pith. sign in

arxiv: 2605.20436 · v1 · pith:PVTSUWDGnew · submitted 2026-05-19 · 💻 cs.CV

Lighting-aware Unified Model for Instance Segmentation

Qisai Liu , Alloy Das , Zhanhong Jiang , Joshua R. Waite , Aditya Balu , Adarsh Krishnamurthy , Soumik Sarkar This is my paper

Pith reviewed 2026-05-21 07:00 UTC · model grok-4.3

classification 💻 cs.CV
keywords instance segmentationlighting robustnesscontrast mapsadapter moduleSAMsynthetic datasetdual-branch architectureillumination invariance
0
0 comments X

The pith

A contrast-map adapter makes instance segmentation robust to real-world lighting without retraining the backbone.

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

Foundation models like SAM lose accuracy when illumination varies across real scenes. This paper adds a lightweight Lighting Convolutional-Attention adapter that runs a second branch on contrast maps derived from the input. The adapter is trained on paired clean and illuminated images using a loss that penalizes output differences between the pairs. A new Unity-generated synthetic dataset supplies controlled lighting variants for training and testing. The result is a model that focuses on structural edges instead of brightness shifts.

Core claim

Lighting Convolutional-Attention (LCA) is an adapter module with a dual-branch architecture that processes RGB features alongside contrast maps. It is optimized through a pairwise training strategy that introduces a targeted loss term penalizing discrepancies between clean images and their illumination variants. A novel Unity-based synthetic dataset replicates complex real-world lighting conditions to support training and evaluation of the architecture.

What carries the argument

Lighting Convolutional-Attention (LCA) adapter, a dual-branch module that combines RGB features with contrast maps to produce lighting-invariant representations for the segmentation head.

If this is right

  • The adapter delivers superior lighting-robust instance segmentation on multiple existing benchmarks.
  • The method bridges the domain gap between synthetic lighting variants and real-world illumination conditions.
  • Performance gains occur without any fine-tuning of the underlying heavy foundation-model backbone.

Where Pith is reading between the lines

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

  • Contrast-map branches could be added to other vision tasks such as detection or depth estimation to gain similar lighting invariance.
  • The pairwise training approach may reduce the volume of real-world annotated data needed when adapting models to new environments.
  • Deployment in outdoor robotics or autonomous vehicles could become more reliable if the adapter proves stable across seasons and weather.

Load-bearing premise

The contrast maps enable physically motivated sensitivity to structural changes rather than illumination artifacts, allowing the dual-branch architecture to generalize from synthetic lighting variants to real-world conditions.

What would settle it

A clear drop in segmentation accuracy on real captured images whose lighting conditions fall outside the range simulated in the Unity dataset would show that the claimed generalization does not hold.

Figures

Figures reproduced from arXiv: 2605.20436 by Adarsh Krishnamurthy, Aditya Balu, Alloy Das, Joshua R. Waite, Qisai Liu, Soumik Sarkar, Zhanhong Jiang.

