Pith. sign in

REVIEW 3 major objections 5 minor 47 references

Parallax Portrait Matting

T0 review · 3 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read A casually captured second frame with slight viewpoint change recovers finer portrait mattes and cleaner foreground colors than single-image methods.

desk verdict Practical two-frame burst matting with a sensible asymmetric fusion design; synthetic numbers are clean, real-world support is still only qualitative. read the letter →

arxiv 2607.11205 v1 pith:QGFOHB7L submitted 2026-07-13 cs.CV

Xin Cai , Jiawen Chen , Lars Jebe , Tianfan Xue , Zhoutong Zhang This is my paper
classification cs.CV
keywords imagemattingportraitparallaxburstphotographyopticalflowalphamatteforegroundcolorcross-attention
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

Image matting tries to separate a portrait subject from its background into an opacity map and a clean foreground color, but the problem is underdetermined when both layers are textured. Single-image networks lean on learned priors and still fail on hair and cluttered scenes; stronger methods demand green screens or other special capture. This paper argues that a second photograph taken a few millimeters away, as already happens in burst photography, supplies useful extra constraints through foreground-background parallax. The method estimates separate motion fields for each layer, then feeds a background-aligned pair into a network for direct fusion while routing a noisier foreground-aligned pair through cross-attention for residual correction. On synthetic and real portrait pairs the approach produces sharper alpha maps and substantially more accurate premultiplied colors than strong single-image and video baselines, without specialized hardware.

What carries the argument

Asymmetric two-branch alignment: the background-aligned image pair is treated as trustworthy pixel-level evidence for direct fusion, while the foreground-aligned pair is injected only through cross-attention as a soft auxiliary cue that can correct residual motion errors without hard pixel fusion.

What would settle it

Capture real portrait bursts with visible hair motion or large subject movement, feed them through the full pipeline, and check whether alpha and foreground-color errors fall below the single-image baseline on the same frames; if the second view no longer helps, the claim fails.

Watch

Extended reading notes

Core claim

A second casually captured frame with slight viewpoint change, processed by direct fusion of the more reliable background-aligned pair and feature-level cross-attention compensation from the noisier foreground-aligned pair, recovers finer alpha details and substantially more accurate premultiplied foreground colors than strong single-image matting baselines on challenging portrait cases.

Load-bearing premise

Real capture can be modeled well enough by two locally smooth warping fields, one for the subject and one for the background, with residual errors handled by the network rather than perfect optical flow.

Editorial extensions

If this is right

  • Casual burst photography already present on phones becomes a usable matting signal without green screens or extra hardware.
  • Foreground color estimates become clean enough for reliable compositing onto new backgrounds.
  • When parallax is weak or alignment fails, the model still falls back to single-image behavior rather than producing worse results.
  • Portrait matting pipelines can treat background alignment as the primary cue and foreground alignment only as soft correction.

Reading between the lines

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

  • The same asymmetric trust pattern—prefer the smoother, farther layer for hard fusion—could transfer to other layered inverse problems such as reflection separation or multi-layer depth.
  • Jointly training the motion estimator with the matting network, rather than freezing an off-the-shelf flow model, is a natural next step the authors flag as future work.
  • If the method generalizes beyond portraits, ordinary two-frame bursts could replace specialized multi-view or focal-stack capture for everyday object matting.
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

3 major / 5 minor

Summary. The paper proposes Parallax Portrait Matting, a two-frame portrait matting method that exploits slight camera-motion-induced foreground/background parallax available in burst capture. Given a base and alternate frame, it estimates trimaps (BiRefNet + fixed erode/dilate), approximates separate foreground and background motion by nearest-neighbor inpainting of off-the-shelf optical flow in uncertain regions, and feeds a dual-branch network: background-aligned views for direct pixel fusion and foreground-aligned views as an auxiliary cue via cross-attention. The network predicts alpha and premultiplied foreground of the base frame. Training is fully synthetic (composites from P3M/HHM foregrounds, BG-20K backgrounds, affine parallax + noise, layer-diffusion pseudo-F). On held-out synthetic tests the method beats trimap-free, trimap-based, and short-video baselines under a shared trimap, with especially large gains in MSE(αF); real-world support is qualitative on fixed-setting RAW pairs and commercial comparisons.

