Pith. sign in

REVIEW 3 major objections 5 minor 34 references

REVIVE restores vandalized AV camera frames by type, then gates them so detection never falls below the unrecovered baseline.

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 · grok-4.5

2026-07-11 04:20 UTC pith:ZNATEN5T

load-bearing objection Solid systems evaluation of type-aware VOA recovery with a real quality-gate result—but the “never worse” guarantee is still an oracle measured against clean frames. the 3 major comments →

arxiv 2607.05649 v1 pith:ZNATEN5T submitted 2026-07-06 cs.CV cs.LG

REVIVE: A Multi-Modal Framework for Vandalism Detection and Recovery in Autonomous Vehicles

Abdullah Tariq Choudhry , Tapadhir Das This is my paper
classification cs.CV cs.LG
keywords physical adversarial attacksautonomous vehiclesvandalismocclusion recoveryimage inpaintingobject detectionquality gatecamera perception
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.

Autonomous-vehicle cameras can be blocked by spray paint, stickers, mud, or covers. Detection alone leaves the perception stream useless. REVIVE is a four-stage pipeline that first flags vandalism, classifies which of five spatial patterns it is, segments the damaged region, and then routes the frame to the recovery method that matches that pattern: adaptive median filtering for scatter noise, generative inpainting for large structured blocks, or direct copy from an aligned clean reference when one exists. The paper’s key operational claim is not that every recovery always improves the image, but that a post-recovery quality gate—checking that recall does not drop, false positives do not rise, and structural similarity stays above a floor—rejects bad candidates so the stream sent to the object detector is never worse than the unrecovered frame. On 500 tracked pairs, unrecovered vandalism cuts detection recall to 0.588; aligned pixel replacement restores it to 0.967, while the gated type-aware path holds recall at 0.608. Generative reconstruction is kept asynchronous because of latency and hallucination risk.

Core claim

Type-aware routing of vandalism-induced occlusions, followed by a reference-available quality gate, yields a camera preprocessing stream whose downstream object-detection performance is never worse than the unrecovered frame, and that can approach clean-frame performance when an aligned clean reference is available.

What carries the argument

REVIVE’s type-aware recovery module plus quality gate: multi-class VOA labels select among median filtering, BLIP-guided Stable Diffusion, or direct pixel replacement; the gate accepts a candidate only if it preserves recall, does not increase false positives, and clears an SSIM floor, otherwise falling back to the unrecovered frame.

Load-bearing premise

The evaluation assumes that synthetic occlusion masks on a driving dataset, together with access to clean reference frames for both replacement and gating, adequately stand in for real paint, mud, and stickers under ordinary vehicle motion.

What would settle it

On paired real-world vandalized and clean camera frames under ordinary driving motion, measure whether the gated type-aware pipeline still keeps per-image detection recall at or above the unrecovered baseline and whether any accepted recovery improves object-level detections without introducing safety-critical hallucinations.

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

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes REVIVE, a four-stage camera-preprocessing pipeline for vandalism-induced occlusion attacks (VOAs) on AV cameras: binary vandalism detection, multi-class VOA pattern classification (random, center-out, top-bottom, bottom-top, targeted), EfficientNet-U-Net segmentation, and type-aware recovery via adaptive median filtering, BLIP-guided Stable Diffusion inpainting, or aligned direct pixel replacement, followed by a quality gate. On BDD100K with synthetic VOAs and 500 tracked clean/vandalized pairs, unrecovered VOAs drop YOLOv8l recall to 0.588; aligned pixel replacement restores recall to 0.967 and F1 to 0.970 (upper bound). Classical/learned inpainters improve SSIM/PSNR more than detection, and Stable Diffusion is demoted to an asynchronous branch. A reference-available quality gate (no recall drop, no FP increase, SSIM≥0.5 vs clean) raises type-aware routing recall from 0.304 without the gate to 0.608 with it, at or above the unrecovered baseline. Online non-generative preprocessing is reported at ~22.7 ms.

