Pith. sign in

REVIEW 3 major objections 4 minor 59 references

TimeTracker: Event-based Continuous Point Tracking for Video Frame Interpolation with Non-linear Motion

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Video frame interpolation under fast, non-linear motion can be handled by tracking local regions through event streams instead of estimating global optical flow.

desk verdict TimeTracker is a genuinely first integration of point tracking into event-based VFI, with consistent gains on fast motion, but its superpixel coherence assumption limits the SOTA claim and needs to be stated more carefully. read the letter →

arxiv 2505.03116 v1 pith:4VZEV5BK submitted 2025-05-06 cs.CV

Haoyue Liu , Jinghan Xu , Yi Chang , Hanyu Zhou , Haozhi Zhao , Lin Wang , Luxin Yan This is my paper · ORCID
classification cs.CV
keywords eventcamerasvideoframeinterpolationpointtrackingopticalflownon-linearmotiondenseany-timesuperpixelsegmentationevent-basedvision
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 tries to establish that event-based video frame interpolation under fast, non-linear motion is better handled by tracking local regions continuously through time than by estimating global optical flow from sparse events or from synthesized intermediate frames. It proposes TimeTracker, which segments each frame into appearance-similar patches, tracks one query point per patch through the event stream, and converts those trajectories into dense optical flow at any requested timestamp. The authors report consistent gains over existing frame- and event-based interpolation methods in both motion estimation and frame quality, and they contribute a real-world dataset featuring fast, non-linear motion.

What carries the argument

The load-bearing object is the per-superpixel continuous trajectory. A superpixel is a small image region whose pixels share similar color and spatial position; one query point per superpixel is initialized from a keypoint detector and tracked through voxelized events by local correlation in a sliding time window. The trajectories are converted into a coarse dense any-time flow, then refined by a global optimization module using a self-supervised consistency loss between interpolation results and re-estimated flow.

What would settle it

Run the method on a high-speed event-image clip of splashing water or smoke with ground-truth frames; if interpolation quality in the fluid regions is no better than plain fusion and visibly worse than in rigid regions, the coherent-motion assumption is the failing part.

Watch

Extended reading notes

Core claim

The paper's central claim is that any-time dense optical flow for event-based video frame interpolation is better obtained by turning flow estimation into local point tracking than by estimating global flow from events or from synthesized frames. TimeTracker segments the boundary frames into appearance-similar superpixels, initializes one query point per superpixel, tracks those points through the event stream over continuous time using local correlation in a sliding temporal window, and then globally optimizes the resulting coarse flow against a frame-interpolation consistency constraint. The authors report that this yields dense flow at any timestamp and outperforms existing frame- and event-based interpolation methods on synthetic and real benchmarks, including their new dataset of fast, non-linear motion.

Load-bearing premise

The load-bearing premise is that pixels within each appearance-similar superpixel move coherently, so that one tracked point can stand for the whole patch; the authors note this breaks down for fluids and other dynamic textures.

Editorial extensions

If this is right

  • Interpolated frames can be generated at any requested timestamp, not just fixed ratios, because the tracked trajectories define dense flow continuously in time.
  • Multi-frame interpolation remains stable when many frames are skipped, since local tracking degrades more gracefully than global flow estimation as displacement grows.
  • Image appearance, not events alone, defines where tracking templates live, so regions with sparse or absent events are still covered by the dense flow.
  • Errors from synthesized intermediate frames no longer feed back into flow estimation, breaking the chicken-and-egg loop that affects synthesis-based any-time flow methods.

Reading between the lines

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

  • By extension, making the region segmentation adaptive—splitting superpixels whose tracked trajectories diverge—could extend the method to non-rigid objects and dynamic textures without abandoning point tracking.
  • By extension, the same continuous-trajectory representation could support estimating acceleration or higher-order motion from event timestamps, which would matter for deblurring and high-frame-rate reconstruction beyond interpolation.
  • By extension, fluid scenes, which the paper names as its weak spot, could be handled by plugging in event-based fluid velocimetry as a specialized tracker for those regions.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes TimeTracker, an event-based video frame interpolation framework that replaces direct dense optical flow estimation with continuous point tracking. The method segments the boundary frames into SLIC superpixels, initializes a query point per superpixel with SIFT, tracks those points through a voxelized event stream using local-attention point tracking, converts the tracks into a coarse dense any-time optical flow, refines that flow with a RAFT-style global optimization module under a self-supervised consistency loss, and finally warps and refines the intermediate frame. The authors evaluate on synthetic datasets (GoPro, SNU-FILM) and real datasets (BS-ERGB and a newly collected fast-motion dataset, named CHMD/CHRD inconsistently), reporting better PSNR and SSIM than several frame-based and event-based baselines. They also provide ablations showing the importance of the superpixel module, point tracking module, global optimization, and frame refinement, and they explicitly acknowledge that the tracking assumption fails for dynamic textures and fluids.

