Pith. sign in

REVIEW 4 major objections 6 minor 75 references

Cast and Attached Shadow Detection via Iterative Light and Geometry Reasoning

T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read This framework claims that cast and attached shadows are detected more accurately by iterating between shadow detection and light-direction estimation, cutting attached-shadow error by at least 33% relative to prior methods.

desk verdict A genuinely new dataset and task framing, with real attached-shadow gains, but the core physics equation is sign-flipped as written and most light labels are heuristic; fix those and this becomes a solid contribution. read the letter →

arxiv 2512.06179 v3 pith:TFUITMQL submitted 2025-12-05 cs.CV

classification cs.CV
keywords castshadowsattachedlightdirectionestimationsurfacenormalsiterativerefinementshadowdetectiondatasetphysics-basedreasoningclosed-looparchitecture
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 claims that cast and attached shadows can be detected separately and more accurately by closing a feedback loop between a shadow detector and a light-direction estimator: the light estimate plus surface normals produce a geometry-derived partial attached-shadow map that guides the next detection pass, and the improved shadows in turn sharpen the light estimate. It further claims this iterative scheme cuts attached-shadow error by at least 33% relative to prior methods on a new 1,458-image benchmark with separate cast and attached annotations, without sacrificing full-shadow or cast-shadow performance. A sympathetic reader would care because attached shadows carry direct 3D-shape and lighting cues that existing cast-focused detectors ignore, and this is the first dedicated dataset-and-model combination for the task.

What carries the argument

The central mechanism is the reciprocal loop between the shadow detection module and the light estimation module, coupled through a physics-derived partial attached shadow map. The map is computed as M_attached = 1{n·ℓ>0}, where ℓ is the estimated unit light direction in a camera-centric frame and n is the surface normal field; it identifies pixels the geometry says should be attached-shadow candidates (though, as written, the inequality selects surfaces facing the light rather than away from it). The map is added as input to the detector's next pass, and the shadow predictions are fed into the light estimator, with losses L_att, L_dir, L_unit aligning the map to ground-truth attached masks,

What would settle it

Render a sphere under a known directional light, run the geometry branch alone with ℓ pointing from light to scene, and check whether 1{n·ℓ>0} selects the dark or the lit hemisphere; the paper's text predicts the dark hemisphere, but the formula as written selects the lit one. Also, on a test set with measured light directions, compare the estimated ℓ to the measured direction—if the 'heuristic' targets are the only supervision, the estimator's true accuracy has not been established.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a dual-module network—one module segmenting {background, cast, attached} and a second estimating a 3D light direction—can be trained in a closed loop, where the estimated light direction and surface normals are combined by n·ℓ>0 into a 'partial attached shadow map' that is fed back into the detector on the next pass. Iterating three times yields the reported results: full BER 6.50, cast BER 5.04, attached BER 12.93 with attached F1 86.49, compared with attached BER above 19 for retrained baselines. The paper attributes the gain to explicit geometry and lighting cues rather than appearance alone.

Load-bearing premise

The load-bearing premise is that a single dominant directional light exists and that the signed-geometry condition stated in Eq. (1)—which, as written, flags surfaces facing the light as attached-shadow pixels—correctly identifies where attached shadows form; if the sign is a typo or the single-light assumption fails, the claimed physical grounding does not hold.

Editorial extensions

If this is right

  • If correct, separate cast and attached masks become a standard output, enabling downstream tasks such as instance-level shadow removal and object-placement reasoning to consume shadow-type information directly.
  • The closed-loop design means performance saturates after about three iterations, so the accuracy gain is available at roughly 3x the non-iterative runtime.
  • The curated 1,458-image benchmark with separate cast/attached annotations and an 'undefined' label for ambiguous shadows provides an evaluation protocol that future methods can be compared on directly.
  • Existing cast-focused detectors, when retrained on the new data, still show attached-shadow BER between 19 and 27, indicating that appearance-only supervision is insufficient and that geometry/light cues are necessary.
  • The method's ability to label 'undefined' shadow regions correctly even when annotations only cover foreground objects suggests a path to weak supervision and pseudo-labeling for scalable data collection.