Significance. If the practical two-frame claim holds under real capture, the work is a useful middle ground between single-image matting and specialized multi-signal setups (green screen, polarization, clean background, arrays). The asymmetric design—trusting background alignment for fusion while treating foreground alignment only as feature-level compensation—is a clear, well-motivated engineering contribution, and the ablations (Table 2, Fig. 8) cleanly separate the two cues and show graceful degradation when parallax is absent or noisy. Stronger foreground-color accuracy is particularly valuable for compositing. The paper is honest about limitations (detached motion, larger motion, residual ambiguity). The main open question is how far the synthetic gains transfer to the casual-burst regime the abstract advertises.

major comments (3)
  1. Sec. 5.1–5.3 and Table 1 vs. Figs. 6–7: Quantitative support for the central claim is confined to synthetic composites (random affine warps, ~10 px noise, layer-diffusion pseudo-F, histogram equalization). Real-world evaluation is only qualitative on a small number of fixed-setting RAW pairs, with no real alpha/F ground truth and no real SAD/MSE/Conn/Grad or MSE(αF). Because the abstract and introduction frame the contribution as a practical burst cue, the domain gap between synthetic motion and real residual subject/hair motion is load-bearing. At minimum the paper needs either (i) a real quantitative protocol (e.g., multi-annotator or multi-view pseudo-GT on a held-out real set) or (ii) a clearly scoped claim that quantitative gains are synthetic and real evidence is qualitative, plus more systematic real failure cases.
  2. Sec. 3 (Mostly static scene) and Sec. 4.2: Motion is frozen off-the-shelf GMFlow with nearest-neighbor inpainting from certain trimap regions; it is never jointly optimized and residual non-rigid motion is left entirely to the network. Table 2 row (5) adds synthetic flow noise and still beats the single-image baseline, which is encouraging, but does not measure real residual motion (wind, hair, slight subject movement) that the paper itself flags as the hard case. A controlled real-motion stress test or an end-to-end / refinement experiment would substantially strengthen the robustness claim that underpins the asymmetric architecture.
  3. Sec. 5.2 baselines: Video methods (MaGGIe, MatAnyone) are evaluated as adjacent-frame references rather than full-sequence temporal models. That is a reasonable framing, but the paper should state more explicitly how many frames and what temporal window each baseline receives, and ideally include a two-frame-only ablation of those methods (or a simple two-frame flow-warped single-image baseline) so that the reported gap can be attributed to explicit parallax reasoning rather than to any extra temporal capacity.
minor comments (5)
  1. Fig. 4 notation is dense (warped trimaps, double-warped companion frame); a short input-tensor legend or equation block listing exact tensors per branch would help reproducibility.
  2. Sec. 4.1: the fixed 100 px erode/dilate (200 px uncertain band) is generous; a short sensitivity note would clarify dependence on BiRefNet quality.
  3. Sec. 5.1: layer-diffusion pseudo-foreground colors are used for supervision; any known bias of that generator on hair/semi-transparent regions should be acknowledged when interpreting MSE(αF).
  4. Related work: a slightly sharper contrast with background-matting and camera-array / color-filter methods would better locate the “no extra setup” claim.
  5. Minor polish: consistent hyphenation of “pre-multiplied” / “premultiplied”; ensure all figure callouts match the final layout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: standard supervised two-frame matting with external baselines and held-out evaluation.

full rationale

This is an empirical computer-vision paper, not a first-principles derivation. The matting equation (I = αF + (1-α)B) and the two-frame extension with warps MF and MB are standard underdetermined inverse problems; the paper does not claim to derive a unique closed-form solution from them. Motion fields are obtained from off-the-shelf GMFlow plus nearest-neighbor inpainting of uncertain trimap regions (Sec. 4.2); they are not fitted to the evaluation metrics and then re-presented as predictions. The network is trained with ordinary L1/Laplacian/gradient/composition losses on synthetic composites (Sec. 4.4–5.1) and is compared against external single-image and video baselines (MODNet, MatteFormer, MatAnyone, etc.) on held-out synthetic sets (Table 1) and real captures (Figs. 6–7). Ablations (Table 2) remove inputs rather than re-label fitted parameters. No uniqueness theorem, self-citation chain, or ansatz is load-bearing for the central claim. The reader’s and skeptic’s concerns about synthetic-only training and qualitative real evaluation are domain-gap / evaluation-strength issues, not circularity. Score 0 is therefore appropriate.