Figure 1
Figure 1. Figure 1: PLAP-LCA. Clean and variant images share a SAM ViT-B encoder with LCA modules injected into the final two blocks. Only LCA weights, scalar gates, and the mask decoder are trained through supervise and consistency losses. Please see the Supplementary Materials for concrete details of the whole framework. tecture. While SAM’s core design consists of a ViT-based image encoder, a lightweight prompt encoder, an… view at source ↗
Figure 2
Figure 2. Figure 2: Unity synthetic data pipeline overview. Top: The lighting rig exposes five physically-grounded parameters manipulated directly in the 3D scene. Middle: Each iteration randomises 16 camera positions then records instance-segmentation masks. Bottom: Every viewpoint is re-rendered under four lighting conditions Clean, Mild, Moderate, and Severe construct pairwise training data with the perfect ground truth ma… view at source ↗
Figure 3
Figure 3. Figure 3: Lighting Convolutional Attention (LCA) module. Channel, spatial, and Lapla￾cian contrast gates are multiplied element-wise and projected separable convolution, then fused through a learned scalar gate. captures the typical activation energy across the spatial extent, and a maximum that captures the strongest localized response. Both are obtained directly from global adaptive pooling operations: davg = AvgP… view at source ↗
Figure 4
Figure 4. Figure 4: Quantitative IoU analysis comparing the SAM baseline and LCA in Cityscape lighting variant. The histogram on the left shows the overall distribution shift; the scatter plot on the right shows per-instance results, where points above the y=x indicate LCA improvement [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sheep in a low-contrast outdoor scene. The baseline (IoU = 0.007) collapses into a near-total false-positive flood across the hillside; LCA (IoU = 0.585) isolates the small target instance despite its low contrast against the surrounding terrain. evaluate four resolutions: 128 × 128, 256 × 256, 512 × 512, and 1024 × 1024. Interestingly, the relationship between resolution and mIoU is non-monotonic. The bes… view at source ↗
Figure 6
Figure 6. Figure 6: On the top is the lighting variant image under GradCam. LCA achieves IoU = 0.268, comparing to the baseline only gets 0.112. LCA concentrates attention on the targeted area, but the decoder only diffuses with the background. between Clean and Variant conditions remains stable (∼0.005) across all res￾olutions, indicating that the lighting robustness of our approach is resolution￾agnostic. LCA Architectural … view at source ↗
Figure 7
Figure 7. Figure 7: Detailed PLAP-LCA training pipeline. Clean and variant images are processed by a shared, frozen ViT-B encoder with LCA modules at the last two blocks. The prompts remain unchanged and are shared with both streams. The supervised loss trains against shared ground-truth masks, while the consistency loss penalizes any prediction divergence caused purely by illumination [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: LCA module architecture.The three gates run in parallel. The channel gate (left) uses global average and max pool statistics fed through a shared MLP to produce a per-channel scalar mask. The spatial gate (center) concatenates channel-wise mean and max maps and convolves them with a 7×7 kernel to produce a per-location scalar mask. The contrast gate (right) projects the feature map to a single-channel gray… view at source ↗
Figure 9
Figure 9. Figure 9: Robustness analysis on COCO across all three severity levels. (a) Average mIoU drop (clean − variant) per severity. (b) Per-image drop distributions; wider spread indicates more variable degradation. (c) Pooled KDE across all severities; a sharper peak at zero indicates greater lighting robustness. (a) IoU histogram — COCO (lighting variant). (b) Per-instance scatter — COCO [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 10
Figure 10. Figure 10: IoU analysis on COCO under lighting variant conditions. Left: distribution shift between SAM baseline and LCA. Right: per-instance comparison; points above y=x indicate LCA improvement [PITH_FULL_IMAGE:figures/full_fig_p027_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: IoU analysis on VOC under lighting variant conditions. (a) IoU histogram — Unity (lighting variant). (b) Per-instance scatter — Unity [PITH_FULL_IMAGE:figures/full_fig_p028_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: IoU analysis on the Unity synthetic dataset under physically rendered lighting variant conditions. Because Unity lighting is rendered rather than augmented, the performance gap is larger, reflecting the additional domain shift [PITH_FULL_IMAGE:figures/full_fig_p028_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Cityscapes (road, sev1, ∆IoU = +0.761). The baseline floods the entire road surface with high confidence (IoU = 0.060, conf = 0.98); LCA recovers a compact mask tightly aligned with the target lane segment (IoU = 0.821) [PITH_FULL_IMAGE:figures/full_fig_p028_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Cityscapes (sky, sev2, ∆IoU = +0.725). Under moderate lighting, the base￾line misidentifies large building façades as sky (IoU = 0.042); LCA correctly isolates the visible sky region (IoU = 0.767) [PITH_FULL_IMAGE:figures/full_fig_p029_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: COCO (truck, sev2, ∆IoU = +0.756). The baseline activates on the lower-half background rather than the vehicle body (IoU = 0.046); LCA produces a well-bounded truck mask (IoU = 0.802), demonstrating robustness to the high-contrast reflective win￾dow surface [PITH_FULL_IMAGE:figures/full_fig_p029_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: COCO (person, sev2, ∆IoU = +0.700). Low ambient stadium lighting causes the baseline to flood a rectangular region of the court (IoU = 0.030); LCA correctly segments the athlete’s silhouette (IoU = 0.729) [PITH_FULL_IMAGE:figures/full_fig_p029_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Unity (object, sev1, ∆IoU = +0.686). In a near-dark synthetic indoor scene, the baseline produces a large false-positive region unrelated to the target object (IoU = 0.128); LCA isolates the dimly-lit target using structural edge cues (IoU = 0.814) [PITH_FULL_IMAGE:figures/full_fig_p030_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Unity (object, sev1, ∆IoU = +0.663). Warm ceiling lighting creates a strong false edge that misleads the baseline into activating across the wall and furniture (IoU = 0.123); LCA suppresses the photometric distraction and recovers the target boundary (IoU = 0.786) [PITH_FULL_IMAGE:figures/full_fig_p030_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: COCO scene. Row 1: clean image and three severity levels (mild, moderate, severe). Row 2: corresponding ground-truth instance masks [PITH_FULL_IMAGE:figures/full_fig_p031_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: COCO scene (second example) [PITH_FULL_IMAGE:figures/full_fig_p031_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Cityscapes scene [PITH_FULL_IMAGE:figures/full_fig_p032_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Cityscapes scene (second example) [PITH_FULL_IMAGE:figures/full_fig_p032_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Unity scene [PITH_FULL_IMAGE:figures/full_fig_p032_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Unity scene (second example) [PITH_FULL_IMAGE:figures/full_fig_p032_24.png] view at source ↗
read the original abstract