Reading between the lines

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

  • Editorial inference: If Eq. (1)'s sign convention is a typo and the intended condition is n·ℓ<0, the reported gains remain interpretable: the network receives a spatial prior that correlates with self-occluded versus lit geometry, and the iterative loop can learn to correct any polarity error; the method would then be demonstrating the value of a geometry cue rather than the exact physical law sta
  • Editorial inference: Because the light-direction targets for SOBA and CUHK are heuristic estimates from cast-shadow centroids and relative depth, the light module is trained on a proxy that has no independent ground-truth validation; a fair test of 'light estimation' needs images with measured or rendered light directions.
  • Editorial inference: The single-directional-light assumption limits generalization to indoor and night scenes with multiple sources; a natural extension is to replace the scalar light direction with a low-dimensional lighting field or a small set of directional lights, which the loop architecture could in principle accommodate.
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

4 major / 6 minor

Summary. The paper proposes a dual-module framework for joint detection of cast and attached shadows. A shadow detection module predicts separate masks, while a light estimation module predicts a scene-wide light direction. Using surface normals and the predicted light direction, the system computes a partial attached-shadow map via Eq. (1) and feeds it back to the detector over several iterations, forming a closed loop. To support the task, the authors curate a 1,458-image dataset with separate cast/attached annotations, sourced from WSRD, SOBA, and CUHK. Experiments report a ≥33% relative reduction in attached-shadow BER over prior methods, with maintained full and cast shadow performance. The paper also includes ablations on the normal input, iterations, and loss components.

Significance. If the empirical results hold, this is a useful contribution to a relatively under-studied problem. The dataset fills a real gap, and the reported attached-shadow improvements are substantial. The iterative light-geometry feedback idea is well motivated by the mutual dependence of shadows and illumination. However, the paper's central claim of 'physics-grounded' reasoning is weakened by the fact that light-direction targets for most images are heuristic estimates derived from the very shadow annotations used for supervision, and because the non-differentiable step in Eq. (1) is never reconciled with the gradient-based training of the light estimator. These issues are addressable and do not in themselves refute the empirical gains, but they must be clarified before the physical-grounding claim can be accepted.

major comments (4)
  1. [§3.2.1, Eq. (1) & Eq. (7)] Eq. (1) defines M_attached via a hard threshold on n·ℓ. This is non-differentiable with respect to the predicted light direction ℓ. Since Latt (Eq. 7) supervises M_attached against the attached-shadow ground truth and is included in the light-estimation loss, gradients must flow through the threshold to update ℓ. The paper does not specify a differentiable relaxation (e.g., sigmoid) or a straight-through estimator. Without this, the proposed joint iterative training is not implementable as described, and the improvement from Latt reported in Table 3 lacks an explanation. Please clarify the exact differentiable formulation and gradient flow.
  2. [§4 (Light Directions) & §3.2.2, Eq. (8)] For 1,248 of the 1,458 images (SOBA and CUHK), the light-direction targets are heuristic estimates computed from cast-shadow geometry (object/shadow centroids and relative depth). Additionally, Latt directly aligns the partial map with the attached-shadow ground truth. Thus the 'physical prior' is partly supervised by the same annotations used for detection, making the claimed grounding less independent than implied. This is not fatal, but the paper should qualify the claim and provide a quantitative assessment of light-direction accuracy, e.g., on the 220 WSRD images with calibrated lighting, or via an ablation that removes Ldir/Latt and compares.
  3. [§5.1, Tables 2 & 3] The paper never evaluates a 'physics-only' baseline: a detector that uses Eq. (1) with the provided (oracle) light directions and normal maps to directly produce an attached-shadow map. Such a baseline would isolate the contribution of the learned modules and directly test the value of the physical prior. Without it, the reader cannot tell whether the gains stem from the geometric prior or from the network's capacity. Please add this baseline using both oracle light directions and the estimated ones.
  4. [§4 (Light Directions)] The heuristic light-direction procedure is under-specified. The step 'infer the third component using the relative depth between the object region and its cast shadow region' lacks a concrete formula, making the dataset construction non-reproducible. Please provide the exact computation, including how the relative depth is converted into a signed z-component and how normalization is performed.
minor comments (6)
  1. [§3.1 & §3.2.1] The sign convention in Eq. (1) is a potential source of confusion. I verified that it is consistent with the text: since ℓ points from the light toward the surface, a surface facing away from the light has its outward normal aligned with ℓ, giving n·ℓ>0. A short sentence explicitly stating this would prevent misreadings.
  2. [§4] The paper states it is the 'first dataset specifically curated for cast and attached shadow detection,' but SILT [65] already annotated attached shadows on the SBU-test set. Consider rephrasing to 'first dataset with separate cast and attached masks for both training and evaluation across multiple sources.'
  3. [§3.2.2, Eq. (2)] The Dice loss is applied to s, which is a logit (LSE(z_cast,z_att) - z_bg). Please clarify that the Dice coefficient is computed after a sigmoid activation, otherwise the expression is ambiguous.
  4. [§5.1, Table 2] The † marker groups methods fine-tuned on the proposed dataset, but the table also includes SBU-pretrained models. The grouping would be clearer if the rows were visually separated, e.g., with a horizontal rule or a sub-header.
  5. [§4] Using color-space subtraction between WSRD shadow and shadow-free images to derive full-shadow masks is a heuristic that may introduce noisy labels. This limitation is mentioned implicitly but could be discussed explicitly in Sec. 6.
  6. [§5.1] No runtime or parameter-count comparison against prior methods is reported, apart from the internal iteration cost. A brief comparison would help position the method's practical efficiency.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the light-geometry prior is an auxiliary supervised cue, not a by-construction prediction.

