Uncertainty-Guided Face Matting for Occlusion-Aware Face Transformation

Hyebin Cho; Jaehyup Lee

REVIEW 4 major objections 5 minor 46 references

Uncertainty-Guided Face Matting for Occlusion-Aware Face Transformation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A trimap-free matting model with uncertainty-guided training separates occluding objects from faces and beats prior matting baselines on occlusion-heavy benchmarks.

desk verdict A real method and a useful new dataset, but the headline RealOcc result does not hold up under scrutiny of label definitions and evaluation protocol. read the letter →

arxiv 2508.03055 v2 pith:H2JQGQPQ submitted 2025-08-05 cs.CV cs.AI

Hyebin Cho , Jaehyup Lee This is my paper · ORCID

classification cs.CVcs.AI

keywords facemattingimagevideotrimap-freeuncertaintyestimationknowledgedistillationocclusionalphamatte

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

Face matting, as defined here, is the task of predicting a per-pixel opacity (alpha) map that separates a face from objects in front of it, such as hands, hair, microphones, and transparent occluders. The paper argues that treating skin as foreground and occlusions as background makes face filters compositing-friendly, and that existing trimap- or mask-dependent matting methods cannot be used in real-time video where such auxiliary inputs are unavailable. FaceMat's answer is a two-stage pipeline: a teacher network jointly learns alpha mattes and per-pixel uncertainty through a negative log-likelihood objective, and that uncertainty then weights a distillation loss so a student network concentrates on ambiguous boundary regions while running with no auxiliary inputs at test time. The authors also construct CelebAMat, a synthetic training and benchmark dataset, and report that FaceMat improves over state-of-the-art matting models, reaching IoU 0.7121 on RealOcc versus 0.4099 for the RVM baseline.

What carries the argument

The mechanism that carries the argument is uncertainty-guided knowledge distillation (UGKD). The teacher is trained so that its predicted per-pixel variance is high exactly where the $\alpha$ matte is ambiguous, via the negative log-likelihood losses in equations (2) and (3); the student is then trained with an uncertainty-weighted L1 loss whose spatial weight map is $w_{\text{unc}} = w_1 + w_2\,\sigma_u^{\text{teacher}}$, with $w_1 = w_2 = 2$. This map performs the role that the unknown region of a trimap plays in conventional matting, but it is predicted by a network, so the student needs no auxiliary input at test time. Alongside it, the CelebAMat pipeline supplies synthetic supervision by compositing occlusion cutouts onto clean faces from CelebAMask-HQ-WO with affine motion augmentation to simulate video.

What would settle it

Take FaceMat and RVM to a real occlusion video corpus with per-pixel alpha annotations and a shared foreground definition for both models; if FaceMat's IoU advantage over RVM shrinks to noise level, the large RealOcc gap is an artifact of synthetic training data or mismatched occlusion labels rather than genuine matting superiority.

Watch

Extended reading notes

Core claim

The paper's central claim is that per-pixel uncertainty estimated during matting training can act as a learned stand-in for a trimap's unknown region, and that using it to spatially weight knowledge distillation yields alpha mattes that preserve facial boundaries while excluding occluders. This is established in the paper by training a teacher with trimap supervision plus NLL losses over both the alpha matte and the uncertainty map, then distilling to a student with an uncertainty-weighted L1 loss. The student, which sees no trimap or mask at inference, is reported to outperform existing matting models on the synthetic CelebAMat benchmark and on the real-world RealOcc dataset, with the largest gains on occlusion-heavy frames.

Load-bearing premise

The load-bearing premise is that occlusions synthesized by pasting cutouts from matting, hand, and texture datasets onto clean faces, with only affine motion augmentation, faithfully represent real-world facial occlusions so that a model trained on them transfers to unconstrained video.

Editorial extensions

If this is right

Face filters can be composited with the alpha matte so the filter stays on visible skin and off occluders such as hands and microphones.
Because no trimap or mask is required at test time, the matte can serve real-time video applications directly.
CelebAMat gives the community a large-scale benchmark with diverse occlusion types for comparing face matting models.
If the RealOcc results hold, the method transfers from synthetic training to unseen real occlusions, which matters for face recognition and face swapping under occlusion as well as for filtering.