Foundation models like the Segment Anything Model (SAM) demonstrate impressive zero-shot generalization but frequently degrade under diverse real-world illumination, particularly for instance segmentation. In this work, we address this limitation by developing \textit{Lighting Convolutional-Attention (\lca{})}, an adapter module that enhances segmentation robustness without fine-tuning the heavy backbone. \lca{} employs a dual-branch architecture to process RGB features alongside contrast maps, enabling physically motivated sensitivity to structural changes rather than illumination artifacts. We optimize \lca{} through a pairwise training strategy, introducing a targeted loss term that explicitly penalizes discrepancies between clean images and their corresponding illumination variants. To evaluate and support this architecture, we conduct a comprehensive empirical study across multiple existing benchmarks and present a novel Unity-based synthetic dataset specifically designed to accurately replicate complex real-world lighting conditions. Extensive experimental results demonstrate that our approach successfully bridges the domain gap, delivering superior lighting-robust segmentation.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes the Lighting Convolutional-Attention (LCA) adapter module to improve the lighting robustness of foundation models such as SAM for instance segmentation. LCA uses a dual-branch architecture that processes RGB features together with contrast maps, is trained via a pairwise strategy with a loss that penalizes discrepancies between clean images and their illumination variants, and is evaluated on existing benchmarks plus a new Unity-generated synthetic dataset designed to replicate complex real-world lighting.

Significance. If the reported gains hold under scrutiny, the work would provide a practical, lightweight adaptation strategy for making segmentation models more reliable under uncontrolled illumination without retraining the backbone. The new synthetic dataset could serve as a useful resource for the community studying domain gaps induced by lighting.