full rationale

The paper's claimed derivation chain does not reduce to its own inputs. The partial attached-shadow map (Eq. 1) is a deterministic function of the estimated light direction and surface normals; it is not fitted to the output. The light-direction targets for SOBA/CUHK (Sec. 4) are heuristic labels derived from cast-shadow centroids and relative depth, but these labels only supervise an auxiliary light-estimation loss (Ldir, Eq. 8). The final attached-shadow predictions come from a learned segmentation module and are evaluated on a manually annotated, held-out test set, so the headline BER reduction is an empirical measurement, not a construction. The Latt loss (Eq. 7) aligns the partial map with attached-shadow ground truth; this is deep supervision of an intermediate representation, not a mechanism that forces the final prediction to equal the label. Ablations (Tab. 3) show each component contributes and performance is not a vacuous identity. Self-citations such as [37] are prior published work used as motivation, not as an unverified premise that forces the result. The apparent sign issue in Eq. (1) is a physics-convention question, not a circularity, and does not constitute a reduction of the derivation to its inputs. Overall, the paper is self-contained against its own benchmark and held-out evaluations, so no circular step is exhibited.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced; the partial attached shadow map is a deterministic feature computed from ℓ and normals, not a new entity with an independent falsifiable handle. The central assumptions are the single-directional-light model, the accuracy of monocular normals, the correctness of Eq. (1), and the heuristic light labels derived from shadows.

free parameters (3)
  • Margin m in Ldist (Eq. 4) = 0.2
    Class-conditional margin chosen by hand; no sensitivity analysis is reported.
  • Loss weights λ_Dice, λ_dist, λ_att, λ_dir, λ_unit = 0.1, 0.2, 0.4, 0.5, 0.1
    Set for training (§5); no ablation varying these weights is presented.
  • Number of feedback iterations = 3
    Ablation shows improvements plateau at 2–3 iterations and 5 iterations hurt class-specific metrics; 3 is chosen as the final setting.
assumptions (5)
  • domain assumption Single scene-wide directional light source shared across each image (§3.1).
    The partial attached map and light estimator assume one dominant light direction; the paper notes in §6 that this fails for multi-source indoor/night settings.
  • domain assumption Surface normals from Depth Anything v2 are accurate enough for attached-shadow reasoning.
    Normals are generated by an off-the-shelf monocular model and converted from relative depth; §6 shows cases where normals are wrong and attached predictions suffer.
  • ad hoc to paper The sign of n·ℓ determines attached shadows (Eq. 1), ignoring visibility and geometric blocking.
    The paper calls this a partial map because occlusion is not modeled, but the sign convention is internally inconsistent with the stated definition of ℓ.
  • domain assumption Heuristic light directions from object/cast-shadow centroids and relative depth approximate true illumination.
    For SOBA/CUHK, ℓ* is estimated rather than measured; no validation of light-direction accuracy is reported.
  • domain assumption Manual foreground-object annotations and 'undefined' masks define the evaluation scope.
    Attached evaluation is limited to object regions and undefined shadows are excluded; the paper states this avoids ambiguous ownership but limits generalization claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cast and Attached Shadow Detection via Iterative Light and Geometry Reasoning." pith.science (2026). https://pith.science/paper/TFUITMQL

@misc{pith2026251206179,
  author       = {Pith},
  title        = {Pith review of: Cast and Attached Shadow Detection via Iterative Light and Geometry Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TFUITMQL}},
  note         = {Machine review of arXiv:2512.06179}
}
read the original abstract

Shadows encode rich information about scene geometry and illumination, yet existing methods either predict a unified shadow mask or overlook attached shadows entirely. We address this gap by proposing a framework for jointly detecting cast and attached shadows through explicit physical modeling of light direction and surface geometry under a dominant directional-light setting. Our approach is grounded in a simple observation: surfaces facing away from the light source tend to fall into shadow. We exploit the reciprocal relationship between shadow formation and light estimation to construct a closed feedback loop, a dual-module architecture in which a shadow detection module and a light estimation module iteratively refine each other. At each pass, updated light estimates, together with surface normals, produce partial attached shadow maps that guide detection, while improved shadow predictions sharpen light estimation. To support training and evaluation, we introduce a dataset of 1,458 images with manually annotated cast and attached shadow masks sourced from three existing benchmarks. Experiments demonstrate that our proposed method outperforms prior methods, with at least a 33% reduction in attached-shadow BER, while maintaining strong full-shadow and cast-shadow performance.

