Pith. sign in

REVIEW 5 major objections 6 minor 35 references

Post-Training Quantization for Video Matting

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Post-training quantization can bring a video matting model from full precision down to 4-bit weights and activations while keeping near-full-precision alpha mattes and cutting FLOPs by 8x, the paper claims.

desk verdict Novel PTQ pipeline for video matting, but the main evidence tables are self-declared placeholders and the OFA ablation contradicts the text, so the paper is not reviewable as submitted. read the letter →

arxiv 2506.10840 v1 pith:EJXBY5T3 submitted 2025-06-12 cs.CV cs.AI

Tianrui Zhu , Houyuan Chen , Ruihao Gong , Michele Magno , Haotong Qin , Kai Zhang This is my paper · ORCID
classification cs.CVcs.AI
keywords post-trainingquantizationvideomattingalphamatteopticalflowtemporalcoherenceglobalaffinecalibrationblock-wiseoptimizationlow-bit
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 targets post-training quantization for video matting, the task of estimating per-pixel foreground opacity (the alpha matte) in every frame. It claims that existing PTQ methods, designed mainly for image classification, lose accuracy and temporal stability when applied to recurrent video matting models. PTQ4VM is presented as the first systematic PTQ framework for video matting: block-wise initial quantization, then a global affine calibration of weights and activations, plus an optical-flow-assisted loss that warps the previous frame's alpha prediction to regularize the current one. On the RVM model, 4-bit PTQ4VM is reported to stay close to the full-precision baseline across alpha errors and temporal-coherence metrics while cutting FLOPs by 8x. The same global calibration can be added to existing PTQ methods, reducing their error on video matting as well.

What carries the argument

The mechanism is the combination of two components. Global Affine Calibration applies per-layer affine transforms $W'_{f,q,i} = \gamma_i W_{f,q,i} + \beta_i$ to the initially quantized folded weights and jointly optimizes activation scales against the final $\alpha$ output, then absorbs the learned parameters into the quantization scales so inference cost is unchanged. Optical Flow Assistance uses the loss $L_{\mathrm{OFA}} = \|\hat{\alpha}_t - \mathrm{Warp}(\hat{\alpha}_{t-1}, F_{t-1\to t})\|_1$, where $F_{t-1\to t}$ is a precomputed RAFT optical flow field on the small calibration set; it regularizes the model toward temporal-semantic coherence without any flow computation at inference time. GAC compensates for the accumulated statistical shift and its reshaping by nonlinearities, while OFA supplies the temporal prior that stabilizes the recurrent structure during calibration.

What would settle it

Run the W4A4 calibration with and without the optical-flow loss on a clip selected for fast occluding motion, and compare per-frame alpha error against the error of the warped prior itself: if the flow-assisted model is worse exactly where the warped prior is wrong, the temporal-prior assumption fails.

Watch

Extended reading notes

Core claim

The central claim is that video matting post-training quantization can be made near-lossless at very low bit-widths by treating quantization error as a global statistical distortion rather than only a per-layer reconstruction problem. After Batch Normalization folding, quantized weights and activations accumulate mean and scale shifts that make the folded weights suboptimal for the activations they actually encounter; PTQ4VM's Global Affine Calibration learns per-layer scale and shift parameters on the quantized folded weights, together with activation scales, optimized end-to-end against the final alpha output. On top of this, Optical Flow Assistance warps the previous frame's predicted alpha with RAFT flow to produce a temporal prior and adds an L1 loss between the current prediction and that prior, pushing the quantized recurrent model toward temporally coherent mattes. The claimed outcome is state-of-the-art PTQ accuracy across 4- to 8-bit settings on the VM video matting dataset and on the held-out D646 image matting dataset, with W4A4 performance close to the full-precision RVM baseline.

Load-bearing premise

The method assumes that moving the previous frame's predicted foreground-opacity map according to measured pixel motion gives a reliable guide to the current frame's true map; when motion is wrong at occlusions, fast movement, or thin structures, the flow-based regularizer steers the quantized model toward the wrong target.

