Pith. sign in

REVIEW 4 major objections 5 minor 48 references

PS4PRO: Pixel-to-pixel Supervision for Photorealistic Rendering and Optimization

T0 review · 4 major / 5 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 The paper claims that adding video-frame-interpolation frames to the training set improves neural rendering reconstruction, because interpolation models carry an implicit world prior.

desk verdict VFI-as-augmentation for neural rendering is a real and cheap idea with positive numbers, but the paper hides the pose-assignment protocol for interpolated frames, so it needs a major revision rather than outright rejection. read the letter →

arxiv 2505.22616 v1 pith:ROSZEBF3 submitted 2025-05-28 cs.CV eess.IV

Yezhi Shen , Qiuchen Zhai , Fengqing Zhu This is my paper
classification cs.CVeess.IV
keywords neuralrenderingvideoframeinterpolationdataaugmentationradiancefields3DGaussiansplattingpixel-to-pixelsupervisionimplicitworldpriorautonomousdrivingscenes
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

Neural rendering reconstructs 3D scenes from 2D images, but quality suffers when input views are sparse, especially in dynamic scenes where some angles are never observed. This paper proposes video frame interpolation as a data augmentation tool: a model synthesizes intermediate frames between neighboring views, and those frames are added to the neural renderer's training set. To make this work, the authors build PS4PRO, a lightweight flow-based interpolation model trained on diverse video data, and argue that it acts as an implicit world prior that encodes camera motion and real 3D geometry. They report that augmenting Lightning-NeRF on Argoverse2 and KITTI, and NeuRAD on NuScenes-mini, with PS4PRO-generated frames improves reconstruction metrics such as PSNR, SSIM, and LPIPS over both the unaugmented baselines and augmentation with existing interpolation models. The claim matters because dense supervision for 3D reconstruction could then be obtained from cheap preprocessing rather than additional capture hardware.

What carries the argument

The load-bearing mechanism is PS4PRO, a lightweight flow-based video frame interpolation model used as a preprocessing module. Given two frames $I_0$ and $I_1$ and a target timestep $t$, it estimates backward motion fields $F_{t\to 0}$ and $F_{t\to 1}$ plus a blending mask $M$, warps both input frames, and fuses them into the intermediate frame $I_t$. Its architecture uses three feature-pyramid levels, a low-resolution base block that receives the timestep, and two refinement blocks that do not, based on the argument that the coarse flow is already accurate and the refinement only needs local detail. The model is trained on Vimeo90K with a teacher optical-flow model providing backward-flow distillation during early epochs. Within the neural rendering pipeline, the generated frames are treated as additional ground-truth views whose photometric loss is minimized, which is how the 'pixel-to-pixel supervision' enters the reconstruction.

What would settle it

Take a driving video with recorded camera poses, train the same renderer once with real captured intermediate frames plus their true poses and once with PS4PRO-generated intermediate frames plus whatever pose rule the paper intends, and compare reconstruction quality; if the real frames clearly outperform the generated ones, the improvement is not evidence of an implicit world prior, and if small pose perturbations erase the gain, the unstated pose assignment is what actually carries the result.

Watch

Extended reading notes

Core claim

The central claim, stated as the authors would state it, is that a video frame interpolation model trained on large and diverse natural video data implicitly learns a world prior—knowledge of how cameras move and how real 3D geometry projects onto pixels—and that this prior can be transferred to neural rendering. The transfer happens by using PS4PRO to synthesize intermediate frames between existing training views, then feeding those synthetic frames into the renderer's photometric loss as if they were real ground-truth views. The generated frames provide what the paper calls pixel-to-pixel supervision: they pin down correspondences between rays, reducing the ambiguity that arises when sparse rays intersect. The paper reports consistent improvements on static scenes reconstructed by Lightning-NeRF over Argoverse2 and KITTI, and on the dynamic scenes of NuScenes-mini reconstructed by NeuRAD, at an augmentation cost measured in fractions of a permille of total training time.

Load-bearing premise

The method assumes that an interpolated frame can be inserted into the neural renderer's training set as a valid ground-truth view, which requires assigning each interpolated frame a camera pose, yet the paper never states what pose is assigned to the generated frames.