Significance. If the results hold, the paper makes a useful contribution by reformulating event-based any-time dense optical flow for VFI as a local point-tracking problem, which is a conceptually clean way to exploit the temporal continuity of events while using image appearance to overcome spatial sparsity. The new real-world dataset with fast, non-linear motion is a valuable resource, and the ablation study is informative. The main limitations are that the central claim of state-of-the-art performance is currently demonstrated only on rigid or near-rigid scenes, that several headline gains over the strongest baseline are small and are reported without error bars or significance tests, and that the self-supervised flow-consistency loss in Eq. (5) may not provide an independent correction signal in exactly the regions where the method is weakest.

major comments (3)
  1. [§3.2, §5 (Limitation)] The central claim of state-of-the-art performance under non-linear motion rests on the assumption that appearance-similar, spatially compact superpixels move coherently. The paper's own Limitation paragraph concedes that for dynamic textures and fluids the tracked features are temporally discontinuous, and that TimeTracker falls back to direct synthesis, which it calls suboptimal. Since no quantitative evaluation on non-rigid or dynamic-texture scenes is reported anywhere in the experiments, the SOTA claim is currently established only for rigid or near-rigid scenes. Please either add an evaluation on a non-rigid or dynamic-texture event dataset, provide a quantitative measure of intra-superpixel optical-flow variance to justify the coherence assumption, or explicitly restrict the claim in the abstract and conclusion to rigid or near-rigid motion.
  2. [§3.3, Eq. (5)] The self-supervised consistency loss in Eq. (5) compares the coarse tracking-based flow with the output of a frame-based flow network applied to the interpolated frames. Those interpolated frames are themselves generated by the same pipeline, so in regions where superpixel tracking fails, both sides of the loss are computed from the same degraded estimates and there is no independent correction signal; the tracking-based flow and the frame-based flow can converge to a mutually consistent but incorrect motion. The manuscript does not report an ablation without Lflow, so the contribution of this term is not quantified. Please report such an ablation, and ideally validate the refined flow against ground-truth optical flow on an event flow benchmark rather than only through the final interpolated-frame metrics.
  3. [§4.2, Tables 1 and 2] The SOTA claim is supported by differences that are small relative to typical run-to-run variation: on GoPro 7-skip, TimeTracker exceeds TimeLens-XL by 0.11 dB PSNR (37.13 vs. 37.02) and 0.003 SSIM, and on BS-ERGB 1-skip by 0.50 dB PSNR. No error bars, multiple seeds, or significance tests are reported anywhere in the paper. Please add standard deviations over at least three training runs or a paired per-sequence test for the headline comparisons, or temper the "outperforms prior arts" claim to the datasets and settings where the gap is clearly outside noise.
minor comments (4)
  1. [§4.1, Table 2, Fig. 6, Fig. 8] The newly collected dataset is named CHMD in the introduction, Fig. 6, and Fig. 8, but CHRD in §4.1 and Table 2; Table 2 also uses the placeholder column header "Ours". Please unify the dataset name and describe its capture setup, size, and train/test split.
  2. [Table 4] Table 4 labels the proposed method "TimeTraker"; the correct spelling is TimeTracker. Also be consistent with the method name PerVFI versus PerFVI across Tables 1 and 4.
  3. [§3.3, Eq. (5)] Eq. (5) uses the notation F^coa, I_t, O_t without specifying how these quantities are sampled from the any-time coarse flow and occlusion fields at a particular training timestamp, or how occluded pixels are excluded from the loss; please clarify the indexing and the masking procedure.
  4. [§3.3, Sec. 3.2] The text says SIFT points are used to initialize a query point for each superpixel segment, but it does not specify the selection rule when multiple SIFT points fall in one superpixel, when no SIFT point is found, or when no event occurs at the chosen location; please provide this detail, since the query-point initialization directly affects the whole tracking pipeline.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: tracking and interpolation are supervised by external ground truth, and the self-supervised flow consistency loss is anchored by reconstruction loss.