Editorial extensions

If this is right

  • Video matting models such as RVM can be deployed at 4-bit weights and activations with alpha quality close to the full-precision version, cutting computation to one eighth of the original FLOPs.
  • The Global Affine Calibration stage can be bolted onto existing PTQ methods; on BRECQ and QDrop it reduces alpha error, with the largest gains at W4A4.
  • A calibration set of only 256 frames, sampled at 4-frame intervals from 64 clips, is enough to drive the whole two-stage calibration, which runs on a single 24 GB GPU.
  • The quantized model keeps its accuracy on D646, an image matting dataset not used in calibration, indicating the calibrated quantization parameters generalize beyond the calibration distribution.
  • Temporal coherence, quantified by DTSSD, improves when the optical-flow loss is active during calibration, which directly addresses the flicker problem of recurrent quantized networks.

Reading between the lines

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

  • Beyond the paper, GAC's premise that PTQ error manifests as a layer-wise affine statistical shift after BN folding should transfer to other recurrent video tasks such as video depth estimation or video segmentation, where the same recurrent instability appears.
  • Beyond the paper, replacing RAFT with a lighter or learned flow estimator would make OFA self-contained and cheaper, at the cost of a less accurate warped prior on fast motion.
  • Beyond the paper, the 4-bit results on one recurrent backbone leave open whether the same recipe holds for transformer-based video matting models, whose attention layers may produce larger activation-range shifts.
  • Beyond the paper, the OFA component implies that calibration clips should be chosen to contain motion and occlusion diversity; static clips would under-exercise the temporal prior the method relies on.
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

5 major / 6 minor

Summary. PTQ4VM is a post-training quantization framework for video matting. It combines a block-wise initial quantization stage (BIQ) that reconstructs blocks against the full-precision model, a Global Affine Calibration stage (GAC) that learns per-layer affine corrections to weights and activation scales by minimizing the task loss on alpha predictions, and an Optical Flow Assistance component (OFA) that warps the previous predicted alpha with RAFT optical flow and adds an L1 temporal-consistency term. The paper evaluates on RVM under W8A8, W4A8, and W4A4 on the VM video matting dataset and the D646 image matting dataset, reporting large gains over MSE-based PTQ, BRECQ, and QDrop, and claims near full-precision performance at 4-bit with 8x FLOP savings.

Significance. Video matting deployment on edge devices is practically important, and PTQ for this task is under-explored. The two-stage design addresses a real stability issue, and GAC is a simple, method-agnostic correction that can be applied on top of existing PTQ methods. The use of a held-out image matting dataset and a temporal coherence metric is appropriate and a strength of the evaluation. However, the evidence as submitted is incomplete: the main tables carry placeholder captions, no variance or significance testing is reported, the OFA ablation is internally inconsistent for QDrop, calibration/evaluation overlap on VM is not ruled out, and only one architecture is evaluated. The core ideas are plausible, but the central quantitative claims are not yet established by the manuscript as it stands.