major comments (3)
  1. [§3.2] §3.2 (LCA architecture): The claim that contrast maps enable 'physically motivated sensitivity to structural changes rather than illumination artifacts' is central to the dual-branch design but is presented without a derivation, photometric justification, or explicit formula for how the contrast map is computed from the input; this leaves the physical motivation as an assumption rather than a demonstrated property.
  2. [§4] §4 (Experiments): While the abstract and introduction assert 'extensive experimental results' and 'superior lighting-robust segmentation' across benchmarks, the manuscript provides no quantitative metrics, error bars, statistical tests, or dataset statistics in the summary sections; without these, the magnitude and reliability of the claimed domain-gap bridging cannot be assessed.
  3. [§3.3] §3.3 (Pairwise training): The targeted loss term that penalizes discrepancies between clean and illumination-variant pairs is load-bearing for the training strategy, yet the manuscript does not specify its exact functional form, weighting relative to the segmentation loss, or ablation isolating its contribution versus the contrast-map branch alone.
minor comments (3)
  1. [§3] The notation for the LCA module and its components should be introduced with a clear diagram or pseudocode early in §3 to aid readability.
  2. [Figure 3] Figure captions for the synthetic dataset examples should explicitly state the range of lighting parameters varied (e.g., light source positions, intensities) so readers can judge how well they approximate real-world conditions.
  3. [§2] A few references to prior work on contrast-based illumination invariance (e.g., in photometric stereo or Retinex theory) appear to be missing from the related-work section.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, clarifying aspects of the manuscript and outlining the revisions we will make.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (LCA architecture): The claim that contrast maps enable 'physically motivated sensitivity to structural changes rather than illumination artifacts' is central to the dual-branch design but is presented without a derivation, photometric justification, or explicit formula for how the contrast map is computed from the input; this leaves the physical motivation as an assumption rather than a demonstrated property.

    Authors: We agree that the physical motivation benefits from explicit support. The contrast map is intended to emphasize local structural variations by normalizing against local intensity statistics. In the revised manuscript we will add the precise computation formula for the contrast map, a short derivation drawing on photometric principles of local contrast, and an explanation of why this formulation reduces sensitivity to global illumination shifts while preserving edge and texture information. revision: yes

  2. Referee: [§4] §4 (Experiments): While the abstract and introduction assert 'extensive experimental results' and 'superior lighting-robust segmentation' across benchmarks, the manuscript provides no quantitative metrics, error bars, statistical tests, or dataset statistics in the summary sections; without these, the magnitude and reliability of the claimed domain-gap bridging cannot be assessed.

    Authors: We acknowledge that the abstract and introduction currently lack numerical summaries. The full experimental section and supplementary material already contain mAP / mIoU tables with standard deviations across multiple runs, statistical significance tests, and statistics for the Unity dataset (image count, lighting variation parameters). In the revision we will insert concise quantitative highlights and dataset descriptors into the abstract and introduction so that the claimed improvements are immediately quantifiable. revision: yes

  3. Referee: [§3.3] §3.3 (Pairwise training): The targeted loss term that penalizes discrepancies between clean and illumination-variant pairs is load-bearing for the training strategy, yet the manuscript does not specify its exact functional form, weighting relative to the segmentation loss, or ablation isolating its contribution versus the contrast-map branch alone.

    Authors: We thank the referee for noting this omission. The targeted loss is a pairwise discrepancy term between feature representations of clean and illumination-augmented images. In the revised §3.3 we will state its exact functional form as an equation, specify the relative weighting hyper-parameter with respect to the primary segmentation loss, and add an ablation experiment that isolates the pairwise loss contribution while holding the contrast-map branch fixed. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical architecture and benchmarks are self-contained

full rationale

The paper proposes an LCA adapter with dual-branch processing of RGB and contrast maps, a pairwise loss penalizing illumination variants, and a new Unity synthetic dataset. These elements are trained and evaluated on existing benchmarks plus the new data, with performance gains reported directly from experiments. No equations, predictions, or uniqueness claims reduce by construction to fitted parameters or self-citations; the central claim rests on empirical validation rather than any self-referential derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that contrast maps isolate structural information independent of illumination and that the synthetic dataset faithfully represents real lighting variations.