Figures

Figures reproduced from arXiv: 2512.06179 by the authors.

Figure 1
Figure 1. Existing shadow detection methods primarily focus on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview. Our network consists of two modules: a shadow detection module and a light estimation module. The shadow detection module simultaneously predicts both cast and attached shadows, while the light estimation module predicts a three-dimensional light direction and generates a partial attached shadow map from the estimated light direction. This partial map is then fed back into the shadow detection module as an… view at source ↗
Figure 3
Figure 3. Problem Setup. The goal is to separately predict cast and attached shadows. Cast shadows appear on external surfaces, while attached shadows form directly on the object itself. 3.2. Framework Overview As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Examples of our proposed dataset. We curate the dataset from the WSRD [50], SOBA [54], and CUHK [16] datasets. Each image is paired with its corresponding normal map, light direction, cast shadow, attached shadow, and “undefined” shadow masks. The light direction is sp…
Figure 5
Figure 5. Figure 5: Qualitative comparison of the proposed framework for detecting cast and attached shadows against retrained BDRAR[ [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Examples where only the foreground object is annotated, [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Limitations. The first row shows errors in attached shadow predictions due to an inaccurate normal map (compare highlighted regions). The second row illustrates detected shadow masks when multiple objects are present. provements to foundation models could address this …
Figure 7
Figure 7. Figure 7: Effectiveness of the iterative learning scheme. Our pro [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

75 extracted references · 2 linked inside Pith

  1. [1]

    Local-to-global panorama in- painting for locale-aware indoor lighting prediction.IEEE Transactions on Visualization and Computer Graphics, 29 (11):4405–4416, 2023

    Jiayang Bai, Zhen He, Shan Yang, Jie Guo, Zhenyu Chen, Yan Zhang, and Yanwen Guo. Local-to-global panorama in- painting for locale-aware indoor lighting prediction.IEEE Transactions on Visualization and Computer Graphics, 29 (11):4405–4416, 2023. 3

  2. [2]

    Sam-adapter: Adapting segment anything in underperformed scenes

    Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3367–3375, 2023. 2

  3. [3]

    A multi-task mean teacher for semi-supervised shadow detection

    Zhihao Chen, Lei Zhu, Liang Wan, Song Wang, Wei Feng, and Pheng-Ann Heng. A multi-task mean teacher for semi-supervised shadow detection. InProceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 5611–5620, 2020. 1, 2, 5, 7

  4. [4]

    Triple-cooperative video shadow detection

    Zhihao Chen, Liang Wan, Lei Zhu, Jia Shen, Huazhu Fu, Wennan Liu, and Jing Qin. Triple-cooperative video shadow detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2715– 2724, 2021. 1, 2

  5. [5]

    Salesin, and Richard Szeliski

    Yung-Yu Chuang, Dan B Goldman, Brian Curless, David H. Salesin, and Richard Szeliski. Shadow matting and com- positing.ACM Transactions on Graphics, 22(3):494–500,

  6. [6]

    Towards ghost-free shadow removal via dual hierarchical aggrega- tion network and shadow matting gan

    Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggrega- tion network and shadow matting gan. InProceedings of the AAAI Conference on Artificial Intelligence, pages 10680– 10687, 2020. 2

  7. [7]

    Ev- erlight: Indoor-outdoor editable hdr lighting estimation

    Mohammad Reza Karimi Dastjerdi, Jonathan Eisenmann, Yannick Hold-Geoffroy, and Jean-Franc ¸ois Lalonde. Ev- erlight: Indoor-outdoor editable hdr lighting estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7420–7429, 2023. 3

  8. [8]

    Argan: Attentive recurrent generative adversarial net- work for shadow detection and removal

    Bin Ding, Chengjiang Long, Ling Zhang, and Chunxia Xiao. Argan: Attentive recurrent generative adversarial net- work for shadow detection and removal. InProceedings of the IEEE/CVF international conference on computer vision, pages 10213–10222, 2019. 2

Show all 75 references
  1. [9]