major comments (5)
  1. [Section 4, Tables 1 and 2] Both table captions state "Results for FLOPs, Params, and metrics are to be filled by the user." This placeholder text is a self-declared statement that the quantitative results are not finalized in the submitted manuscript. Since the abstract's 10-20% error-reduction claim and the near-full-precision 4-bit claim rest entirely on these tables, the central evidence is missing as presented. The authors must replace the placeholder captions with final reported values and state the exact configuration used for every number.
  2. [Section 4.2, Table 2] The text states that adding OFA to the second-stage calibration yields "further improvements in accuracy ... for both methods." For QDrop, the reported numbers contradict this claim: at W4A8, QDrop+GAC+OFA has higher MSE (4.54 vs 4.43) and higher Grad (3.49 vs 3.36) than QDrop+GAC; at W4A4, MSE increases from 11.85 to 13.80 and Grad from 6.90 to 7.48. The paper should correct the claim or specify which metrics support it, and should analyze why OFA helps BRECQ but hurts QDrop on these metrics.
  3. [Sections 4.1-4.2] All reported results are single-run point estimates with no error bars, number of random seeds, or significance tests. Many cross-method differences at W8A8 are small (e.g., Table 1 shows MAD 6.03 vs 6.08 and MSE 1.29 vs 1.27), so it is unclear which differences are meaningful. Please report mean and standard deviation over at least three independent calibration runs, or perform a paired test on held-out video clips.
  4. [Section 4, Data Construction] The calibration set is sampled "from the VM dataset," and evaluation is also performed on the VM video matting dataset, but the manuscript never states whether the 64 calibration clips are disjoint from the evaluation clips. If calibration frames come from the test portion of VM, the reported gains on VM could be inflated by calibration-set leakage. Please specify the exact train/test split used and confirm that the calibration clips are excluded from evaluation.
  5. [Section 4 and Abstract] The framework is described as general and as the first systematic PTQ approach for video matting, but the experiments cover only RVM. A single architecture is insufficient to support the "general framework" claim. Please add at least one additional video matting architecture (e.g., VMFormer or another recurrent matting model) or explicitly restrict the claims to RVM.
minor comments (6)
  1. [Section 4.1] The text mentions "a reduction of approximately 20%" for W4A4 on VM, but Table 1 shows SAD 20.33 for PTQ4VM vs 24.36 for QDrop, which is about a 16.5% reduction; please report the relative improvements precisely and state which metrics are used for each percentage.
  2. [Appendix A.4] The description of the Stage-2 objective says "implicitly the influence of the OFA loss." Please write the total loss explicitly, e.g., L = L_alpha + lambda * L_OFA, and state how L_alpha is computed.
  3. [Appendix A.3, Figure 7] The per-frame error curves would be more informative with axis labels, a legend, error bars or multiple clip traces, and a statement of how many test sequences the average is taken over.
  4. [Equation (2)] The notation W'_{f,q,i} = gamma_i * W_{f,q,i} + beta_i suggests the affine correction is applied after quantization; please clarify whether gamma_i and beta_i are folded into the quantized weight scale and zero-point or applied in floating point during inference.
  5. [Section 1] The statement that the authors are "the first to focus on the BN layer" in PTQ is too strong. Prior bias-correction and BatchNorm-related calibration methods (including the cross-layer equalization and bias correction cited as [32]) should be acknowledged more precisely.
  6. [References] The citation for Quantization-Aware Training in the introduction is [35], which is BiMatting, a binarization method rather than a general QAT reference. Please cite an appropriate QAT work or remove the misleading citation.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: GAC and OFA are standard calibration/regularization, and the self-citations are non-load-bearing baselines; the table placeholders are an evidence gap, not a circular step.

full rationale

The derivation chain is self-contained with respect to the reported claims. BIQ is block-wise MSE reconstruction, GAC fits layer-wise affine parameters by minimizing the final task loss against ground-truth alpha mattes (Section 3.3, Eqs. 2-3), and OFA adds an L1 temporal-consistency term between the current prediction and the flow-warped previous prediction (Section 3.4, L_OFA = ||alpha_t - Warp(alpha_{t-1}, F)||_1). None of these steps define the evaluation metrics in terms of the fitted parameters. The GAC parameters are fitted to a 256-frame calibration set, which is standard PTQ calibration rather than a disguised prediction: the evaluation is reported on VM and on D646, which the paper explicitly identifies as an unseen generalization set. The OFA loss is self-referential in form, since the prior is the model's own previous output, but it is a consistency regularizer, not a definition of the target; final errors are computed against ground-truth alpha, so the headline accuracy claim does not reduce to LOFA by construction. The self-citations to BRECQ and QDrop are used as comparison baselines from the same group, but the paper does not rely on them to justify the correctness of its own framework, and no uniqueness theorem or ansatz is imported from prior self-cited work. Two evidentiary problems are real but non-circular: Tables 1 and 2 captions state 'Results for FLOPs, Params, and metrics are to be filled by the user,' which is a missing-evidence artifact, and Section 4.2's claim that OFA yields 'further improvements in accuracy' for both methods conflicts with Table 2, where QDrop+GAC+OFA has higher MSE and Grad than QDrop+GAC at both W4A8 and W4A4. These issues affect the reliability of the quantitative conclusions, but they are not instances of a derivation reducing to its own inputs. Score 1 reflects the presence of minor non-load-bearing self-citations, not substantive circularity.