axioms (1)
  • domain assumption Contrast maps capture structural changes independent of illumination artifacts
    Invoked in the description of the dual-branch architecture processing RGB features alongside contrast maps.
invented entities (1)
  • Lighting Convolutional-Attention (LCA) adapter module no independent evidence
    purpose: Enhance segmentation robustness to lighting without backbone fine-tuning
    New module introduced with dual-branch design and targeted loss.

pith-pipeline@v0.9.0 · 5702 in / 1225 out tokens · 42245 ms · 2026-05-21T07:00:23.133034+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 3 internal anchors

  1. [1]

    arXiv preprint arXiv:2505.09274 (2025)

    Bougourzi, F., Hadid, A.: Recent advances in medical imaging segmentation: A survey. arXiv preprint arXiv:2505.09274 (2025)

    work page arXiv 2025

  2. [2]

    Diagnostics15(21), 2762 (2025)

    Chankhachon,S.,Kansomkeat,S.,Bhurayanontachai,P.,Intajag,S.:Deeplearning network with illuminant augmentation for diabetic retinopathy segmentation using comprehensive anatomical context integration. Diagnostics15(21), 2762 (2025)

    work page 2025

  3. [3]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chaturvedi, S., Ren, M., Hold-Geoffroy, Y., Liu, J., Dorsey, J., Shu, Z.: Synthlight: Portrait relighting with diffusion model by learning to re-render synthetic faces. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 369–379 (2025)

    work page 2025

  4. [4]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Chen, T., Zhu, L., Deng, C., Cao, R., Wang, Y., Zhang, S., Li, Z., Sun, L., Zang, Y., Mao, P.: Sam-adapter: Adapting segment anything in underperformed scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3367–3375 (2023)

    work page 2023

  5. [5]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., Schiele, B.: The cityscapes dataset for semantic urban scene understanding. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 3213–3223 (2016)

    work page 2016

  6. [6]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Cubuk,E.D.,Zoph,B.,Mane,D.,Vasudevan,V.,Le,Q.V.:Autoaugment:Learning augmentation strategies from data. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 113–123 (2019)

    work page 2019

  7. [7]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops

    Cubuk, E.D., Zoph, B., Shlens, J., Le, Q.V.: Randaugment: Practical automated data augmentation with a reduced search space. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 702–703 (2020)

    work page 2020

  8. [8]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  9. [9]

    Everingham, M., Eslami, S.A., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.:Thepascalvisualobjectclasseschallenge:Aretrospective.Internationaljournal of computer vision111(1), 98–136 (2015)

    work page 2015

  10. [10]

    Benchmarking Neural Network Robustness to Common Corruptions and Perturbations

    Hendrycks,D.,Dietterich,T.:Benchmarkingneuralnetworkrobustnesstocommon corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)

    work page internal anchor Pith review Pith/arXiv arXiv 1903

  11. [11]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Hoyer, L., Dai, D., Van Gool, L.: Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9924– 9935 (2022)

    work page 2022

  12. [12]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Hu, J., Shen, L., Sun, G.: Squeeze-and-excitation networks. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7132–7141 (2018)

    work page 2018

  13. [13]

    In: Proceedings of the IEEE international conference on computer vision

    Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501–1510 (2017)

    work page 2017

  14. [14]

    arXiv e-prints pp

    Jin, S., Wang, L., Temming, B., Pokorny, F.T.: Physically-based lighting augmen- tation for robotic manipulation. arXiv e-prints pp. arXiv–2508 (2025)

    work page 2025

  15. [15]

    Advances in Neural Information Processing Systems36, 29914–29934 (2023) 16 Q

    Ke, L., Ye, M., Danelljan, M., Tai, Y.W., Tang, C.K., Yu, F., et al.: Segment anything in high quality. Advances in Neural Information Processing Systems36, 29914–29934 (2023) 16 Q. Liu et al

    work page 2023

  16. [16]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4015–4026 (2023)

    work page 2023

  17. [17]