Assumptions & free parameters 5 free parameters · 6 assumptions · 1 invented entities

The claim rests on the classical linear matting equation plus three operating-regime assumptions (parallax, mostly static scene, consistent camera settings) and an engineering approximation that uncertain-region motion equals nearest-neighbor extrapolation of certain-region flow. Network and trimap hyperparameters are free choices that affect absolute numbers but are not fitted to the real-world qualitative claims. No new physical entities are postulated.

free parameters (5)
  • trimap uncertain-band width (erode/dilate 100 px)
    Hand-chosen generous band around BiRefNet mask; used at inference and affects which pixels are treated as uncertain.
  • training trimap random dilate/erode iterations (60–120)
    Hand range used to synthesize trimaps during training.
  • patch size 448 and overlap 224 with Gaussian merge
    Hand-chosen tiling parameters for high-res training/inference.
  • synthetic motion noise (~10 px) and affine simulation
    Hand-chosen noise level so the network never sees perfect alignment in training.
  • optimizer / LR / epochs (AdamW 5e-5, 50 epochs, 100k patches/epoch)
    Standard training hyperparameters that determine the fitted network weights.
assumptions (6)
  • standard math Linear compositing: I = αF + (1−α)B (and the two-frame warped form in Sec. 3).
    Classical matting equation assumed throughout formulation and composition loss.
  • domain assumption Foreground is closer than background so small camera motion induces usable parallax; when MF = MB the extra equation is linearly dependent.
    Sec. 3 operating regime; without parallax the second frame adds no new constraint.
  • domain assumption Scene is mostly static so layer motion can be modeled by two warping fields; residual non-rigid motion is tolerable.
    Sec. 3 “Mostly static scene”; load-bearing for the two-field motion model.
  • ad hoc to paper Motion in uncertain (mixed) regions can be approximated by nearest-neighbor extrapolation from certain regions of a single off-the-shelf flow field.
    Sec. 4.2; common smoothness assumption applied specifically as NN inpainting of GMFlow.
  • domain assumption Background alignment is systematically more reliable than foreground alignment in typical burst portraits, justifying asymmetric fusion.
    Central design premise stated in introduction and Sec. 4.3.
  • domain assumption Both frames share consistent exposure, white balance, and tone (or can be rendered so).
    Sec. 3 “Consistent camera settings”; required for direct photometric use of the second view.
invented entities (1)
  • Asymmetric dual-branch parallax matting network (bg-aligned direct fusion + fg-aligned cross-attention compensation)
    purpose: Exploit two-frame parallax while remaining robust to imperfect foreground motion estimates.
    Architectural construct introduced by the paper; not an external physical entity. Independent evidence is only the paper’s own ablations and benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Parallax Portrait Matting." pith.science (2026). https://pith.science/paper/QGFOHB7L

@misc{pith2026260711205,
  author       = {Pith},
  title        = {Pith review of: Parallax Portrait Matting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QGFOHB7L}},
  note         = {Machine review of arXiv:2607.11205}
}
read the original abstract

Image matting is highly ill-posed, especially when both the foreground and background are richly textured. While single-image matting methods learn strong priors from data, they often struggle on these challenging cases. Existing approaches improve results by requiring additional signals such as green screens, polarized lighting, or clean background images, but these typically rely on specialized capture setups. We present Parallax Portrait Matting, a practical two-frame matting method that uses a second image captured with slight viewpoint change. Such a setting arises naturally in burst photography, where small camera motion induces foreground-background parallax and provides complementary observations for matting. Our pipeline estimates trimaps and foreground/background motion, then constructs aligned views for prediction. To handle imperfect motion estimation, the network uses the background-aligned pair for direct fusion and the foreground-aligned cue through cross-attention for error compensation. Experiments show that our method recovers finer details and more accurate foreground colors than strong single-image matting baselines on challenging portrait cases.

Figures

Figures reproduced from arXiv: 2607.11205 by the authors.