Reading between the lines

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

A direct extension the paper leaves implicit is that the same teacher-uncertainty-to-student-weighting recipe could be applied to other structured foregrounds such as animal fur or transparent objects, wherever a trimap-trained teacher is available.
The paper's RealOcc comparison uses segmentation-style binary labels and a foreground definition that differs from RVM's, so part of the reported IoU gain may reflect label disagreement rather than matte quality; a fair comparison would re-annotate with the paper's occlusion definitions.
The uncertainty map itself is a testable artifact: crop test regions by predicted uncertainty and check whether the student's boundary accuracy improves most in the high-uncertainty crops, isolating the contribution of the $w_2$ term.
Temporal inpainting of occluded facial regions is flagged as unsolved, suggesting that the next bottleneck for occlusion-aware face transformation is not alpha estimation but consistent video inpainting.

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.

Desk Editor's Note

read the letter

Nothing here is a scientific breakthrough, but the paper is a genuine engineering contribution. It defines a sensible new task—face matting with occlusions as background—and proposes a two-stage training idea that is plausible: train a teacher with NLL uncertainty, then distill to a student with spatially adaptive weights. The synthetic dataset CelebAMat is released with code, and the ablations (dataset composition, occlusion ratio, schedule) are more thorough than most papers in this space. That part deserves credit.

The problems are in the evaluation. Your stress-test note is correct, and it lands. The RealOcc comparison in Table 3 is the paper's headline real-world evidence, but it is not verifiable as reported: the paper never says how the continuous alpha maps were thresholded to compute IoU/accuracy/recall, and Table 8 shows the occlusion definitions genuinely mismatch (the paper counts beard and eyeglass frames as occlusions; RealOcc does not count beards and counts colorless lenses as occlusions). So the 0.71 vs 0.41 margin is confounded by label-alignment, not necessarily matting quality. On the author-constructed CelebAMat, the gains are small and mixed: UGKD improves MSE/SAD over the no-trimap Stage 1 baseline but actually has lower IoU and accuracy (0.8408 vs 0.8579). That undercuts the 'balance' claim. And the occlusion ratio and uncertainty weights were chosen by evaluating on the test benchmarks, which leaks test information into model selection.

None of this is fatal to the core idea. The method is coherent, the dataset is useful, and the weaknesses are standard evaluation discipline problems that a serious revision can fix. The authors should specify the thresholding rule for RealOcc, ideally using the RealOcc evaluation protocol, add a quantitative FaceOcc comparison, and select hyperparameters on validation splits.

Bottom line: this is a paper I'd send to referees, not desk reject. If I were reviewing, I'd ask for major revision and re-evaluation of the real-world claims, but the contribution is real. Would I use it in my own work? Probably not directly, but the dataset and code make it a reasonable reference for anyone working on matting or occlusion handling.

Referee Report

4 major / 5 minor

Summary. The paper introduces FaceMat, a trimap-free face matting framework for occlusion-aware face transformation. It defines face matting as estimating alpha mattes that separate facial skin (foreground) from occlusions such as hands, hair, and accessories (background). The method uses a two-stage training scheme: a teacher network is trained with a negative log-likelihood loss to predict alpha mattes and per-pixel uncertainty, and a student network is trained via uncertainty-weighted knowledge distillation using the teacher's uncertainty map as a spatial weight. The authors also construct a synthetic dataset, CelebAMat, by compositing occlusions from SIMD, AM2k, HIU, and DTD onto clean face images from CelebAMask-HQ-WO with motion augmentation. They report improvements over RVM and other matting methods on CelebAMat, and a large margin over RVM on the RealOcc dataset.