    Advances in neural information processing systems29(2016)

    Kondor, R., Pan, H.: The multiscale laplacian graph kernel. Advances in neural information processing systems29(2016)

    work page 2016

  18. [18]

    In: European conference on computer vision

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: European conference on computer vision. pp. 740–755. Springer (2014)

    work page 2014

  19. [19]

    In: Proceedings of the european conference on computer vision (ECCV)

    Pan, X., Luo, P., Shi, J., Tang, X.: Two at once: Enhancing learning and gen- eralization capacities via ibn-net. In: Proceedings of the european conference on computer vision (ECCV). pp. 464–479 (2018)

    work page 2018

  20. [20]

    SAM 2: Segment Anything in Images and Videos

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., et al.: Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  21. [21]

    In: Pro- ceedings of the IEEE/CVF international conference on computer vision

    Sakaridis, C., Dai, D., Gool, L.V.: Guided curriculum model adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. In: Pro- ceedings of the IEEE/CVF international conference on computer vision. pp. 7374– 7383 (2019)

    work page 2019

  22. [22]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Sakaridis, C., Dai, D., Van Gool, L.: Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10765–10775 (2021)

    work page 2021

  23. [23]

    Engineering, Technology & Applied Science Research15(6), 30119–30129 (2025)

    Sermwuthisarn, P., Phumeechanya, S.: Integration of u-net and fastsam for accu- rate leaf image segmentation in complex backgrounds. Engineering, Technology & Applied Science Research15(6), 30119–30129 (2025)

    work page 2025

  24. [24]

    Journal of big data6(1), 1–48 (2019)

    Shorten, C., Khoshgoftaar, T.M.: A survey on image data augmentation for deep learning. Journal of big data6(1), 1–48 (2019)

    work page 2019

  25. [25]

    Advances in neural information processing systems34, 237–250 (2021)

    Wang, H., Xiao, C., Kossaifi, J., Yu, Z., Anandkumar, A., Wang, Z.: Augmax: Adversarial composition of random augmentations for robust training. Advances in neural information processing systems34, 237–250 (2021)

    work page 2021

  26. [26]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wang, J., Liu, J., Sun, X., Singh, K.K., Shu, Z., Zhang, H., Yang, J., Zhao, N., Wang, T.Y., Chen, S.S., et al.: Comprehensive relighting: Generalizable and con- sistent monocular human relighting and harmonization. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 380–390 (2025)

    work page 2025

  27. [27]

    Neurocomputing 312, 135–153 (2018)

    Wang, M., Deng, W.: Deep visual domain adaptation: A survey. Neurocomputing 312, 135–153 (2018)

    work page 2018

  28. [28]

    In: Proceedings of the European conference on computer vision (ECCV)

    Woo, S., Park, J., Lee, J.Y., Kweon, I.S.: Cbam: Convolutional block attention module. In: Proceedings of the European conference on computer vision (ECCV). pp. 3–19 (2018)

    work page 2018

  29. [29]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wu, X., Wu, Z., Guo, H., Ju, L., Wang, S.: Dannet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15769– 15778 (2021)

    work page 2021

  30. [30]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Xiong, Y., Varadarajan, B., Wu, L., Xiang, X., Xiao, F., Zhu, C., Dai, X., Wang, D., Sun, F., Iandola, F., et al.: Efficientsam: Leveraged masked image pretraining for efficient segment anything. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16111–16121 (2024)

    work page 2024

  31. [31]

    Personalize segment anything model with one shot

    Zhang, R., Jiang, Z., Guo, Z., Yan, S., Pan, J., Ma, X., Dong, H., Gao, P., Li, H.: Personalize segment anything model with one shot. arXiv preprint arXiv:2305.03048 (2023) Abbreviated paper title 17 A Supplementary Materials A.1 Additional Related Work Attention Mechanisms for Visual Robustness.Attention mechanisms have been widely adopted to enhance vis...

    work page arXiv 2023