    Combining color and geometry for the active, visual recognition of shadows

    Gareth Funka-Lea and Ruzena Bajcsy. Combining color and geometry for the active, visual recognition of shadows. In Proceedings of IEEE International Conference on Computer Vision, pages 203–209. IEEE, 1995. 2

  2. [10]

    Deep parametric indoor lighting estimation

    Marc-Andr ´e Gardner, Yannick Hold-Geoffroy, Kalyan Sunkavalli, Christian Gagn ´e, and Jean-Franc ¸ois Lalonde. Deep parametric indoor lighting estimation. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 7175–7183, 2019. 3

  3. [11]

    Shadowd- iffusion: When degradation prior meets diffusion model for shadow removal

    Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowd- iffusion: When degradation prior meets diffusion model for shadow removal. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 1404...

  4. [12]

    Paired regions for shadow detection and removal.IEEE transactions on pattern analysis and machine intelligence, 35(12):2956–2967, 2012

    Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal.IEEE transactions on pattern analysis and machine intelligence, 35(12):2956–2967, 2012. 2

  5. [13]

    Shadow removal refinement via material- consistent shadow edges.2025 IEEE/CVF Winter Confer- ence on Applications of Computer Vision (WACV), pages 2631–2641, 2024

    Shilin Hu, Hieu Le, ShahRukh Athar, Sagnik Das, and Dim- itris Samaras. Shadow removal refinement via material- consistent shadow edges.2025 IEEE/CVF Winter Confer- ence on Applications of Computer Vision (WACV), pages 2631–2641, 2024. 2

  6. [14]

    Direction-aware spatial context features for shadow detection

    Xiaowei Hu, Lei Zhu, Chi-Wing Fu, Jing Qin, and Pheng- Ann Heng. Direction-aware spatial context features for shadow detection. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 7454–7462,

  7. [15]

    Mask-shadowgan: Learning to remove shadows from unpaired data

    Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. InProceedings of the IEEE/CVF interna- tional conference on computer vision, pages 2472–2481,

  8. [16]

    Revisiting shadow de- tection: A new benchmark dataset for complex world.IEEE Transactions on Image Processing, 30:1925–1934, 2021

    Xiaowei Hu, Tianyu Wang, Chi-Wing Fu, Yitong Jiang, Qiong Wang, and Pheng-Ann Heng. Revisiting shadow de- tection: A new benchmark dataset for complex world.IEEE Transactions on Image Processing, 30:1925–1934, 2021. 1, 2, 4, 5, 6, 7

  9. [17]

    Des3: Adaptive attention-driven self and soft shadow removal using vit similarity

    Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. InProceedings of the AAAI Conference on Artificial Intelligence, pages 2634–2642,

  10. [18]

    Pakorn KaewTraKulPong and Richard Bowden. An im- proved adaptive background mixture model for real-time tracking with shadow detection.Video-based surveillance systems: Computer vision and distributed processing, pages 135–144, 2002. 1

  11. [19]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InICLR, 2015. 5

  12. [20]

    Detecting ground shadows in outdoor con- sumer photographs

    Jean-Franc ¸ois Lalonde, Alexei A Efros, and Srinivasa G Narasimhan. Detecting ground shadows in outdoor con- sumer photographs. InComputer Vision–ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part II 11, p...

  13. [21]

    Shadow removal via shadow image decomposition

    Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. InProceedings of the IEEE/CVF In- ternational Conference on Computer Vision, pages 8578– 8587, 2019. 2

  14. [22]

    From shadow segmentation to shadow removal

    Hieu Le and Dimitris Samaras. From shadow segmentation to shadow removal. InECCV, 2020

  15. [23]

    Physics-based shadow im- age decomposition for shadow removal.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9088– 9101, 2021

    Hieu Le and Dimitris Samaras. Physics-based shadow im- age decomposition for shadow removal.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9088– 9101, 2021. 2

  16. [24]

    Hieu Le, Vu Nguyen, Chen-Ping Yu, and D. Samaras. Geodesic distance histogram feature for video segmentation. ACCV, 2016. 2

  17. [25]

    Co-localization with category-consistent features 9 and geodesic distance propagation

    Hieu Le, Chen-Ping Yu, Gregory Zelinsky, and Dimitris Samaras. Co-localization with category-consistent features 9 and geodesic distance propagation. InICCV 2017 Workshop on CEFRL: Compact and Efficient Feature Representation and Learning in Computer Vision, 2017. 2

  18. [26]

    A+ d net: Training a shadow de- tector with adversarial shadow attenuation

    Hieu Le, Tomas F Yago Vicente, Vu Nguyen, Minh Hoai, and Dimitris Samaras. A+ d net: Training a shadow de- tector with adversarial shadow attenuation. InProceedings of the European Conference on Computer Vision (ECCV), pages 662–678, 2018. 2