Significance. If the results were reproducible, the paper would make a useful contribution: the face matting task formulation is practical, the uncertainty-guided distillation idea is reasonable, and the release of code and dataset supports community reuse. However, the central evidence is weakened by the RealOcc evaluation protocol (no thresholding rule, label disagreement) and by an internal inconsistency in Table 2 that contradicts the paper's interpretation. The method's gains on the author-constructed CelebAMat are small and mixed, and the large RealOcc margin is not verifiable as reported. With a tightened evaluation protocol, the paper could still be a valid contribution.

major comments (4)

[Sec. 5.2, Table 3] The RealOcc evaluation is underspecified and confounded. The paper reports IoU/accuracy/recall for continuous alpha predictions on a binary segmentation benchmark, but it never states the binarization threshold (global, per-image, or per-method). Given the very large reported margin (IoU 0.7121 vs 0.4099), this omission makes the result non-reproducible and potentially an artifact. Additionally, Table 8 of the supplementary shows that the paper's occlusion definition disagrees with RealOcc's on beard, colorless transparent lenses, and eyeglass frames; the model is therefore scored against labels it was not trained to predict. The real-world generalization claim should be supported with a clear thresholding protocol and, ideally, a label-aligned evaluation.
[Table 2, Sec. 5.3] The claimed balance of UGKD is not supported by the numbers. Stage1 Extended without trimap achieves IoU 0.8579 ± 0.0036 and accuracy 0.9602 ± 0.0017, while Stage2 (UGKD) achieves IoU 0.8408 ± 0.0181 and accuracy 0.9529 ± 0.0067. Thus UGKD does not improve global semantic consistency over the no-trimap baseline; it improves MSE/SAD (0.0182/8.43 vs 0.0236/16.38) at the expense of IoU/accuracy. The text describes the opposite trade-off and should be corrected or the method should be reframed accordingly.
[Table 2, Stage 1] The NLL-based multi-task learning is not consistently beneficial. Adding NLL to the RVM baseline reduces IoU from 0.8249 to 0.6017 and accuracy from 0.9528 to 0.8180, while connectivity error increases from 7.0865 to 13.3785. The paper states that 'dual supervision leads to performance improvements,' but the improvements are limited to MSE/SAD. Given the large standard deviations (e.g., IoU 0.6017 ± 0.0998), the authors should explain the mechanism or provide a more careful ablation.
[Sec. 5.2, Tables 4–6] Several hyperparameters (fixed occlusion ratio 0.25, w1=2, w2=2, beta, EMA decay) are tuned on the same CelebAMat test set used for the final comparison, with no validation split or cross-validation reported. This creates a risk of overfitting to the benchmark and inflating the reported gains. Please clarify how hyperparameters were selected and whether the reported numbers are on a truly held-out test set.

minor comments (5)

[Sec. 4.2] 'during training of inference' should be 'during training or inference'; the same typo appears as 'durinig' in other places.
[Table 1] The header contains 'Quatitative' (spelling), and the caption says 'under various occlusion types' but the table has no occlusion-type column.
[Eq. (5), Sec. 4.2] The values w1=2 and w2=2 are presented without justification; a sensitivity analysis or explanation for these constants should be included.
[Supplementary Table 10] The table formatting appears misaligned (the EMA and metric columns do not line up with the row labels), which makes it hard to interpret.
[Supplementary Fig. 12 caption] The caption contains 'our our method', a typo.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the teacher–student losses target ground-truth alpha mattes, and no reported result reduces to its own inputs by construction.

full rationale

The paper's derivation chain is not circular. The teacher is trained with a negative log-likelihood loss to jointly predict alpha and per-pixel uncertainty, and the student is trained with an uncertainty-weighted L1 loss L_soft_l1 = ||w_unc ⊙ (α − α_gt)||_1, where α_gt is the ground-truth alpha and w_unc is derived from the teacher's predicted variance. The student's objective is regression to ground truth, not to the teacher's alpha, so the uncertainty map reweights supervision but does not define the prediction. None of the reported gains (Table 2 on CelebAMat, Table 3 on RealOcc) are forced by construction: they are empirical comparisons against an external baseline (RVM) and an external dataset (RealOcc). The CelebAMat benchmark is self-constructed, but evaluation uses held-out test splits and the same composition pipeline is applied to all compared models, which is a standard benchmark-design choice rather than a circular reduction. The RealOcc comparison has protocol concerns — continuous alpha outputs must be thresholded for IoU/accuracy/recall and the paper's occlusion definition (Table 8) differs from RealOcc's labels — but these are validity and evaluation-fairness issues, not circularity: the numbers do not reduce to the model's own fitted parameters or to a self-citation. Hyperparameters such as w1=2, w2=2, the fixed 0.25 occlusion ratio, and EMA decay were selected via ablations reported on the benchmarks, which weakens the strength of the empirical claims but does not make the method's output equivalent to its input. There are no load-bearing self-citations: the cited prior work (e.g., RVM, CelebAMask-HQ, RealOcc) is external, and the paper does not invoke any uniqueness theorem or author-derived constraint to force its design. Overall, the central claim — that uncertainty-guided distillation improves occlusion-aware face matting — is supported by an independent training objective and external evaluation, so the appropriate circularity score is 0.