Significance. If the results hold under more realistic conditions, the work is a useful systems contribution: it couples VOA-type routing with recovery selection, evaluates restoration by downstream detection rather than image similarity alone, and shows that unfiltered generative recovery can harm perception while a gate can prevent net degradation. Strengths include honest demotion of pixel replacement to a reference upper bound, alignment-sensitivity analysis (Table III), aggregate YOLO TP/FP/FN (Table VI), quality-gate accept rates (Table VII), online latency profiling (Table IV), and a raindrop mask-geometry stress test. The central operational claim—that the forwarded stream is never worse than the unrecovered frame—is currently demonstrated only under a clean-frame oracle gate, which limits deployability claims but does not erase the value of the empirical pipeline and evaluation design.

major comments (3)
  1. [§III-F, §IV-D, §IV-F] §III-F and §IV-D: The strongest operational claim—that type-aware recovery with the quality gate ensures the forwarded stream is never worse than the unrecovered frame (recall 0.608 vs 0.588; without gate 0.304)—is evaluated only with a reference-available gate that uses clean-frame detections and SSIM against the paired clean image. Limitations (§IV-F) state that the reference-free instantiation remains future work. Either reframe the claim as an oracle upper bound on safe recovery selection, or provide a deployable gate proxy (temporal consistency / multi-sensor / confidence) with the same never-worse evaluation; otherwise the deployment guarantee is overstated.
  2. [§IV-A, §IV-C, Limitations] §IV-A and Limitations: All primary VOAs are synthetic programmatic masks (10–30% occlusion) on BDD100K. The raindrop transfer test (§IV-C) only partially addresses real texture/opacity/boundary geometry and lacks paired real vandalism recovery targets. Because routing and recovery choice depend on spatial structure of the five VOA categories, the paper needs either paired real vandalism data or a clearer scope statement that results are synthetic-mask evidence only; this is load-bearing for the claim of a practical AV recovery framework.
  3. [Table I, Table VI, Table VII, §IV-D] Table I / Table VI vs type-aware routing: Under the paper’s own routing (random→median, structured→Stable Diffusion), unfiltered recovery is net-harmful (recall 0.304). LaMa achieves higher aggregate recall (0.667) than gated type-aware routing (0.608) with much higher accept rate (86.4%). The manuscript should justify why type-aware routing to median/SD remains preferred over routing structured VOAs to LaMa (or another learned inpainter) as the primary online candidate, or revise the routing policy and re-report gated end-to-end metrics.
minor comments (5)
  1. [Abstract] Abstract and opening: duplicate wording (“present present” / “This paper presents present”) and a few grammar slips should be cleaned.
  2. [§III, Eq. (1)] Eq. (1) notation is dense; define binary/multi/S/R/OD more cleanly and state that OD is evaluation-only, not part of the recovery controller.
  3. [§III-D, §III-E] Report α, β for L_seg (Eq. 4), training details for binary/multi-class CNNs, and exact Stable Diffusion/BLIP settings for reproducibility.
  4. [Table V] Table V is a single-scene case study; make that more prominent in the caption so it is not read as aggregate evidence (Table VI already serves that role).
  5. [Table I, Table VI] Clarify whether mAP50 and per-image precision/recall use the same matching protocol everywhere, and whether Stable Diffusion’s n=400 subset is reweighted when compared to n=500 methods.

Circularity Check

0 steps flagged

No significant circularity: empirical systems evaluation against held-out clean frames and an external detector, not a derivation that reduces to its own inputs.