Figure 1
Figure 1. Our matting method exploits camera-motion-induced parallax between the foreground and the background. It takes two frames as input, each taken with a slightly different camera location, and predicts both a pre-multiplied foreground image and an alpha map. Trained on public datasets, our method produces a cleaner foreground with more details than closed-source commercial solutions like remove.bg. line of work tackles… view at source ↗
Figure 2
Figure 2. Our motion estimation pipeline. Given two frames, we first estimate optical flow between the two, together with a trimap generated as described in Sec. 4.1. Since flow estimation in overlapping areas is often incorrect, we inpaint them by their nearest neighbor in non-overlapping regions. Background-aligned examples Foreground-aligned examples [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. An example of foreground and background alignment. Background-aligned patches only contain foreground subject motion, and the dis-occluded background in the alternate frame provides more context. Foreground-aligned patches fixate on the subject, which directly helps in separating the foreground. background motion MB 0→1 , we get an image I B 0→1 that differs only from I1 in fore￾ground regions. Regions that are orig… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Overview of our framework. The bottom branch is the primary prediction path and uses the base frame together with a background-aligned alternate frame, which provides reliable pixel-aligned evidence for matting. The top branch uses a foreground￾aligned view as an auxil…
Figure 5
Figure 5. Figure 5: Qualitative results on synthetic test sets. color accuracy (MSE of αF), where our method reduces error by 35–45% over the best baseline—critical for downstream compositing. Qualitative Results [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results on real-world images. Our method produces higher-quality alpha mattes, more accurate foreground colors, and finer structural details compared to existing methods. 5.3 Evaluation on Real-World Data We further assess the robustness of our method in re…
Figure 7
Figure 7. Figure 7: Comparison against closed-source commercial solutions (Adobe Photoshop and Remove.bg) on real-world images. Despite competing with proprietary systems, our method recovers significantly more detail in fine structures such as hair strands. 5.4 Ablation Study [PITH_FULL…
Figure 8
Figure 8. Figure 8: Ablation on real-world cases. Removing background-aligned burst loses fine details; removing foreground-aligned burst causes foreground-background confusion. firming graceful degradation when parallax is absent. Finally, we perturb the estimated foreground flow by intr…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 2 canonical work pages

  1. [1]

    ACM Transactions on Graphics (TOG)35(5), 152:1–152:12 (2016)

    Aksoy, Y., Aydin, T.O., Pollefeys, M., Smolić, A.: Interactive high-quality green- screen keying via color unmixing. ACM Transactions on Graphics (TOG)35(5), 152:1–152:12 (2016)

    2016

  2. [2]

    In: ACM SIGGRAPH Asia 2008 Papers

    Bando, Y., Chen, B.Y., Nishita, T.: Extracting depth and matte using a color- filtered aperture. In: ACM SIGGRAPH Asia 2008 Papers. SIGGRAPH Asia ’08, Association for Computing Machinery, New York, NY, USA (2008).https://doi. org/10.1145/1457515.1409087,https://doi.org/10.1145/1457515.1409087

    work page doi:10.1145/1457515.1409087 2008

  3. [3]

    Bhat,G.,Danelljan,M.,VanGool,L.,Timofte,R.:Deepburstsuper-resolution.In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 9209–9218 (2021)

    2021

  4. [4]

    In: NeurIPS (2024)

    Cai, X., You, Z., Zhang, H., Liu, W., Gu, J., Xue, T.: PhoCoLens: Photorealistic and consistent reconstruction in lensless imaging. In: NeurIPS (2024)

    2024

  5. [5]

    In: CVPR (2026)

    Cai, X., You, Z., Zhang, Z., Xue, T.: DA-VAE: Plug-in latent compression for diffusion via detail alignment. In: CVPR (2026)

    2026

  6. [6]

    In: Proceedings of the 26th ACM international conference on Multimedia

    Chen, Q., Ge, T., Xu, Y., Zhang, Z., Yang, X., Gai, K.: Semantic human matting. In: Proceedings of the 26th ACM international conference on Multimedia. pp. 618– 626 (2018)

    2018

  7. [7]

    In: Proceedings of the 29th annual conference on Computer graphics and interactive techniques

    Chuang, Y.Y., Agarwala, A., Curless, B., Salesin, D.H., Szeliski, R.: Video matting of complex scenes. In: Proceedings of the 29th annual conference on Computer graphics and interactive techniques. pp. 243–248 (2002)

    2002

  8. [8]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    arXiv 2010

Show all 47 references
  1. [9]

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

    Enomoto, K., Rhodes, T., Price, B., Miller, G.: Polarmatte: Fully computational ground-truth-quality alpha matte extraction for images and video using polarized screen matting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3901–3909 (2024)

    2024

  2. [10]

    Englewood Cliffs, NJ: Prentice-Hall (1977)

    Fry, R., Fourzon, P.: The saga of special effects. Englewood Cliffs, NJ: Prentice-Hall (1977)

    1977

  3. [11]

    In: CVPR (2011)

    He, K., Rhemann, C., Rother, C., Tang, X., Sun, J.: A global sampling method for alpha matting. In: CVPR (2011)

    2011

  4. [12]

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

    Huynh, C., Oh, S.W., Shrivastava, A., Lee, J.Y.: Maggie: Masked guided grad- ual human instance matting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3870–3879 (2024)