Editorial extensions

If this is right

  • Any neural rendering method trained from video frames can be augmented with interpolated frames essentially for free, turning sparse viewpoint coverage into denser supervision without new capture hardware.
  • Because the augmentation is a preprocessing step that is independent of the renderer, it should compose with other enhancement techniques such as monocular depth priors or diffusion-based regularizers.
  • The reported gains on dynamic scenes suggest the world prior captures object motion as well as camera motion, pointing toward deformable-scene reconstruction as a natural continuation.
  • The step-control design update—passing the timestep only to the base block—could carry over to other flow-based interpolation networks that need arbitrary-time synthesis.

Reading between the lines

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

  • A direct test the paper leaves open is whether PS4PRO's intermediate frames are geometrically consistent with the interpolated camera trajectory; comparing VFI output against genuinely captured frames at the same timestamps on a dataset with dense video would settle it.
  • The paper claims generality over both NeRF and 3D Gaussian Splatting, but its experiments only cover NeRF-style methods (Lightning-NeRF and NeuRAD); testing 3DGS directly would check the breadth of the claim.
  • If the gains are robust to pose perturbation, the mechanism might be less about world knowledge and more about providing any extra collinear image to stabilize optimization; ablating the pose rule would distinguish these readings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes PS4PRO, a flow-based video frame interpolation (VFI) model, and uses it to augment neural rendering training sets by generating intermediate frames between existing views. The method is evaluated on standard VFI benchmarks (Vimeo90K, UCF101, Middlebury, SNU-FILM, HD, X4K1000FPS) and as a data-augmentation module for Lightning-NeRF on Argoverse2 and KITTI and for NeuRAD on NuScenes-mini. The authors report consistent but small improvements in PSNR/SSIM/LPIPS over baselines and attribute them to an implicit world prior learned by the VFI model.

Significance. If substantiated, the idea is practically attractive: a lightweight, training-free augmentation that improves sparse-view neural rendering at negligible computational cost. The VFI model itself is described in detail and benchmarked against strong baselines, and the downstream neural-rendering evaluation is a genuine external test rather than a circular validation. However, the current evidence is not sufficient to support the central claim because the pose-assignment protocol for inserted frames is missing, no variance or ablation controls are provided, and the proposed 'pixel-to-pixel supervision' is not actually enforced as a loss in the reconstruction optimization.

major comments (4)
  1. [4.5, Tables 3 and 4] The manuscript never specifies how the interpolated frames are associated with camera poses when inserted into the L-NeRF and NeuRAD training sets. Neural rendering requires a camera pose for every supervision image, yet the paper only states that 'the training set is interpolated with one extra frame in between each frame' and that '5x interpolation' is used for NuScenes. For L-NeRF, one needs to know whether poses are linearly interpolated in SE(3) or in some other parameterization. For NeuRAD on NuScenes, where moving agents are present and the text itself notes that roll/pitch information is discarded, assigning a single valid pose to a synthetic intermediate frame is not geometrically well defined. Without this protocol, the experiments cannot be reproduced and the reported gains cannot be attributed to the described method.
  2. [4.5, Tables 3 and 4] No variance or number of runs is reported for the neural-rendering results, and the improvements over the strongest non-PS4PRO baselines are small: +0.15 PSNR over EMA-s on Argoverse2, +0.07 PSNR over M2M on KITTI, and +0.41 PSNR over M2M on NuScenes-mini. Since Section 4.4 indicates that multiple runs were used for timing, the same practice should be applied to PSNR/SSIM/LPIPS, with standard deviations or confidence intervals, so that the reader can judge whether the differences are meaningful.
  3. [3.2 and 4.5] The paper claims to impose 'pixel-to-pixel supervision' on neural rendering, but no new loss term or explicit correspondence constraint is added to the reconstruction objective: Eq. (1) remains the standard photometric loss, and the interpolated images are simply added as extra training targets. As written, the mechanism in Eq. (3) is not enforced during optimization. The authors should clarify how this differs from generic data augmentation and should include an ablation that replaces VFI frames with a non-learned baseline (e.g., linearly blended frames or duplicated existing frames) to isolate the contribution of the learned VFI model.
  4. [3.3 and 4.5] The claim that PS4PRO captures 'an implicit world prior consisting of camera motion and 3D geometry' and that it 'comprehends 3D geometry better' than other VFI models is not directly measured. The neural-rendering improvements could simply reflect the higher perceptual quality of the synthetic images, as measured on VFI benchmarks, rather than a distinct geometric prior. A test such as geometry/depth consistency on held-out views, or an experiment controlling for interpolation image quality, is needed to support the world-prior interpretation.
minor comments (5)
  1. [Table 1] The caption states that 'our model achieves the best performance in center frame interpolation tasks,' but PS4PRO is not the best on UCF101 or on the Hard/Extreme subsets of SNU-FILM; please qualify this statement.
  2. [4.2, Eqs. (6)-(8)] The norm expressions use comma notation (e.g., ||Ît, It||1); this should be written as a difference, e.g., ||Ît - It||1.
  3. [4.4, Table 2] The text should explicitly state that the reported proportions are in permille (‰), not percent, to avoid confusion with the header 'Proportion of Time.'
  4. [4.5] The dataset splits, the number of scenes, and the exact definition of 'keyframes' for NuScenes-mini are not specified; please provide this information for reproducibility.
  5. [Figure 5] The caption refers to 'sequences 1 to 5' without identifying the KITTI sequences; please list the sequence names or provide a supplementary description.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PS4PRO is trained on an external video dataset and the neural-rendering augmentation is evaluated on held-out driving datasets with fixed VFI weights.