full rationale

REVIVE is an engineering pipeline paper (binary/multi-class CNNs, EfficientNet-U-Net, type-aware routing among median filter / LaMa / Telea / Navier-Stokes / BLIP-Stable-Diffusion / aligned pixel copy, plus a post-hoc quality gate). All load-bearing numbers (SSIM/PSNR, YOLOv8l TP/FP/FN, recall 0.588 unrecovered vs 0.967 pixel-replace vs 0.608 gated type-aware) are computed by matching detections and pixels against independent clean-frame references on 500 tracked BDD100K pairs; the gate criteria themselves (recall not below unrecovered, no FP increase, SSIM≥0.5) are evaluation filters, not definitions that force the reported improvement by construction. Direct pixel replacement is explicitly labeled a reference-based upper bound and is stress-tested under deliberate misalignment (Table III). No equation equates a claimed prediction to a fitted quantity; no uniqueness theorem or ansatz is imported via self-citation; citations are to external attack, inpainting, and dataset literature. The acknowledged limitation that the gate is currently reference-available (reference-free instantiation left as future work) is a deployment gap, not circular derivation. The paper is therefore self-contained against its stated external benchmarks.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central claim rests on an empirical pipeline and evaluation protocol rather than a closed-form derivation. Load-bearing inputs are the synthetic VOA taxonomy, the assumption that clean-frame YOLO detections define recovery success, hand-chosen gate thresholds, and the availability of aligned clean references for the upper-bound and gate evaluation. No new physical entity is postulated; the invented pieces are operational constructs (REVIVE stages, VOA categories, gate criteria).

free parameters (5)
  • quality-gate SSIM floor = 0.5
    Candidates must clear SSIM ≥ 0.5 before acceptance; this threshold is chosen by the authors and directly controls accept/reject behavior.
  • VOA occlusion fraction range = 10-30%
    Synthetic masks use 10–30% occlusion; recovery and detection numbers depend on this chosen corruption severity band.
  • adaptive median dark threshold bounds = clamp [10,50]; kernels 3-9
    T_adaptive = max(10, min(50, μ_noise − 2σ_noise)) and multi-scale kernels [3×3…9×9] are hand-designed controls for the random-VOA branch.
  • segmentation loss weights α, β
    L_seg = α L_BCE + β L_Dice; weights are free design choices affecting mask quality and thus recovery.
  • detection matching IoU threshold = 0.50
    Recovered detections are matched to clean-frame detections at IoU ≥ 0.50; this defines reported recall/F1/mAP50.
axioms (5)
  • domain assumption Threat model: attacker can physically occlude one camera lens but cannot modify AV software, stored frames, or other sensors.
    Stated in §III-A; scopes the problem to image-level recovery rather than full stack compromise.
  • ad hoc to paper Five programmatic VOA categories (random, center-out, top-bottom, bottom-top, targeted) adequately represent practical vandalism spatial structure for recovery routing.
    Defined in §III-C and used throughout training and type-aware branch selection.
  • domain assumption Clean-frame YOLOv8l detections are a valid external reference for whether recovery restores perception-relevant objects.
    §III-G and Tables I/VI; recovery success is defined relative to this detector behavior.
  • domain assumption An aligned clean reference frame can exist for short-horizon upper-bound pixel replacement and for reference-available gate evaluation.
    §III-E3 and §III-F; authors also show the assumption fails under modest shifts (Table III).
  • standard math Standard image-processing and deep-learning training assumptions (BCE/CCE/Dice losses, ImageNet normalization, BDD100K as driving proxy) hold for this task.
    Used in Stages 1–3 and experimental setup without novel theoretical derivation.
invented entities (3)
  • REVIVE pipeline (binary detect → multi-class VOA type → conditional U-Net mask → type-aware recovery → quality gate) no independent evidence
    purpose: Organize end-to-end camera-stream recovery for VOAs and enforce no-worse-than-unrecovered forwarding.
    The named framework and routing logic are the paper’s operational construct; components are known, the composition is new.
  • Five-class VOA pattern taxonomy for recovery routing no independent evidence
    purpose: Map occlusion spatial structure to median filter, generative inpainting, or pixel replacement.
    Categories are defined by the authors for this pipeline rather than taken from a standard vandalism ontology.
  • Reference-available recovery quality gate (no recall drop, no FP increase, SSIM floor) no independent evidence
    purpose: Filter recovered candidates so the forwarded stream is never worse than the unrecovered frame.
    Gate criteria are paper-defined; evaluated with clean references, with reference-free form left to future work.