    2024

  5. [13]

    ACM Transactions on Graphics (TOG)25(3), 779–786 (2006)

    Joshi, N., Matusik, W., Avidan, S.: Natural video matting using camera arrays. ACM Transactions on Graphics (TOG)25(3), 779–786 (2006)

    2006

  6. [14]

    IEEE Computer Graphics and Applications27(2), 43–52 (2007)

    Joshi, N., Matusik, W., Avidan, S., Pfister, H., Freeman, W.T.: Exploring defo- cus matting: Nonparametric acceleration, super-resolution, and off-center matting. IEEE Computer Graphics and Applications27(2), 43–52 (2007)

    2007

  7. [15]

    In: AAAI (2022)

    Ke, Z., Sun, J., Li, K., Yan, Q., Lau, R.W.: MODNet: Real-time trimap-free por- trait matting via objective decomposition. In: AAAI (2022)

    2022

  8. [16]

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

    Li, J., Goel, V., Ohanyan, M., Navasardyan, S., Wei, Y., Shi, H.: Vmformer: End- to-end video matting with transformer. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 6678–6687 (2024)

    2024

  9. [17]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Li, J., Jain, J., Shi, H.: Matting anything. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 1775–1785 (2024)

    2024

  10. [18]

    In: Pro- ceedings of the 29th ACM international conference on multimedia

    Li, J., Ma, S., Zhang, J., Tao, D.: Privacy-preserving portrait matting. In: Pro- ceedings of the 29th ACM international conference on multimedia. pp. 3501–3509 (2021) Parallax Portrait Matting 17

    2021

  11. [19]

    International Journal of Computer Vision130(2), 246–266 (2022)

    Li, J., Zhang, J., Maybank, S.J., Tao, D.: Bridging composite and real: towards end-to-end deep image matting. International Journal of Computer Vision130(2), 246–266 (2022)

    2022

  12. [20]

    In: AAAI (2020)

    Li, Y., Lu, H.: Natural image matting via guided contextual attention. In: AAAI (2020)

    2020

  13. [21]

    In: 2020 IEEE International Conference on Multimedia and Expo (ICME)

    Li, Y., Zhang, J., Zhao, W., Jiang, W., Lu, H.: Inductive guided filter: Real-time deep matting with weakly annotated masks on mobile devices. In: 2020 IEEE International Conference on Multimedia and Expo (ICME). pp. 1–6. IEEE (2020)

    2020

  14. [22]

    In: CVPR (2021)

    Lin, S., Ryabtsev, A., Sengupta, S., Curless, B.L., Seitz, S.M., Kemelmacher- Shlizerman, I.: Real-time high-resolution background matting. In: CVPR (2021)

    2021

  15. [23]

    In: WACV (2022)

    Lin, S., Yang, L., Saleemi, I., Sengupta, S.: Robust high-resolution video matting with temporal guidance. In: WACV (2022)

    2022

  16. [24]

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

    Liu, J., Yao, Y., Hou, W., Cui, M., Xie, X., Zhang, C., Hua, X.s.: Boosting seman- tic human matting with coarse annotations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8563–8572 (2020)

    2020

  17. [25]

    In: ICCV (2019)

    Lu, H., Dai, Y., Shen, C., Xu, S.: Context-aware image matting for simultaneous foreground and alpha estimation. In: ICCV (2019)

    2019

  18. [26]

    IJCV131(8), 2172–2197 (2023)

    Ma, S., Li, J., Zhang, J., Zhang, H., Tao, D.: Rethinking portrait matting with privacy preserving. IJCV131(8), 2172–2197 (2023)

    2023

  19. [27]

    In: CVPR (2022)

    Park, G., Son, S., Yoo, J., Kim, S., Kwak, N.: Matteformer: Transformer-based image matting via prior-tokens. In: CVPR (2022)

    2022

  20. [28]

    In: CVPR (2009)

    Rhemann,C.,Rother,C.,Wang,J.,Gelautz,M.,Kohli,P.,Rott,P.:Aperceptually motivated online benchmark for image matting. In: CVPR (2009)

    2009

  21. [29]

    In: CVPR (2020)

    Sengupta, S., Jayaram, V., Curless, B., Seitz, S.M., Kemelmacher-Shlizerman, I.: Background matting: The world is your green screen. In: CVPR (2020)

    2020

  22. [30]

    In: Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14

    Shen, X., Tao, X., Gao, H., Zhou, C., Jia, J.: Deep automatic portrait matting. In: Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14. pp. 92–107. Springer (2016)

    2016

  23. [31]

    In: Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques

    Smith, A.R., Blinn, J.F.: Blue screen matting. In: Proceedings of the 23rd Annual Conference on Computer Graphics and Interactive Techniques. p. 259–268. SIG- GRAPH ’96, Association for Computing Machinery, New York, NY, USA (1996). https://doi.org/10.1145/237170.237263,https:...