full rationale

The paper's derivation chain is self-contained against external benchmarks rather than circular. The point-tracking model is trained with ground-truth point trajectories from Tap-Vid and MultiFlow (Section 3.5), and the VFI model is trained with ground-truth frames from GoPro via the reconstruction loss Lrec (Eq. 10). The self-supervised consistency term Lflow (Eq. 5) compares the coarse tracking flow with optical flow computed from the interpolated frames, but this is a mutual training signal, not a fitted prediction, and it is anchored by Lrec and by the supervised tracking loss. The global optimization module is an architectural component (RAFT-style) trained with these losses, not a parameter fitted to the evaluation data. The appearance-coherence assumption underlying SLIC superpixels (Section 3.2) is a heuristic, and the paper's own Limitation paragraph concedes it fails for dynamic textures such as fluids, but this is a scope limitation rather than a circular reduction: the claimed SOTA results are demonstrated on the datasets tested, and no equation reduces to its own input by construction. The only author self-citation, [28] in the related-work list, is not load-bearing to the central derivation. No fitted input is renamed as a prediction, and no uniqueness theorem or ansatz is smuggled in via self-citation. Therefore the circularity score is 0.

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

The framework rests on standard event-camera models and on several domain assumptions about motion coherence, point-to-flow densification, and self-supervised consistency. There are no invented physical entities. The main free parameters are training hyperparameters and the voxel bin size, which appears to be dataset-dependent.

free parameters (4)
  • Loss weights lambda_1, lambda_2, lambda_3 = 1, 1, 0.8
    Set by hand in Section 3.5 (Eq. 11-12). They weight tracking, reconstruction, and flow consistency losses; no search or principle is given.
  • Voxel bin size = Not explicitly stated; Fig. 8 shows dataset-dependent optimum (about 0.005 s for BS-ERGB, smaller for CHMD)
    Section 4.3 analyzes reconstruction quality versus bin size and implies the optimal value differs by dataset. If the bin size was chosen after seeing test performance, this is a fit-to-eval-set issue.
  • Sliding window length L and iterations M, N = L=10, M=5, N=10
    Implementation hyperparameters given in Section 3.5; the paper does not report sensitivity or selection procedure.
  • SLIC superpixel parameters = Not stated
    Superpixel compactness and number of segments are not reported; these directly control template size and tracking behavior, yet no values are provided.
assumptions (5)
  • standard math Event generation model: DeltaL = logI(...) = pC and constant-illumination approximation (Eq. 2)
    Standard model from event camera literature [15]; used to justify that events trace motion edges and to build the motion mask.
  • domain assumption Local appearance similarity implies local motion consistency (SLIC superpixels)
    Load-bearing assumption of the SARS module in Section 3.2. The paper provides qualitative support in Fig. 4, not a quantitative guarantee, and it fails for dynamic textures (acknowledged in Limitations).
  • domain assumption Tracking a sparse set of query points can be densified into an accurate dense any-time flow
    CTME tracks one query point per superpixel and then densifies the result to a dense flow via global optimization. There is no formal argument that the number of points or the interpolation is sufficient for accurate dense flow.
  • domain assumption RAFT's frame-based flow is a reliable teacher for the self-supervised consistency loss (Eq. 5)
    The global optimization assumes that RAFT flow computed from interpolated frames provides a valid training signal. If RAFT fails in the same large-motion or non-linear regimes, the loss can reinforce rather than correct errors.
  • domain assumption ESIM-simulated events on GoPro and Tap-Vid transfer to real event camera data
    The tracking and VFI models are trained on simulated events (Section 3.5), while evaluation includes real datasets (BS-ERGB, CHMD). No domain-adaptation mechanism is described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TimeTracker: Event-based Continuous Point Tracking for Video Frame Interpolation with Non-linear Motion." pith.science (2026). https://pith.science/paper/4VZEV5BK

@misc{pith2026250503116,
  author       = {Pith},
  title        = {Pith review of: TimeTracker: Event-based Continuous Point Tracking for Video Frame Interpolation with Non-linear Motion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4VZEV5BK}},
  note         = {Machine review of arXiv:2505.03116}
}
read the original abstract