full rationale

The paper's central claim is that intermediate frames generated by a video frame interpolation model can be used as data augmentation to improve neural rendering reconstruction (Section 4.5, Tables 3-4). The VFI model PS4PRO is trained on Vimeo90K (Section 4.2) and then applied without fine-tuning to KITTI, Argoverse2, and NuScenes-mini scenes, so the downstream comparison is not fitted to the evaluated reconstruction metrics. The photometric equations in Section 3 are standard rendering and pixel-correspondence definitions, and they do not force the reported gains by construction. The phrase 'implicit world prior' is an interpretive overclaim rather than a mathematical premise of the derivation: no step assumes that the VFI model already knows the target scene. The only self-citation, reference [43] used for the 'accuracy to precision' block design, is not load-bearing for the central augmentation claim. The Vimeo90K column in Table 1 overlaps with training data, which is a benchmarking weakness, but the neural-rendering conclusions rest on unseen driving datasets. No circular step meeting the quoting-and-reduction standard was found.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the domain assumptions that VFI-generated frames can serve as ground truth and that they can be placed into the neural rendering pipeline with valid poses. The latter assumption is unstated. The only free parameter directly tied to the central claim is the interpolation factor, which is not ablated.

free parameters (1)
  • Interpolation factor (number of inserted frames per input pair) = 1x for Argoverse2/KITTI, 5x for NuScenes-mini
    Chosen per dataset without reported ablation; directly controls the amount of pseudo-supervision and may affect the magnitude of improvement.
assumptions (4)
  • domain assumption Interpolated frames are treated as ground-truth supervision for neural rendering
    Section 4.5 adds interpolated frames to the training set without accounting for interpolation errors; the method assumes these pseudo-images are reliable photometric supervision.
  • domain assumption Interpolated frames can be assigned valid camera poses (implied, not stated)
    Neural rendering requires per-image poses; the paper never describes how poses are obtained for synthesized frames. This is essential for the experiments in Section 4.5.
  • domain assumption The VFI model trained on Vimeo90K generalizes to driving scenes without fine-tuning
    The paper claims generalizability (Sections 1 and 4.5) but only evaluates on autonomous driving datasets; no cross-domain analysis or testing on general scenes is provided.
  • standard math Photometric consistency I0(p) = I1(p') holds for corresponding pixels
    Equation (3) assumes color constancy along motion trajectories, a standard assumption in optical flow and VFI but not always valid under lighting changes or occlusions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PS4PRO: Pixel-to-pixel Supervision for Photorealistic Rendering and Optimization." pith.science (2026). https://pith.science/paper/ROSZEBF3