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

No new physical entities (forces, particles, dimensions) are introduced. The 'face matting task' is a definitional reframing of existing matting and segmentation notions, and the per-pixel uncertainty map is an internal supervisory signal without a falsifiable handle outside the paper's own training pipeline. The free parameters listed above are the main empirical knobs the central claim depends on.

free parameters (5)

beta in NLL losses
Exponent in the NLL losses (Eq. 2, 3) controlling the shape of the uncertainty loss; its value is not reported in the main text or supplementary.
w1 and w2 in uncertainty weight map = 2, 2
Weights in the linear weighting function w_unc = w1 + w2 * sigma_teacher_u (Sec. 4.2). These are hand-set constants, chosen without a described validation procedure.
EMA decay = 0.97
Exponential moving average decay for the teacher model updates (Supplementary Table 10).
Occlusion ratio in CelebAMat training = 0.25
The fixed fraction of occlusion area in training compositions was selected as the best value on the test set (Table 5), which introduces selection bias.
Number of HIU and DTD samples = 200 each
Randomly selected subsamples from the larger HIU and DTD datasets to balance training (Sec. 5.2); arbitrary and potentially influential.

assumptions (4)

domain assumption Synthetic compositions of occlusion cutouts onto clean faces adequately represent the distribution of real occlusions (Sec. 3.2, Fig. 5).
The entire training set and benchmark are synthetic; the claim that models trained on it generalize to real video depends on this assumption.
ad hoc to paper Skin is always the foreground and occlusions such as hair, ears, hands, and beards are always background, regardless of image context (Sec. 3.1, Table 8).
This definitional choice drives the labels and the compositing formulation, but the paper acknowledges it contradicts physical depth ordering ('the physical placement in the scene contradicts this assumption').
domain assumption Gaussian blur applied to binary occlusion masks produces plausible soft alpha boundaries for training (Fig. 5).
The ground truth alpha mattes are derived by blending binary masks with Gaussian blur at boundaries; the fidelity of this approximation to real matting is unverified.
standard math The standard matting compositing equation I_i = alpha_i F_i + (1 - alpha_i) B_i holds for the synthetic composites (Eq. 1).
The paper relies on the conventional alpha compositing model, which is standard for matting.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Uncertainty-Guided Face Matting for Occlusion-Aware Face Transformation." pith.science (2026). https://pith.science/paper/H2JQGQPQ

@misc{pith2026250803055,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-Guided Face Matting for Occlusion-Aware Face Transformation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H2JQGQPQ}},
  note         = {Machine review of arXiv:2508.03055}
}

read the original abstract

Face filters have become a key element of short-form video content, enabling a wide array of visual effects such as stylization and face swapping. However, their performance often degrades in the presence of occlusions, where objects like hands, hair, or accessories obscure the face. To address this limitation, we introduce the novel task of face matting, which estimates fine-grained alpha mattes to separate occluding elements from facial regions. We further present FaceMat, a trimap-free, uncertainty-aware framework that predicts high-quality alpha mattes under complex occlusions. Our approach leverages a two-stage training pipeline: a teacher model is trained to jointly estimate alpha mattes and per-pixel uncertainty using a negative log-likelihood (NLL) loss, and this uncertainty is then used to guide the student model through spatially adaptive knowledge distillation. This formulation enables the student to focus on ambiguous or occluded regions, improving generalization and preserving semantic consistency. Unlike previous approaches that rely on trimaps or segmentation masks, our framework requires no auxiliary inputs making it well-suited for real-time applications. In addition, we reformulate the matting objective by explicitly treating skin as foreground and occlusions as background, enabling clearer compositing strategies. To support this task, we newly constructed CelebAMat, a large-scale synthetic dataset specifically designed for occlusion-aware face matting. Extensive experiments show that FaceMat outperforms state-of-the-art methods across multiple benchmarks, enhancing the visual quality and robustness of face filters in real-world, unconstrained video scenarios. The source code and CelebAMat dataset are available at https://github.com/hyebin-c/FaceMat.git