  19. [27]

    Weakly labeling the antarctic: The penguin colony case

    Hieu Le, Bento Goncalves, Dimitris Samaras, and Heather Lynch. Weakly labeling the antarctic: The penguin colony case. InCVPR Workshops, 2019. 2

  20. [28]

    Hieu Le, Dimitris Samaras, and Heather J. Lynch. A con- volutional neural network architecture designed for the auto- mated survey of seabird colonies.Remote Sensing in Ecology and Conservation, 8(2):251–262, 2022. 2

  21. [29]

    Leveraging inpainting for single-image shadow removal

    Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 13055–13064, 2023. 2

  22. [30]

    A convnet for the 2020s.Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2022

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s.Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), 2022. 5

  23. [31]

    Brightness correction and shadow removal for video change detection with uavs

    Thomas M ¨uller and Bastian Erdn ¨ueß. Brightness correction and shadow removal for video change detection with uavs. InAutonomous Systems: Sensors, Processing, and Security for Vehicles and Infrastructure 2019, page 1100906. SPIE,

  24. [32]

    Attached shadow coding: Estimating surface normals from shadows under unknown reflectance and lighting conditions

    Takahiro Okabe, Imari Sato, and Yoichi Sato. Attached shadow coding: Estimating surface normals from shadows under unknown reflectance and lighting conditions. In2009 IEEE 12th International Conference on Computer Vision, pages 1693–1700. IEEE, 2009. 2

  25. [33]

    Robust shadow and illumination estimation using a mixture model

    Alexandros Panagopoulos, Dimitris Samaras, and Nikos Paragios. Robust shadow and illumination estimation using a mixture model. In2009 IEEE Conference on Computer Vi- sion and Pattern Recognition, pages 651–658. IEEE, 2009. 1

  26. [34]

    Illumination estimation from shadow borders

    Alexandros Panagopoulos, Tom ´as F Yago Vicente, and Dim- itris Samaras. Illumination estimation from shadow borders. In2011 IEEE International Conference on Computer Vision Workshops (ICCV Workshops), pages 798–805. IEEE, 2011

  27. [35]

    Illumination estimation and cast shadow detection through a higher-order graphical model

    Alexandros Panagopoulos, Chaohui Wang, Dimitris Sama- ras, and Nikos Paragios. Illumination estimation and cast shadow detection through a higher-order graphical model. In CVPR 2011, pages 673–680. IEEE, 2011

  28. [36]

    Estimating shadows with the bright channel cue

    Alexandros Panagopoulos, Chaohui Wang, Dimitris Sama- ras, and Nikos Paragios. Estimating shadows with the bright channel cue. InTrends and Topics in Computer Vision: ECCV 2010 Workshops, Heraklion, Crete, Greece, Septem- ber 10-11, 2010, Revised Selected Papers, Part II 11, p...

  29. [37]

    Simultaneous cast shadows, illu- mination and geometry inference using hypergraphs.IEEE transactions on pattern analysis and machine intelligence, 35(2):437–449, 2012

    Alexandros Panagopoulos, Chaohui Wang, Dimitris Sama- ras, and Nikos Paragios. Simultaneous cast shadows, illu- mination and geometry inference using hypergraphs.IEEE transactions on pattern analysis and machine intelligence, 35(2):437–449, 2012. 1, 2, 3

  30. [38]

    Pytorch: An im- perative style, high-performance deep learning library.Ad- vances in neural information processing systems, 32, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An im- perative style, high-performance deep learning library.Ad- vances in neural information processing systems, ...

  31. [39]

    Diffusionlight: Light probes for free by painting a chrome ball

    Pakkapon Phongthawee, Worameth Chinchuthakun, Non- taphat Sinsunthithet, Varun Jampani, Amit Raj, Pramook Khungurn, and Supasorn Suwajanakorn. Diffusionlight: Light probes for free by painting a chrome ball. InProceed- ings of the IEEE/CVF conference on computer vision and pat...