@misc{pith2026250522616,
  author       = {Pith},
  title        = {Pith review of: PS4PRO: Pixel-to-pixel Supervision for Photorealistic Rendering and Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ROSZEBF3}},
  note         = {Machine review of arXiv:2505.22616}
}
read the original abstract

Neural rendering methods have gained significant attention for their ability to reconstruct 3D scenes from 2D images. The core idea is to take multiple views as input and optimize the reconstructed scene by minimizing the uncertainty in geometry and appearance across the views. However, the reconstruction quality is limited by the number of input views. This limitation is further pronounced in complex and dynamic scenes, where certain angles of objects are never seen. In this paper, we propose to use video frame interpolation as the data augmentation method for neural rendering. Furthermore, we design a lightweight yet high-quality video frame interpolation model, PS4PRO (Pixel-to-pixel Supervision for Photorealistic Rendering and Optimization). PS4PRO is trained on diverse video datasets, implicitly modeling camera movement as well as real-world 3D geometry. Our model performs as an implicit world prior, enriching the photo supervision for 3D reconstruction. By leveraging the proposed method, we effectively augment existing datasets for neural rendering methods. Our experimental results indicate that our method improves the reconstruction performance on both static and dynamic scenes.

Figures

Figures reproduced from arXiv: 2505.22616 by the authors.

Figure 1
Figure 1. Reconstruction results from NeuRAD with different VFI enhancements on the NuScene-mini dataset. Our proposed method [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the overall framework of the proposed method. Given input frames [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Illustration of reduction of the reconstruction uncertainty [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Architecture of Base and Refinement Blocks in PS4PRO. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The visual comparison of KITTI dataset reconstructed using Lightning-NeRF without (left column) and with (right column) the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 31 canonical work pages

  1. [1]

    A database and evaluation methodology for optical flow

    Simon Baker, Daniel Scharstein, James P Lewis, Stefan Roth, Michael J Black, and Richard Szeliski. A database and evaluation methodology for optical flow. International journal of computer vision, 92:1–31, 2011. 6

    work page 2011

  2. [2]

    Depth-aware video frame interpolation

    Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 3703–3712, 2019. 3

    work page 2019

  3. [3]

    Memc-net: Motion estimation and motion compensation driven neural network for video inter- polation and enhancement

    Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video inter- polation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933–948, 2019. 6

    work page 2019

  4. [4]

    Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 5855–5864,

  5. [5]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 2

    2022

  6. [6]

    nuscenes: A multi- modal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020. 6, 8

    work page 2020

  7. [7]

    Light- ning nerf: Efficient hybrid scene representation for au- tonomous driving

    Junyi Cao, Zhichao Li, Naiyan Wang, and Chao Ma. Light- ning nerf: Efficient hybrid scene representation for au- tonomous driving. arXiv preprint arXiv:2403.05907, 2024. 3, 6, 7

    arXiv 2024

  8. [8]

    Pascal Chang, Jingwei Tang, Markus Gross, and Vinicius C. Azevedo. How i warped your noise: a temporally-correlated noise prior for diffusion models. InThe Twelfth International Conference on Learning Representations, 2024. 2

    work page 2024

Show all 48 references
  1. [9]

    Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo

    Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast general- izable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14124–14133, 2021. 4

    2021

  2. [10]

    A survey on 3d gaussian splatting

    Guikun Chen and Wenguan Wang. A survey on 3d gaussian splatting. arXiv preprint arXiv:2401.03890, 2024. 2

    2024 arXiv

  3. [11]

    Motion-compensated frame interpolation us- ing bilateral motion estimation and adaptive overlapped block motion compensation

    Byeong-Doo Choi, Jong-Woo Han, Chang-Su Kim, and Sung-Jea Ko. Motion-compensated frame interpolation us- ing bilateral motion estimation and adaptive overlapped block motion compensation. IEEE Transactions on Circuits and Systems for Video Technology, 17(4):407–416, 2007. 3

    2007

  4. [12]

    Channel attention is all you need for video frame interpolation

    Myungsub Choi, Heewon Kim, Bohyung Han, Ning Xu, and Kyoung Mu Lee. Channel attention is all you need for video frame interpolation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10663–10671, 2020. 6