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

The central claim depends on a handful of hand-set hyperparameters and learned calibration parameters, all fitted on 256 calibration frames; the method itself is empirical, so the ledger is modest. No new physical entities are introduced; GAC affine parameters and the OFA loss are numerical and methodological components.

free parameters (5)
  • OFA loss weight lambda = 0.05
    Set by hand (Appendix A.4); controls strength of temporal regularization and affects reported ablations.
  • GAC per-layer affine parameters gamma_i, beta_i = Not reported per layer; distributions shown in Appendix A.2
    Learned per convolutional layer by minimizing alpha MSE on the 256-image calibration set (Section 3.3); central to the GAC component.
  • Stage-2 activation scales s'_a,i = Not reported
    Optimized jointly with gamma and beta during global calibration (Eq. 3, Section 3.3); final values not tabulated.
  • BIQ block partition = InvertedResidual, LRASPP, BottleneckBlock, UpsamplingBlock, OutputBlock, Projection
    Chosen manually for RVM (Section 4); no principled selection rule, affects Stage 1 outcome.
  • Stage hyperparameters = LR 4e-5 for BIQ, LR 1e-4 for GAC/OFA, 20k iterations per block, 50 epochs
    Hand-selected; no sensitivity analysis reported (Appendix A.4).
assumptions (5)
  • standard math The uniform affine quantization and dequantization formulas (Eq. 1) and BN folding (Section 3.1) are mathematically exact models of inference-time behavior.
    The paper relies on these identities to justify quantizing folded weights and activations.
  • domain assumption A calibration set of 256 frames (4 frames from 64 VM clips) is representative enough to tune quantization parameters that transfer to held-out VM and D646 data.
    Small calibration sets are standard in PTQ, but representativeness is asserted, not demonstrated (Appendix A.4).
  • domain assumption RVM is a representative video matting architecture, so results generalize to other video matting models.
    The framework is called 'general' but is tested only on RVM (Section 4).
  • domain assumption Alpha mattes are transported by optical flow: Warp(alpha_{t-1}, F_{t-1->t}) is a valid prior for alpha_t.
    OFA loss (Section 3.4) assumes pixel motion from RAFT moves matte boundaries correctly across frames, which can fail at occlusions and fast motion.
  • ad hoc to paper One scalar scale and one scalar shift per layer can compensate accumulated quantization distortion.
    The GAC method (Eq. 2) is a modeling choice, not derived from a quantization error bound or statistical theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Post-Training Quantization for Video Matting." pith.science (2026). https://pith.science/paper/EJXBY5T3

@misc{pith2026250610840,
  author       = {Pith},
  title        = {Pith review of: Post-Training Quantization for Video Matting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EJXBY5T3}},
  note         = {Machine review of arXiv:2506.10840}
}
read the original abstract

