026. Accessed: June 2026

    2026

  40. [40]

    Is nano banana pro a low-level vision all-rounder? a comprehensive evaluation on 14 tasks and 40 datasets.arXiv preprint arXiv:2512.15110, 2025

    Jialong Zuo, Haoyou Deng, Hanyu Zhou, Jiaxin Zhu, Yicheng Zhang, Yiwei Zhang, Yongxin Yan, Kaixing Huang, Weisen Chen, Yongtai Deng, et al. Is nano banana pro a low-level vision all-rounder? a comprehensive evaluation on 14 tasks and 40 datasets.arXiv preprint arXiv:2512.15110, 2025

    arXiv 2025

  41. [41]

    GPT-Image-2: Next-Generation Image Generation and Editing Model

    OpenAI. GPT-Image-2: Next-Generation Image Generation and Editing Model. Available via OpenAI API and ChatGPT, 2026. Accessed: June 2026

    2026

  42. [42]

    Gpt-image-2 in the wild: A twitter dataset of self-reported ai-generated images from the first week of deployment.arXiv preprint arXiv:2604.25370, 2026

    Kidus Zewde, Simiao Ren, Xingyu Shen, Jiaqi Wu, Yuchen Zhou, Tommy Duong, Zikang Zhang, Ethan Traister, and Kewen Xie. Gpt-image-2 in the wild: A twitter dataset of self-reported ai-generated images from the first week of deployment.arXiv preprint arXiv:2604.25370, 2026

    Pith/arXiv arXiv 2026

  43. [43]

    Hayes and Klaus Krippendorff

    Andrew F. Hayes and Klaus Krippendorff. Answering the call for a standard reliability measure for coding data.Communication Methods and Measures, 1(1):77–89, 2007

    2007

  44. [2025]

    CS231N: Deep Learning for Computer Vision