    2020

  5. [13]

    Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps

    Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, De- jia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps. arXiv preprint arXiv:2311.17245, 2023. 4

    2023 arXiv

  6. [14]

    Colmap-free 3d gaussian splat- ting

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A Efros, and Xiaolong Wang. Colmap-free 3d gaussian splat- ting. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 20796–20805,

  7. [15]

    Egolifter: Open-world 3d seg- mentation for egocentric perception

    Qiao Gu, Zhaoyang Lv, Duncan Frost, Simon Green, Julian Straub, and Chris Sweeney. Egolifter: Open-world 3d seg- mentation for egocentric perception. In European Confer- ence on Computer Vision , pages 382–400. Springer, 2024. 2

    2024

  8. [16]

    Free-surgs: Sfm-free 3d gaussian splatting for surgical scene reconstruction

    Jiaxin Guo, Jiangliu Wang, Di Kang, Wenzhen Dong, Went- ing Wang, and Yun-hui Liu. Free-surgs: Sfm-free 3d gaussian splatting for surgical scene reconstruction. In In- ternational Conference on Medical Image Computing and Computer-Assisted Intervention , pages 350–360. Springer,

  9. [17]

    Many-to-many splatting for efficient video frame interpola- tion

    Ping Hu, Simon Niklaus, Stan Sclaroff, and Kate Saenko. Many-to-many splatting for efficient video frame interpola- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 3553–3562,

  10. [18]

    Real-time intermediate flow estimation for video frame interpolation

    Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In European Conference on Com- puter Vision, pages 624–642. Springer, 2022. 3, 6

    2022

  11. [19]

    Ocai: Improving optical flow estimation by occlusion and consistency aware interpolation

    Jisoo Jeong, Hong Cai, Risheek Garrepalli, Jamie Menjay Lin, Munawar Hayat, and Fatih Porikli. Ocai: Improving optical flow estimation by occlusion and consistency aware interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages ...

    2024

  12. [20]

    Super slomo: High quality estimation of multiple intermediate frames for video interpolation

    Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages ...

    2018

  13. [21]

    Alignerf: High-fidelity neural radiance fields via alignment- aware training

    Yifan Jiang, Peter Hedman, Ben Mildenhall, Dejia Xu, Jonathan T Barron, Zhangyang Wang, and Tianfan Xue. Alignerf: High-fidelity neural radiance fields via alignment- aware training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 46–55,

  14. [22]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1,

  15. [23]

    Ifrnet: Intermediate feature refine network for efficient frame interpolation

    Lingtong Kong, Boyuan Jiang, Donghao Luo, Wenqing Chu, Xiaoming Huang, Ying Tai, Chengjie Wang, and Jie Yang. Ifrnet: Intermediate feature refine network for efficient frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , p...

    1969

  16. [24]

    Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d

    Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292–3310, 2022. 6

    2022

  17. [25]

    3dgs- enhancer: Enhancing unbounded 3d gaussian splatting with view-consistent 2d diffusion priors

    Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgs- enhancer: Enhancing unbounded 3d gaussian splatting with view-consistent 2d diffusion priors. arXiv preprint arXiv:2410.16266, 2024. 2, 3

    2024 arXiv

  18. [26]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , pages 11976–11986,

  19. [27]

    Decoupled weight decay regularization

    I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6

    2017 arXiv

  20. [28]

    Urban radiance field representa- tion with deformable neural mesh primitives

    Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representa- tion with deformable neural mesh primitives. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 465–476, 2023. 6

    2023

  21. [29]

    Deformable neural radiance fields using rgb and event cameras

    Qi Ma, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Deformable neural radiance fields using rgb and event cameras. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision , pages 3590–3600,

  22. [30]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM , 65(1):99–106, 2021. 1

    2021

  23. [31]

    Bench- marking neural radiance fields for autonomous robots: An overview

    Yuhang Ming, Xingrui Yang, Weihan Wang, Zheng Chen, Jinglun Feng, Yifan Xing, and Guofeng Zhang. Bench- marking neural radiance fields for autonomous robots: An overview. arXiv preprint arXiv:2405.05526, 2024. 2