  32. [40]

    Deshadownet: A multi-context embedding deep network for shadow removal

    Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 4067–4075, 2017. 2

  33. [41]

    Dimitrios Samaras and Dimitris Metaxas. Incorporating il- lumination constraints in deformable models for shape from shading and light direction estimation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(2):247–264,

  34. [42]

    Illumination from shadows.IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(3):290–300, 2003

    Imari Sato, Yoichi Sato, and Katsushi Ikeuchi. Illumination from shadows.IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(3):290–300, 2003. 2

  35. [43]

    Hdr environment map estimation for real-time augmented reality

    Gowri Somanath and Daniel Kurz. Hdr environment map estimation for real-time augmented reality. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11298–11306, 2021. 3

  36. [44]

    Detec- tion of moving cast shadows for object segmentation.IEEE Transactions on multimedia, 1(1):65–76, 1999

    Jurgen Stander, Roland Mech, and J ¨orn Ostermann. Detec- tion of moving cast shadows for object segmentation.IEEE Transactions on multimedia, 1(1):65–76, 1999. 2

  37. [45]

    Nan Su, Ye Zhang, Shu Tian, Yiming Yan, and Xinyuan Miao. Shadow detection and removal for occluded ob- ject information recovery in urban high-resolution panchro- matic satellite images.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 9:2568– ...

  38. [46]

    Generalised dice overlap as a deep learning loss function for highly unbalanced seg- mentations

    Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced seg- mentations. InInternational Workshop on Deep Learning in Medical Image Analysis, pages 240–248. Springer, 2017. 4

  39. [47]

    Adaptive illumination mapping for shadow detection in raw images

    Jiayu Sun, Ke Xu, Youwei Pang, Lihe Zhang, Huchuan Lu, Gerhard Hancke, and Rynson Lau. Adaptive illumination mapping for shadow detection in raw images. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 12709–12718, 2023. 2

  40. [48]

    Visi- bility subspaces: Uncalibrated photometric stereo with shad- ows

    Kalyan Sunkavalli, Todd Zickler, and Hanspeter Pfister. Visi- bility subspaces: Uncalibrated photometric stereo with shad- ows. InComputer Vision–ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part II 11, pages 251–

  41. [49]

    Kulkarni

    Shailaja Surkutlawar and R. Kulkarni. Shadow suppression using rgb and hsv color space in moving object detection.In- ternational Journal of Advanced Computer Science and Ap- plications, 4:164–169, 2013. 1 10

  42. [50]

    Ntire 2024 im- age shadow removal challenge report

    Florin-Alexandru Vasluianu, Tim Seizinger, Zhuyun Zhou, Zongwei Wu, Cailian Chen, Radu Timofte, Wei Dong, Han Zhou, Yuqiong Tian, Jun Chen, et al. Ntire 2024 im- age shadow removal challenge report. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogn...

  43. [51]

    Eclipse: Disam- biguating illumination and materials using unintended shad- ows

    Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Bar- ron, Todd Zickler, and Pratul P Srinivasan. Eclipse: Disam- biguating illumination and materials using unintended shad- ows. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 7...

  44. [52]

    Large-scale training of shadow de- tectors with noisily-annotated shadow examples

    Tom ´as F Yago Vicente, Le Hou, Chen-Ping Yu, Minh Hoai, and Dimitris Samaras. Large-scale training of shadow de- tectors with noisily-annotated shadow examples. InCom- puter Vision–ECCV 2016: 14th European Conference, Am- sterdam, The Netherlands, October 11-14, 2016, Proceed...

  45. [53]

    Stacked condi- tional generative adversarial networks for jointly learning shadow detection and shadow removal

    Jifeng Wang, Xiang Li, and Jian Yang. Stacked condi- tional generative adversarial networks for jointly learning shadow detection and shadow removal. InProceedings of the IEEE conference on computer vision and pattern recog- nition, pages 1788–1797, 2018. 2

  46. [54]

    Instance shadow detection

    Tianyu Wang, Xiaowei Hu, Qiong Wang, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1880–1889, 2020. 1, 2, 4, 5

  47. [55]

    Single-stage instance shadow detection with bidirec- tional relation learning

    Tianyu Wang, Xiaowei Hu, Chi-Wing Fu, and Pheng-Ann Heng. Single-stage instance shadow detection with bidirec- tional relation learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1–11, 2021

  48. [56]

    Instance shadow detection with a single-stage detector

    Tianyu Wang, Xiaowei Hu, Pheng-Ann Heng, and Chi-Wing Fu. Instance shadow detection with a single-stage detector. IEEE Transactions on Pattern Analysis and Machine Intelli- gence, 45(3):3259–3273, 2022. 2

  49. [57]

    People as scene probes

    Yifan Wang, Brian L Curless, and Steven M Seitz. People as scene probes. InComputer Vision–ECCV 2020: 16th Eu- ropean Conference, Glasgow, UK, August 23–28, 2020, Pro- ceedings, Part X 16, pages 438–454. Springer, 2020. 1

  50. [58]

    Swinshadow: Shifted window for ambiguous adjacent shadow detection.ACM Transactions on Multime- dia Computing, Communications and Applications, 2024