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 42 canonical work pages

[1]

Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. 2017. SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 39, 12 (2017), 2481–2495. doi:10.1109/TPAMI.2016.2644615

work page arXiv 2017
[2]

Xavier P Burgos-Artizzu, Pietro Perona, and Piotr Dollár. 2013. Robust face landmark estimation under occlusion. In Proceedings of the IEEE international conference on computer vision . 1513–1520

work page 2013
[3]

Huanqia Cai, Fanglei Xue, Lele Xu, and Lili Guo. 2022. Transmatting: Enhanc- ing transparent objects matting with transformers. In European Conference on Computer Vision. Springer, 253–269

work page 2022
[4]

Guowei Chen, Yi Liu, Jian Wang, Juncai Peng, Yuying Hao, Lutao Chu, Shiyu Tang, Zewu Wu, Zeyu Chen, Zhiliang Yu, et al. 2022. PP-matting: high-accuracy natural image matting. arXiv preprint arXiv:2204.09433 (2022)

work page Pith review arXiv 2022
[5]

Steve Chiu, Zhong Jin, and Yingjie Gu. 2015. Active learning combining uncer- tainty and diversity for multi-class image classification. IET Computer Vision 9 (06 2015), 400–407. doi:10.1049/iet-cvi.2014.0140

work page arXiv 2015
[6]

Jiwoong Choi, Dayoung Chun, Hyun Kim, and Hyuk-Jae Lee. 2019. Gaussian YOLOv3: An Accurate and Fast Object Detector Using Localization Uncertainty for Autonomous Driving. 2019 IEEE/CVF International Conference on Computer Vi- sion (ICCV) (2019), 502–511. https://api.semanticscholar.org/CorpusID:104292012

work page 2019
[7]

Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and An- drea Vedaldi. 2014. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition . 3606–3613

2014
[8]

Kaiwen Cui, Rongliang Wu, Fangneng Zhan, and Shijian Lu. 2023. Face Trans- former: Towards High Fidelity and Accurate Face Swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. 668–677

work page 2023

Show all 46 references

[9]

Armen Der Kiureghian and Ove Ditlevsen. 2009. Aleatory or Epistemic? Does It Matter? Structural Safety 31 (03 2009), 105–112. doi:10.1016/j.strusafe.2008.06.020

2009 doi
[10]

Paul Goldberg, Christopher Williams, and Christopher Bishop. 1997. Regression with Input-dependent Noise: A Gaussian Process Treatment. In Advances in Neural Information Processing Systems , M. Jordan, M. Kearns, and S. Solla (Eds.), Vol. 10. MIT Press. https://proceedings.neu...

1997
[11]

Xiaobin Hu, Xu Peng, Donghao Luo, Xiaozhong Ji, Jinlong Peng, Zhengkai Jiang, Jiangning Zhang, Taisong Jin, Chengjie Wang, and Rongrong Ji. 2024. DiffuMat- ting: Synthesizing Arbitrary Objects with Matting-Level Annotation. InEuropean Conference on Computer Vision . Springer, 396–413

2024
[12]

Yihan Hu, Yiheng Lin, Wei Wang, Yao Zhao, Yunchao Wei, and Humphrey Shi
[13]

Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson WH Lau. 2022. Modnet: Real-time trimap-free portrait matting via objective decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 36. 1140–1147

2022
[14]

Alex Kendall and Yarin Gal. 2017. What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?. In Advances in Neural In- formation Processing Systems , I. Guyon, U. Von Luxburg, S. Bengio, H. Wal- lach, R. Fergus, S. Vishwanathan, and R. Garnett (Eds.), Vol. 3...