Video matting is crucial for applications such as film production and virtual reality, yet deploying its computationally intensive models on resource-constrained devices presents challenges. Quantization is a key technique for model compression and acceleration. As an efficient approach, Post-Training Quantization (PTQ) is still in its nascent stages for video matting, facing significant hurdles in maintaining accuracy and temporal coherence. To address these challenges, this paper proposes a novel and general PTQ framework specifically designed for video matting models, marking, to the best of our knowledge, the first systematic attempt in this domain. Our contributions include: (1) A two-stage PTQ strategy that combines block-reconstruction-based optimization for fast, stable initial quantization and local dependency capture, followed by a global calibration of quantization parameters to minimize accuracy loss. (2) A Statistically-Driven Global Affine Calibration (GAC) method that enables the network to compensate for cumulative statistical distortions arising from factors such as neglected BN layer effects, even reducing the error of existing PTQ methods on video matting tasks up to 20%. (3) An Optical Flow Assistance (OFA) component that leverages temporal and semantic priors from frames to guide the PTQ process, enhancing the model's ability to distinguish moving foregrounds in complex scenes and ultimately achieving near full-precision performance even under ultra-low-bit quantization. Comprehensive quantitative and visual results show that our PTQ4VM achieves the state-of-the-art accuracy performance across different bit-widths compared to the existing quantization methods. We highlight that the 4-bit PTQ4VM even achieves performance close to the full-precision counterpart while enjoying 8x FLOP savings.

Figures

Figures reproduced from arXiv: 2506.10840 by the authors.

Figure 1
Figure 1. Visual comparison of our PTQ4VM against Full-Precision (RVM) and QDrop . Our method [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our PTQ4VM. In Stage 1 (Block-wise Initial Quantization), we optimize functional blocks sequentially to accelerate convergence and enhance stability. In Stage 2 (Cross-Block Joint Calibration), we employ Global Affine Calibration (GAC) to compensate for distributional shifts; meanwhile, Optical Flow Assistance (OFA) guides the model to learn temporal-semantic coherence. The core optimization objective of… view at source ↗
Figure 3
Figure 3. Comparison of PTQ4VM with Ground Truth (true), Full-Precision (FP) RVM, MSE, [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Convergence comparison of Alpha error for Block-wise Initial Quantization (BIQ) versus [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Histograms of learned affine calibration parameters [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Box plots of deviations for learned affine calibration parameters [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Per-frame average Alpha error comparison for W4A4 quantization with (red curve) and [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 26 canonical work pages

  1. [1]

    Robust high-resolution video matting with temporal guidance,

    S. Lin, L. Yang, I. Saleemi, and S. Sengupta, “Robust high-resolution video matting with temporal guidance,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022, pp. 238–247

    work page 2022

  2. [2]

    Designing effective inter-pixel information flow for natural image matting,

    Y . Aksoy, T. Ozan Aydin, and M. Pollefeys, “Designing effective inter-pixel information flow for natural image matting,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 29–37

    work page 2017

  3. [3]

    A geodesic framework for fast interactive image and video segmentation and matting,

    X. Bai and G. Sapiro, “A geodesic framework for fast interactive image and video segmentation and matting,” in2007 IEEE 11th International Conference on Computer Vision. IEEE, 2007, pp. 1–8

    work page 2007

  4. [4]

    Knn matting,

    Q. Chen, D. Li, and C.-K. Tang, “Knn matting,”IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 9, pp. 2175–2188, 2013

    work page 2013

  5. [5]

    A bayesian approach to digital matting,

    Y .-Y . Chuang, B. Curless, D. H. Salesin, and R. Szeliski, “A bayesian approach to digital matting,” inProceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, vol. 2. IEEE, 2001, pp. II–II

    work page 2001

  6. [6]

    A cluster sampling method for image matting via sparse coding,

    X. Feng, X. Liang, and Z. Zhang, “A cluster sampling method for image matting via sparse coding,” inComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Nether- lands, October 11-14, 2016, Proceedings, Part II 14. Springer, 2016, pp. 204–219

    work page 2016

  7. [7]

    Vmformer: End-to-end video matting with transformer,

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

    work page 2024

  8. [8]

    Deep video matting via spatio-temporal alignment and aggregation,

    Y . Sun, G. Wang, Q. Gu, C.-K. Tang, and Y .-W. Tai, “Deep video matting via spatio-temporal alignment and aggregation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 6975–6984

    work page 2021

Show all 35 references
  1. [9]