pith-pipeline@v1.1.0-grok45 · 20378 in / 3971 out tokens · 46115 ms · 2026-07-11T04:20:51.606428+00:00 · methodology

0 comments
read the original abstract

Autonomous vehicles (AVs) face increasing threats from vandalism-induced occlusion attacks (VOAs) that compromise camera-based perception. While detection frameworks can identify vandalized images, restoring camera-stream utility after physical occlusion remains underexplored. This paper presents present the Recovery and Enhancement of Vandalized Images for Vision Excellence (REVIVE) framework, a vandalism recovery pipeline integrating: (1) binary VOA detection, (2) multi-class VOA pattern identification, (3) EfficientNet-based U-Net segmentation, and (4) type-aware recovery using Bootstrapping Language-Image Pre-training (BLIP)-guided Stable Diffusion inpainting, direct pixel replacement, or adaptive median filtering. Stable Diffusion shows variable reconstruction performance (per-pattern SSIM 0.667-0.867, PSNR 15.4-26.7dB) across VOA patterns, while aligned direct pixel replacement achieves near-identical reconstruction under the aligned-reference condition. On 500 tracked clean/vandalized image pairs, unrecovered VOAs reduce YOLOv8l object-detection recall to 0.588, while direct pixel replacement restores recall to 0.967 and F1-score to 0.970 under that aligned-reference condition. LaMa, Telea, and Navier-Stokes baselines improve image similarity but provide more limited downstream detection recovery, and Stable Diffusion is treated as an asynchronous recovery branch subject to a quality gate rather than a blocking real-time perception step. We evaluate a reference-available quality gate that filters recovered candidates before downstream use: without it, type-aware routing degrades per-image recall to 0.304, whereas with it, recall returns to 0.608, at or above the unrecovered baseline, ensuring the forwarded stream is never worse than the unrecovered frame. REVIVE therefore, provides a structured recovery framework from VOAs in AVs.

Figures

Figures reproduced from arXiv: 2607.05649 by Abdullah Tariq Choudhry, Tapadhir Das.

Figure 1
Figure 1. Figure 1: Example camera-frame degradation from a VOA, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: REVIVE pipeline architecture. Clean frames bypass recovery; vandalized frames are classified, segmented, and recovered [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Five VOA categories used in the framework. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Binary vandalism detector performance [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Multi-class VOA detector performance achieves 99% precision and 100% recall, and targeted attacks achieve 99% precision and 97% recall. This classification accuracy enables precise recovery strategy selection further down the REVIVE pipeline. Following detection and classification, the EfficientNet￾based U-Net model achieves strong performance as shown in [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Final epoch performance of EfficientNet-U-Net [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Segmentation results showing vandalized images (left) [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Representative bottom-top VOA recovery example showing clean reference, vandalized input, generative recovery, and [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Representative recovery examples for center-out and targeted VOAs. Center-out attacks obscure semantically important [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Representative recovery examples for random and top-bottom VOAs. Random attacks are treated as noise-like [PITH_FULL_IMAGE:figures/full_fig_p008_10.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

34 extracted references · 4 linked inside Pith

  1. [1]

    alliedmarketresearch.com/autonomous-vehicle-market, Jan 2025

    Allied Market Research, “Autonomous vehicle market by level of automation (level 3, level 4, level 5), by component (hardware, soft- ware), by vehicle type (passenger cars, commercial vehicles): Global opportunity analysis and industry forecast, 2030-2040.” https://www. alliedmarketresearch.com/autonomous-vehicle-market, Jan 2025. Ac- cessed: October 20, 2025

    2030

  2. [2]

    Remote attacks on automated vehicles sensors: Experiments on camera and lidar,

    J. Petit, B. Stottelaar, M. Feiri, and F. Kargl, “Remote attacks on automated vehicles sensors: Experiments on camera and lidar,”Black Hat Europe, vol. 11, no. 2015, p. 995, 2015

    2015

  3. [3]

    A crowd destroyed a waymo robotaxi in san francisco

    J. Kastrenakes, “A crowd destroyed a waymo robotaxi in san francisco.” The Verge, Feb 2024. Accessed: September 29, 2025

    2024

  4. [4]