2017
[15]

Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. 2020. Maskgan: Towards diverse and interactive facial image manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 5549–5558

2020
[16]

Jizhizi Li, Jing Zhang, Stephen J Maybank, and Dacheng Tao. 2022. Bridging composite and real: towards end-to-end deep image matting.International Journal of Computer Vision 130, 2 (2022), 246–266

2022
[17]

Jizhizi Li, Jing Zhang, and Dacheng Tao. 2021. Deep automatic natural image matting. arXiv preprint arXiv:2107.07235 (2021)

2021 arXiv
[18]

Yaoyi Li and Hongtao Lu. 2020. Natural image matting via guided contextual attention. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 34. 11450–11457

2020
[19]

Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. 2021. Real-time high-resolution back- ground matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8762–8771

2021
[20]

Shanchuan Lin, Linjie Yang, Imran Saleemi, and Soumyadip Sengupta. 2022. Robust high-resolution video matting with temporal guidance. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision . 238–247

2022
[21]

Qinglin Liu, Shengping Zhang, Quanling Meng, Ru Li, Bineng Zhong, and Liqiang Nie. 2023. Rethinking Context Aggregation in Natural Image Matting. arXiv preprint arXiv:2304.01171 (2023)

2023 arXiv
[22]

Sebastian Lutz, Konstantinos Amplianitis, and Aljosa Smolic. 2018. Alphagan: Generative adversarial networks for natural image matting. arXiv preprint arXiv:1807.10088 (2018)

2018 arXiv
[23]

Yuval Nirkin, Iacopo Masi, Anh Tran Tuan, Tal Hassner, and Gerard Medioni
[24]

GyuTae Park, SungJoon Son, JaeYoung Yoo, SeHo Kim, and Nojun Kwak. 2022. Matteformer: Transformer-based image matting via prior-tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 11696– 11706

2022
[25]

Kwanyong Park, Sanghyun Woo, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. 2023. Mask-guided Matting in the Wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 1992–2001

2023
[26]

Yu Qiao, Yuhao Liu, Xin Yang, Dongsheng Zhou, Mingliang Xu, Qiang Zhang, and Xiaopeng Wei. 2020. Attention-guided hierarchical structure aggregation for image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 13676–13685

2020
[27]

Hongje Seong, Seoung Wug Oh, Brian Price, Euntai Kim, and Joon-Young Lee
[28]

Lingxue Song, Dihong Gong, Zhifeng Li, Changsong Liu, and Wei Liu. 2019. Occlusion robust face recognition based on mask learning with pairwise differ- ential siamese network. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 773–782

2019
[29]

Yanan Sun, Chi-Keung Tang, and Yu-Wing Tai. 2021. Semantic image matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 11120–11129

2021
[30]

Yanan Sun, Guanzhi Wang, Qiao Gu, Chi-Keung Tang, and Yu-Wing Tai. 2021. Deep video matting via spatio-temporal alignment and aggregation. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 6975–6984

2021
[31]

Zhonglin Sun, Chen Feng, Ioannis Patras, and Georgios Tzimiropoulos. 2024. LAFS: Landmark-based Facial Self-supervised Learning for Face Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 1639–1649

2024
[32]

Kenny TR Voo, Liming Jiang, and Chen Change Loy. 2022. Delving into high- quality synthetic face occlusion segmentation datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 4711–4720

2022
[33]

Zidu Wang, Xiangyu Zhu, Tianshuo Zhang, Baiqin Wang, and Zhen Lei. 2024. 3D Face Reconstruction with the Geometric Guidance of Facial Part Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 1672–1682

2024
[34]

W. A. Wright. 1999. Bayesian approach to neural-network modeling with input uncertainty. IEEE transactions on neural networks 10 6 (1999), 1261–70. https: //api.semanticscholar.org/CorpusID:28445688

1999
[35]

Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. 2017. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition . 2970–2979

2017
[36]

Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaoou Tang. 2016. Wider face: A face detection benchmark. In Proceedings of the IEEE conference on computer vision and pattern recognition . 5525–5533

2016
[37]

Xiangnan Yin and Liming Chen. 2022. FaceOcc: A diverse, high-quality face occlusion dataset for human face extraction. arXiv preprint arXiv:2201.08425 (2022)