    Yonghui Wang, Shaokai Liu, Li Li, Wengang Zhou, and Houqiang Li. Swinshadow: Shifted window for ambiguous adjacent shadow detection.ACM Transactions on Multime- dia Computing, Communications and Applications, 2024. 2

  51. [59]

    Importance-based token merging for efficient image and video generation

    Haoyu Wu, Jingyi Xu, Hieu Le, and Dimitris Samaras. Importance-based token merging for efficient image and video generation. InICCV, 2025. 2

  52. [60]

    Generating representative samples for few-shot classification

    Jingyi Xu and Hieu Le. Generating representative samples for few-shot classification. InCVPR, 2022

  53. [61]

    Variational feature disentangling for fine- grained few-shot classification

    Jingyi Xu, Hieu Le, Mingzhen Huang, ShahRukh Athar, and Dimitris Samaras. Variational feature disentangling for fine- grained few-shot classification. InICCV, 2021

  54. [62]

    Zero-shot object counting.CVPR, pages 15548– 15557, 2023

    Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting.CVPR, pages 15548– 15557, 2023

  55. [63]

    Generating fea- tures with increased crop-related diversity for few-shot ob- ject detection

    Jingyi Xu, Hieu Le, and Dimitris Samaras. Generating fea- tures with increased crop-related diversity for few-shot ob- ject detection. InCVPR, 2023

  56. [64]

    Assessing sample quality via the latent space of generative models

    Jingyi Xu, Hieu Le, and Dimitris Samaras. Assessing sample quality via the latent space of generative models. InECCV,

  57. [65]

    Silt: Shadow-aware iterative label tuning for learning to detect shadows from noisy labels

    Han Yang, Tianyu Wang, Xiaowei Hu, and Chi-Wing Fu. Silt: Shadow-aware iterative label tuning for learning to detect shadows from noisy labels. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 12687–12698, 2023. 1, 2, 5, 6, 7

  58. [66]

    Depth any- thing v2.arXiv preprint arXiv:2406.09414, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2.arXiv preprint arXiv:2406.09414, 2024. 2, 5, 8

  59. [67]

    All-weather deep outdoor lighting estimation

    Jinsong Zhang, Kalyan Sunkavalli, Yannick Hold-Geoffroy, Sunil Hadap, Jonathan Eisenman, and Jean-Franc ¸ois Lalonde. All-weather deep outdoor lighting estimation. In Proceedings of the IEEE/CVF conference on Computer Vi- sion and Pattern Recognition, pages 10158–10166, 2019. 3

  60. [68]

    Improving shadow suppression for illumination robust face recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 41:611–624, 2019

    Wuming Zhang, Xi Zhao, Jean-Marie Morvan, and Liming Chen. Improving shadow suppression for illumination robust face recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 41:611–624, 2019. 1

  61. [69]

    Bilateral refer- ence for high-resolution dichotomous image segmentation

    Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral refer- ence for high-resolution dichotomous image segmentation. arXiv preprint arXiv:2401.03407, 2024. 5

  62. [70]

    Distraction-aware shadow detection

    Quanlong Zheng, Xiaotian Qiao, Ying Cao, and Rynson WH Lau. Distraction-aware shadow detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5167–5176, 2019. 2, 5, 7

  63. [71]

    Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection

    Lei Zhu, Zijun Deng, Xiaowei Hu, Chi-Wing Fu, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. InProceedings of the European Con- ference on Computer Vision (ECCV), pages 121–136, 20...

  64. [72]

    Miti- gating intensity bias in shadow detection via feature decom- position and reweighting

    Lei Zhu, Ke Xu, Zhanghan Ke, and Rynson WH Lau. Miti- gating intensity bias in shadow detection via feature decom- position and reweighting. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4702– 4711, 2021. 2, 5, 6, 7

  65. [73]

    Single image shadow detection via com- plementary mechanism

    Yurui Zhu, Xueyang Fu, Chengzhi Cao, Xi Wang, Qibin Sun, and Zheng-Jun Zha. Single image shadow detection via com- plementary mechanism. InProceedings of the 30th ACM International Conference on Multimedia, pages 6717–6726,

  66. [74]

    Object-based cloud and cloud shadow detection in landsat imagery.Remote sensing of environment, 118:83–94, 2012

    Zhe Zhu and Curtis E Woodcock. Object-based cloud and cloud shadow detection in landsat imagery.Remote sensing of environment, 118:83–94, 2012. 1 11

  67. [2003]

    Sepcial Issue of the SIGGRAPH 2003 Proceedings. 1

Pith tools

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