    Video shows gang tagging waymo robotaxis with passengers inside

    G. Leahy, “Video shows gang tagging waymo robotaxis with passengers inside.” The San Francisco Standard, Sep 2024. Accessed: September 29, 2025

    2024

  5. [5]

    Deep learning sensor fusion for autonomous vehicle perception and localization: A review,

    J. Fayyad, M. A. Jaradat, D. Gruyer, and H. Najjaran, “Deep learning sensor fusion for autonomous vehicle perception and localization: A review,”Sensors, vol. 20, no. 15, p. 4220, 2020

    2020

  6. [6]

    A survey on sensor failures in autonomous vehicles: Challenges and solutions,

    F. Matos, J. Bernardino, J. Dur ˜aes, and J. Cunha, “A survey on sensor failures in autonomous vehicles: Challenges and solutions,”Sensors, vol. 24, no. 16, p. 5108, 2024

    2024

  7. [7]

    Efficientnet: Rethinking model scaling for con- volutional neural networks,

    M. Tan and Q. Le, “Efficientnet: Rethinking model scaling for con- volutional neural networks,” inInternational conference on machine learning, pp. 6105–6114, PMLR, 2019

    2019

  8. [8]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,

    J. Li, D. Li, C. Xiong, and S. Hoi, “Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,” inInternational conference on machine learning, pp. 12888–12900, PMLR, 2022

    2022

  9. [9]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pp. 10684–10695, 2022

    2022

  10. [10]

    Intriguing properties of neural networks,

    C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan, I. Goodfellow, and R. Fergus, “Intriguing properties of neural networks,”arXiv preprint arXiv:1312.6199, 2013

    Pith/arXiv arXiv 2013

  11. [11]

    Explaining and harnessing adversarial examples,

    I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and harnessing adversarial examples,”arXiv preprint arXiv:1412.6572, 2014

    Pith/arXiv arXiv 2014

  12. [12]

    Robust physical-world attacks on deep learning visual classification,

    K. Eykholt, I. Evtimov, E. Fernandes, B. Li, A. Rahmati, C. Xiao, A. Prakash, T. Kohno, and D. Song, “Robust physical-world attacks on deep learning visual classification,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 1625–1634, 2018

    2018

  13. [13]

    Adversarial patch,

    T. B. Brown, D. Man ´e, A. Roy, M. Abadi, and J. Gilmer, “Adversarial patch,”arXiv preprint arXiv:1712.09665, 2017

    Pith/arXiv arXiv 2017

  14. [14]

    Shapeshifter: Robust physical adversarial attack on faster r-cnn object detector,

    S.-T. Chen, C. Cornelius, J. Martin, and D. H. Chau, “Shapeshifter: Robust physical adversarial attack on faster r-cnn object detector,” inJoint European Conference on Machine Learning and Knowledge Discovery in Databases, pp. 52–68, Springer, 2018

    2018

  15. [15]

    Fooling automated surveil- lance cameras: adversarial patches to attack person detection,

    S. Thys, W. Van Ranst, and T. Goedem ´e, “Fooling automated surveil- lance cameras: adversarial patches to attack person detection,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pp. 0–0, 2019

    2019

  16. [16]

    Lavan: Localized and visible adversarial noise,

    D. Karmon, D. Zoran, and Y . Goldberg, “Lavan: Localized and visible adversarial noise,” inInternational conference on machine learning, pp. 2507–2515, PMLR, 2018

    2018

  17. [17]

    Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition,

    M. Sharif, S. Bhagavatula, L. Bauer, and M. K. Reiter, “Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition,” inProceedings of the 2016 ACM SIGSAC conference on computer and communications security, pp. 1528–1540, 2016

    2016

  18. [18]

    Towards deep learning models resistant to adversarial attacks,

    A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu, “Towards deep learning models resistant to adversarial attacks,”arXiv preprint arXiv:1706.06083, 2017

    Pith/arXiv arXiv 2017

  19. [19]