2022 arXiv
[38]

Xiangnan Yin, Di Huang, Zehua Fu, Yunhong Wang, and Liming Chen. 2023. Segmentation-reconstruction-guided facial image de-occlusion. In 2023 IEEE 17th International Conference on Automatic Face and Gesture Recognition (FG) . IEEE, 1–8

2023
[39]

Qihang Yu, Jianming Zhang, He Zhang, Yilin Wang, Zhe Lin, Ning Xu, Yutong Bai, and Alan Yuille. 2021. Mask guided matting via progressive refinement network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 1154–1163

2021
[40]

Xiong Zhang, Hongsheng Huang, Jianchao Tan, Hongmin Xu, Cheng Yang, Guozhu Peng, Lei Wang, and Ji Liu. 2021. Hand image understanding via deep multi-task learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 11281–11292

2021
[41]

Yunke Zhang, Chi Wang, Miaomiao Cui, Peiran Ren, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, Qixing Huang, and Weiwei Xu. 2021. Attention-guided tempo- rally coherent video object matting. In Proceedings of the 29th ACM International Conference on Multimedia. 5128–5137. Supplemen...

2021
[45]

Segmentation Deep shadow, Strong makeup, Tattoo, Tongue, Eyeglass frame, Color lens, Mirror reflection Colorless transparent lens, Beard, Scalp
[46]

Uncertainty-Guided Face Matting for Occlusion-Aware Face Transformation

Segmentation Sunglass, Ear, Eyeglass including the colorless transparent lens Shadow, Strong makeup, Tattoo, Beard, Scalp Ours Matting Beard, Color lens, Ear, Eyeglass frame Colorless transparent lens, Shadow, Strong makeup, Tattoo, Scalp Table 9: Partition of each datasets da...
[2018]

In 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018)

On face segmentation, face swapping, and face perception. In 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018) . IEEE, 98–105

2018
[2022]

In European Conference on Computer Vision

One-trimap video matting. In European Conference on Computer Vision . Springer, 430–448
[2024]

In European Conference on Computer Vision

Diffusion for natural image matting. In European Conference on Computer Vision. Springer, 181–199

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. 2017. SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 39, 12 (2017), 2481–2495. doi:10.1109/TPAMI.2016.2644615

work page arXiv 2017

[2] [2]

Xavier P Burgos-Artizzu, Pietro Perona, and Piotr Dollár. 2013. Robust face landmark estimation under occlusion. In Proceedings of the IEEE international conference on computer vision . 1513–1520

work page 2013

[3] [3]

Huanqia Cai, Fanglei Xue, Lele Xu, and Lili Guo. 2022. Transmatting: Enhanc- ing transparent objects matting with transformers. In European Conference on Computer Vision. Springer, 253–269

work page 2022

[4] [4]

Guowei Chen, Yi Liu, Jian Wang, Juncai Peng, Yuying Hao, Lutao Chu, Shiyu Tang, Zewu Wu, Zeyu Chen, Zhiliang Yu, et al. 2022. PP-matting: high-accuracy natural image matting. arXiv preprint arXiv:2204.09433 (2022)

work page Pith review arXiv 2022

[5] [5]

Steve Chiu, Zhong Jin, and Yingjie Gu. 2015. Active learning combining uncer- tainty and diversity for multi-class image classification. IET Computer Vision 9 (06 2015), 400–407. doi:10.1049/iet-cvi.2014.0140

work page arXiv 2015

[6] [6]

Jiwoong Choi, Dayoung Chun, Hyun Kim, and Hyuk-Jae Lee. 2019. Gaussian YOLOv3: An Accurate and Fast Object Detector Using Localization Uncertainty for Autonomous Driving. 2019 IEEE/CVF International Conference on Computer Vi- sion (ICCV) (2019), 502–511. https://api.semanticscholar.org/CorpusID:104292012

work page 2019

[7] [7]

Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and An- drea Vedaldi. 2014. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition . 3606–3613

2014

[8] [8]

Kaiwen Cui, Rongliang Wu, Fangneng Zhan, and Shijian Lu. 2023. Face Trans- former: Towards High Fidelity and Accurate Face Swapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. 668–677

work page 2023