    2024 arXiv

  24. [32]

    Instant neural graphics primitives with a mul- tiresolution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a mul- tiresolution hash encoding. In ACM transactions on graphics (TOG), pages 1–15, 2022. 2

    2022

  25. [33]

    A comprehensive survey on video frame interpolation techniques

    Anil Singh Parihar, Disha Varshney, Kshitija Pandya, and Ashray Aggarwal. A comprehensive survey on video frame interpolation techniques. The Visual Computer, 38(1):295– 319, 2022. 2, 3

    2022

  26. [34]

    Bmbc: Bilateral motion estimation with bilateral cost vol- ume for video interpolation

    Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost vol- ume for video interpolation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV 16 , pages 109–125. Spri...

    2020

  27. [35]

    Asymmetric bilateral motion estimation for video frame interpolation

    Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14539–14548, 2021. 4, 5

    2021

  28. [36]

    Fpo++: efficient encoding and rendering of dynamic neural radiance fields by analyzing and enhancing fourier plenoctrees

    Saskia Rabich, Patrick Stotko, and Reinhard Klein. Fpo++: efficient encoding and rendering of dynamic neural radiance fields by analyzing and enhancing fourier plenoctrees. The Visual Computer, 40(7):4777–4788, 2024. 2

    2024

  29. [37]

    Xvfi: extreme video frame interpolation

    Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In Proceedings of the IEEE/CVF international conference on computer vision , pages 14489–14498, 2021. 6

    2021

  30. [38]

    Ucf101: A dataset of 101 human actions classes from videos in the wild

    K Soomro. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402 ,

    arXiv

  31. [39]

    Neurad: Neural rendering for autonomous driving

    Adam Tonderski, Carl Lindstr ¨om, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14895–14904, 2024. 2, 3, 7, 8

    2024

  32. [40]

    Argoverse 2: Next generation datasets for self-driving perception and forecasting

    Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, et al. Argoverse 2: Next generation datasets for self-driving perception and forecasting. arXiv preprint arXiv:230...

    2023 arXiv

  33. [41]

    Recent advances in 3d gaussian splatting

    Tong Wu, Yu-Jie Yuan, Ling-Xiao Zhang, Jie Yang, Yan- Pei Cao, Ling-Qi Yan, and Lin Gao. Recent advances in 3d gaussian splatting. Computational Visual Media, 10(4):613– 642, 2024. 2

    2024

  34. [42]

    Diffusionerf: Regularizing neural radiance fields with denoising diffu- sion models

    Jamie Wynn and Daniyar Turmukhambetov. Diffusionerf: Regularizing neural radiance fields with denoising diffu- sion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4180– 4189, 2023. 3

    2023

  35. [43]

    Pose guided portrait viewinterpolation from dual cameras with a long baseline

    Weichen Xu, Yezhi Shen, Qian Lin, Jan Allebach, and Fengqing Zhu. Pose guided portrait viewinterpolation from dual cameras with a long baseline. IEEE 26th international workshop on Multimedia Signal Processing (MMSP) , 2024. 5

    2024

  36. [44]

    Video enhancement with task-oriented flow

    Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision, 127:1106– 1125, 2019. 5

    2019

  37. [45]

    Unisim: A neural closed-loop sensor simulator

    Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Mani- vasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Ur- tasun. Unisim: A neural closed-loop sensor simulator. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition , pages 1389–1399, 2023. 2, 3

    2023

  38. [46]

    pixelnerf: Neural radiance fields from one or few images

    Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4578–4587, 2021. 2

    2021

  39. [47]

    Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis.arXiv preprint arXiv:2409.02048, 2024

    Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis.arXiv preprint arXiv:2409.02048, 2024. 2, 3

    2024 arXiv

  40. [48]

    Extracting motion and ap- pearance via inter-frame attention for efficient video frame interpolation

    Guozhen Zhang, Yuhan Zhu, Haonan Wang, Youxin Chen, Gangshan Wu, and Limin Wang. Extracting motion and ap- pearance via inter-frame attention for efficient video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 5...

    2023

Pith tools

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