    Distillation as a defense to adversarial perturbations against deep neural networks,

    N. Papernot, P. McDaniel, X. Wu, S. Jha, and A. Swami, “Distillation as a defense to adversarial perturbations against deep neural networks,” in2016 IEEE Symposium on Security and Privacy (SP), pp. 582–597, IEEE, 2016

    2016

  20. [20]

    Context encoders: Feature learning by inpainting,

    D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, and A. A. Efros, “Context encoders: Feature learning by inpainting,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 2536– 2544, 2016

    2016

  21. [21]

    Generative image inpainting with contextual attention,

    J. Yu, Z. Lin, J. Yang, X. Shen, X. Lu, and T. S. Huang, “Generative image inpainting with contextual attention,”Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 5505–5514, 2018

    2018

  22. [22]

    Edge- connect: Structure guided image inpainting using edge prediction,

    K. Nazeri, E. Ng, T. Joseph, F. Z. Qureshi, and M. Ebrahimi, “Edge- connect: Structure guided image inpainting using edge prediction,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, Oct 2019

    2019

  23. [23]

    Resolution-robust large mask inpainting with fourier convolutions,

    R. Suvorov, E. Logacheva, A. Mashikhin, A. Remizova, A. Ashukha, A. Silvestrov, N. Kong, H. Goka, K. Park, and V . Lempitsky, “Resolution-robust large mask inpainting with fourier convolutions,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 2149–2159, 2022

    2022

  24. [24]

    Image inpainting for irregular holes using partial convolutions,

    G. Liu, F. A. Reda, K. J. Shih, T.-C. Wang, A. Tao, and B. Catanzaro, “Image inpainting for irregular holes using partial convolutions,” in Proceedings of the European conference on computer vision (ECCV), pp. 85–100, 2018

    2018

  25. [25]

    Free- form image inpainting with gated convolution,

    J. Yu, Z. Lin, J. Yang, X. Shen, X. Lu, and T. S. Huang, “Free- form image inpainting with gated convolution,” inProceedings of the IEEE/CVF international conference on computer vision, pp. 4471–4480, 2019

    2019

  26. [26]

    Repaint: Inpainting using denoising diffusion probabilistic models,

    A. Lugmayr, M. Danelljan, L. V . Gool, and R. Timofte, “Repaint: Inpainting using denoising diffusion probabilistic models,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11461–11471, 2022

    2022

  27. [27]

    Palette: Image-to-image diffusion models,

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. Denton, K. S. Ghasemipour, B. K. Ayan, S. S. Mahdavi, R. G. Lopes,et al., “Palette: Image-to-image diffusion models,” inACM SIGGRAPH 2022 Confer- ence Proceedings, pp. 1–10, 2022

    2022

  28. [28]

    Ultralytics yolov8,

    G. Jocheret al., “Ultralytics yolov8,” 2023. Accessed: 2025-10-20

    2023

  29. [29]

    Image quality assessment: From error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,”IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  30. [30]

    The pascal visual object classes (voc) challenge,

    M. Everingham, L. Van Gool, C. K. Williams, J. Winn, and A. Zisser- man, “The pascal visual object classes (voc) challenge,”International journal of computer vision, vol. 88, no. 2, pp. 303–338, 2010

    2010

  31. [31]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning,

    F. Yu, H. Chen, X. Wang, W. Xian, Y . Chen, F. Liu, V . Madhavan, and T. Darrell, “Bdd100k: A diverse driving dataset for heterogeneous multitask learning,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 2636–2645, 2020

    2020

  32. [32]

    Are we ready for autonomous driving? the kitti vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” in2012 IEEE conference on computer vision and pattern recognition, pp. 3354–3361, IEEE, 2012

    2012

  33. [33]

    nuscenes: A multi- modal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Krishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multi- modal dataset for autonomous driving,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11621– 11631, 2020

    2020

  34. [34]

    Raindrops on windshield dataset,

    V . Soboleva and O. Shipitko, “Raindrops on windshield dataset,” 2021

    2021