    Attention- guided temporally coherent video object matting,

    Y . Zhang, C. Wang, M. Cui, P. Ren, X. Xie, X.-S. Hua, H. Bao, Q. Huang, and W. Xu, “Attention- guided temporally coherent video object matting,” inProceedings of the 29th ACM International Conference on Multimedia, 2021, pp. 5128–5137

    2021

  2. [10]

    Real-time high-resolution background matting,

    S. Lin, A. Ryabtsev, S. Sengupta, B. L. Curless, S. M. Seitz, and I. Kemelmacher-Shlizerman, “Real-time high-resolution background matting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 8762–8771

    2021

  3. [11]

    Background matting: The world is your green screen,

    S. Sengupta, V . Jayaram, B. Curless, S. M. Seitz, and I. Kemelmacher-Shlizerman, “Background matting: The world is your green screen,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 2291–2300

    2020

  4. [12]

    Spherical space feature decomposition for guided depth map super-resolution,

    Z. Zhao, J. Zhang, X. Gu, C. Tan, S. Xu, Y . Zhang, R. Timofte, and L. Van Gool, “Spherical space feature decomposition for guided depth map super-resolution,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 12 547–12 558

    2023

  5. [13]

    Discrete cosine transform network for guided depth map super-resolution,

    Z. Zhao, J. Zhang, S. Xu, Z. Lin, and H. Pfister, “Discrete cosine transform network for guided depth map super-resolution,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5697–5707

    2022

  6. [14]

    Deep convolutional sparse coding networks for interpretable image fusion,

    Z. Zhao, J. Zhang, H. Bai, Y . Wang, Y . Cui, L. Deng, K. Sun, C. Zhang, J. Liu, and S. Xu, “Deep convolutional sparse coding networks for interpretable image fusion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 2369–2377

    2023

  7. [15]

    Efficient and model-based infrared and visible image fusion via algorithm unrolling,

    Z. Zhao, S. Xu, J. Zhang, C. Liang, C. Zhang, and J. Liu, “Efficient and model-based infrared and visible image fusion via algorithm unrolling,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 3, pp. 1186–1196, 2021

    2021

  8. [16]

    Ddfm: denoising diffusion model for multi-modality image fusion,

    Z. Zhao, H. Bai, Y . Zhu, J. Zhang, S. Xu, Y . Zhang, K. Zhang, D. Meng, R. Timofte, and L. Van Gool, “Ddfm: denoising diffusion model for multi-modality image fusion,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 8082–8093. 10

    2023

  9. [17]

    Vitmatte: Boosting image matting with pre-trained plain vision transformers,

    J. Yao, X. Wang, S. Yang, and B. Wang, “Vitmatte: Boosting image matting with pre-trained plain vision transformers,”Information Fusion, vol. 103, p. 102091, 2024

    2024

  10. [18]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference,

    B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam, and D. Kalenichenko, “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 2704–2713

    2018

  11. [19]

    A white paper on neural network quantization,

    M. Nagel, M. Fournarakis, R. A. Amjad, Y . Bondarenko, M. Van Baalen, and T. Blankevoort, “A white paper on neural network quantization,”arXiv preprint arXiv:2106.08295, 2021

    2021 arXiv

  12. [20]

    A survey of quantiza- tion methods for efficient neural network inference,

    A. Gholami, S. Kim, Z. Dong, Z. Yao, M. W. Mahoney, and K. Keutzer, “A survey of quantiza- tion methods for efficient neural network inference,” inLow-power computer vision. Chapman and Hall/CRC, 2022, pp. 291–326

    2022

  13. [21]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” inInternational conference on machine learning. pmlr, 2015, pp. 448–456

    2015

  14. [22]

    Raft: Recurrent all-pairs field transforms for optical flow,

    Z. Teed and J. Deng, “Raft: Recurrent all-pairs field transforms for optical flow,” inCom- puter Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16. Springer, 2020, pp. 402–419