    1996 doi

  24. [32]

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

    Sun, Y., Tang, C.K., Tai, Y.W.: Human instance matting via mutual guidance and multi-instance refinement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2647–2656 (2022)

    2022

  25. [33]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition

    Sun, Y., Tang, C.K., Tai, Y.W.: Ultrahigh resolution image/video matting with spatio-temporal sparsity. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. pp. 14112–14121 (2023)

    2023

  26. [34]

    Sun, Y., Wang, G., Gu, Q., Tang, C.K., Tai, Y.W.: Deep video matting via spatio- temporalalignmentandaggregation.In:ProceedingsoftheIEEE/CVFConference on Computer Vision and Pattern Recognition. pp. 6975–6984 (2021)

    2021

  27. [35]

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

    Tang, J., Aksoy, Y., Oztireli, C., Gross, M., Aydin, T.O.: Learning-based sam- pling for natural image matting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3055–3063 (2019)

    2019

  28. [36]

    In: ECCV (2020)

    Teed, Z., Deng, J.: RAFT: Recurrent all-pairs field transforms for optical flow. In: ECCV (2020)

    2020

  29. [37]

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

    Wang, T., Liu, S., Tian, Y., Li, K., Yang, M.H.: Video matting via consistency- regularized graph neural networks. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4902–4911 (2021) 18 X. Cai et al

    2021

  30. [38]

    In: ACM SIGGRAPH 2024 Conference Papers

    Wang, Z., Li, B., Wang, J., Liu, Y.L., Gu, J., Chuang, Y.Y., Satoh, S.: Matting by generation. In: ACM SIGGRAPH 2024 Conference Papers. pp. 1–11 (2024)

    2024

  31. [39]

    In: Proceedings of the 36th International Conference on Neural Information Processing Systems

    Weinzaepfel, P., Leroy, V., Lucas, T., Brégier, R., Cabon, Y., Arora, V., Antsfeld, L., Chidlovskii, B., Csurka, G., Revaud, J.: Croco: self-supervised pre-training for 3d vision tasks by cross-view completion. In: Proceedings of the 36th International Conference on Neural Inf...

    2022

  32. [40]

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

    Xu, H., Zhang, J., Cai, J., Rezatofighi, H., Tao, D.: Gmflow: Learning optical flow via global matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8121–8130 (2022)

    2022

  33. [41]

    In: CVPR (2017)

    Xu, N., Price, B., Cohen, S., Huang, T.: Designing effective inter-pixel information flow for natural image matting. In: CVPR (2017)

    2017

  34. [42]

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

    Yaman, D., Ekenel, H.K., Waibel, A.: Alpha matte generation from single input for portrait matting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 696–705 (2022)

    2022

  35. [43]

    In: CVPR (2025)

    Yang, P., Zhou, S., Zhao, J., Tao, Q., Loy, C.C.: MatAnyone: Stable video matting with consistent memory propagation. In: CVPR (2025)

    2025

  36. [44]

    Information Fusion103, 102091 (2024)

    Yao, J., Wang, X., Yang, S., Wang, B.: Vitmatte: Boosting image matting with pre-trained plain vision transformers. Information Fusion103, 102091 (2024)

    2024

  37. [45]

    In: CVPR (2021)

    Yu, Q., Zhang, J., Zhang, H., Wang, Y., Lin, Z., Xu, N., Bai, Y., Yuille, A.: Mask guided matting via progressive refinement network. In: CVPR (2021)

    2021

  38. [46]

    arXiv preprint arXiv:2402.17113 (2024)

    Zhang, L., Agrawala, M.: Transparent image layer diffusion using latent trans- parency. arXiv preprint arXiv:2402.17113 (2024)

    2024 arXiv

  39. [47]

    CAAI Artificial Intelligence Research3, 9150038 (2024)

    Zheng, P., Gao, D., Fan, D.P., Liu, L., Laaksonen, J., Ouyang, W., Sebe, N.: Bilat- eral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research3, 9150038 (2024)

    2024

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.