    2020

  15. [23]

    Determining optical flow,

    B. K. Horn and B. G. Schunck, “Determining optical flow,”Artificial intelligence, vol. 17, no. 1-3, pp. 185–203, 1981

    1981

  16. [24]

    Blue screen matting,

    A. R. Smith and J. F. Blinn, “Blue screen matting,” inProceedings of the 23rd annual conference on Computer graphics and interactive techniques, 1996, pp. 259–268

    1996

  17. [25]

    Video matting of complex scenes,

    Y .-Y . Chuang, A. Agarwala, B. Curless, D. H. Salesin, and R. Szeliski, “Video matting of complex scenes,” inProceedings of the 29th annual conference on Computer graphics and interactive techniques, 2002, pp. 243–248

    2002

  18. [26]

    Up or down? adaptive rounding for post-training quantization,

    M. Nagel, R. A. Amjad, M. Van Baalen, C. Louizos, and T. Blankevoort, “Up or down? adaptive rounding for post-training quantization,” inInternational conference on machine learning. PMLR, 2020, pp. 7197–7206

    2020

  19. [27]

    Brecq: Pushing the limit of post-training quantization by block reconstruction,

    Y . Li, R. Gong, X. Tan, Y . Yang, P. Hu, Q. Zhang, F. Yu, W. Wang, and S. Gu, “Brecq: Pushing the limit of post-training quantization by block reconstruction,”arXiv preprint arXiv:2102.05426, 2021

    2021 arXiv

  20. [28]

    Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization,

    X. Wei, R. Gong, Y . Li, X. Liu, and F. Yu, “Qdrop: Randomly dropping quantization for extremely low-bit post-training quantization,”arXiv preprint arXiv:2203.05740, 2022

    2022 arXiv

  21. [29]

    An iterative image registration technique with an application to stereo vision,

    B. D. Lucas and T. Kanade, “An iterative image registration technique with an application to stereo vision,” inIJCAI’81: 7th international joint conference on Artificial intelligence, vol. 2, 1981, pp. 674–679

    1981

  22. [30]

    Flownet: Learning optical flow with convolutional networks,

    A. Dosovitskiy, P. Fischer, E. Ilg, P. Hausser, C. Hazirbas, V . Golkov, P. Van Der Smagt, D. Cremers, and T. Brox, “Flownet: Learning optical flow with convolutional networks,” in Proceedings of the IEEE international conference on computer vision, 2015, pp. 2758–2766

    2015

  23. [31]

    Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume,

    D. Sun, X. Yang, M.-Y . Liu, and J. Kautz, “Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 8934–8943

    2018

  24. [32]

    Data-free quantization through weight equalization and bias correction,

    M. Nagel, M. v. Baalen, T. Blankevoort, and M. Welling, “Data-free quantization through weight equalization and bias correction,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 1325–1334

    2019

  25. [33]

    Quantizing deep convolutional networks for efficient inference: A whitepa- per,

    R. Krishnamoorthi, “Quantizing deep convolutional networks for efficient inference: A whitepa- per,”arXiv preprint arXiv:1806.08342, 2018. 11

    2018 arXiv

  26. [34]

    Attention-guided hierarchical structure aggregation for image matting,

    Y . Qiao, Y . Liu, X. Yang, D. Zhou, M. Xu, Q. Zhang, and X. Wei, “Attention-guided hierarchical structure aggregation for image matting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 13 676–13 685

    2020

  27. [35]

    Bimatting: Efficient video matting via binarization,

    H. Qin, L. Ke, X. Ma, M. Danelljan, Y .-W. Tai, C.-K. Tang, X. Liu, and F. Yu, “Bimatting: Efficient video matting via binarization,”Advances in Neural Information Processing Systems, vol. 36, pp. 43 307–43 321, 2023. A Technical Appendices and Supplementary Material A.1 Analy...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.