Video frame interpolation (VFI) that leverages the bio-inspired event cameras as guidance has recently shown better performance and memory efficiency than the frame-based methods, thanks to the event cameras' advantages, such as high temporal resolution. A hurdle for event-based VFI is how to effectively deal with non-linear motion, caused by the dynamic changes in motion direction and speed within the scene. Existing methods either use events to estimate sparse optical flow or fuse events with image features to estimate dense optical flow. Unfortunately, motion errors often degrade the VFI quality as the continuous motion cues from events do not align with the dense spatial information of images in the temporal dimension. In this paper, we find that object motion is continuous in space, tracking local regions over continuous time enables more accurate identification of spatiotemporal feature correlations. In light of this, we propose a novel continuous point tracking-based VFI framework, named TimeTracker. Specifically, we first design a Scene-Aware Region Segmentation (SARS) module to divide the scene into similar patches. Then, a Continuous Trajectory guided Motion Estimation (CTME) module is proposed to track the continuous motion trajectory of each patch through events. Finally, intermediate frames at any given time are generated through global motion optimization and frame refinement. Moreover, we collect a real-world dataset that features fast non-linear motion. Extensive experiments show that our method outperforms prior arts in both motion estimation and frame interpolation quality.

Figures

Figures reproduced from arXiv: 2505.03116 by the authors.

Figure 1
Figure 1. Visual comparison of our method with other SOTA meth [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of (a) four flow-based VFI paradigms and (b) their comparison results. Image-based methods like SuperSlomo [ [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The overall architecture of the TimeTracker includes a Scene-Aware Region Segmentation (SARS) module, a Continuous [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Two continuous motion estimation paradigms. (a) Tra [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Visual comparison of the proposed method and other SOTA methods across different datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Ablation studies of TimeTracker. The first row and the [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 52 canonical work pages

  1. [1]

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

    Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. InIEEE Conf. Comput. Vis. Pattern Recog., pages 8934–8943, 2018. 1, 3

    work page 2018

  2. [2]

    Raft: Recurrent all-pairs field transforms for optical flow

    Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. InEur. Conf. Comput. Vis., pages 402–419, 2020. 1, 3, 5, 6

    work page 2020

  3. [3]

    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. InIEEE Conf. Comput. Vis. Pattern Recog., pages 9000–9008, 2018. 1, 2, 3, 6, 7

    work page 2018

  4. [4]

    Depth-aware video frame interpolation

    Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiy- ong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. InIEEE Conf. Comput. Vis. Pattern Recog., pages 3703–3712, 2019

    work page 2019

  5. [5]

    Softmax splatting for video frame interpolation

    Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation. InIEEE Conf. Comput. Vis. Pattern Recog., pages 5437–5446, 2020

    work page 2020

  6. [6]

    Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation

    Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. InEur. Conf. Comput. Vis., pages 109–125, 2020

    work page 2020

  7. [7]

    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. InEur. Conf. Comput. Vis., pages 624–642, 2022. 1, 3

    work page 2022

  8. [8]

    Quadratic video interpolation.Adv

    Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming- Hsuan Yang. Quadratic video interpolation.Adv. Neural Inform. Process. Syst., 32, 2019. 1, 3

    work page 2019

Show all 59 references
  1. [9]

    All at once: Temporally adaptive multi-frame interpolation with advanced motion modeling

    Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at once: Temporally adaptive multi-frame interpolation with advanced motion modeling. InEur. Conf. Comput. Vis., pages 107–123, 2020

    2020

  2. [10]

    Video frame interpolation without temporal priors.Adv

    Youjian Zhang, Chaoyue Wang, and Dacheng Tao. Video frame interpolation without temporal priors.Adv. Neural Inform. Process. Syst., pages 13308–13318, 2020

    2020

  3. [11]

    Iq-vfi: Implicit quadratic motion estimation for video frame interpolation

    Mengshun Hu, Kui Jiang, Zhihang Zhong, Zheng Wang, and Yinqiang Zheng. Iq-vfi: Implicit quadratic motion estimation for video frame interpolation. InIEEE Conf. Comput. Vis. Pattern Recog., pages 6410–6419, 2024. 1, 3

    2024

  4. [12]

    Time lens: Event-based video frame interpolation

    Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. InIEEE Conf. Comput. Vis. Pattern Recog., pages 16155–16164, 2021. 2, 3, 6, 7, 8

    2021

  5. [13]

    Timelens-xl: Real-time event-based video frame interpo- lation with large motion

    Yongrui Ma, Shi Guo, Yutian Chen, Tianfan Xue, and Jinwei Gu. Timelens-xl: Real-time event-based video frame interpo- lation with large motion. InEur. Conf. Comput. Vis., pages 178–194, 2025. 2, 3, 6, 7, 8

    2025

  6. [14]

    Lichtsteiner, C

    P. Lichtsteiner, C. Posch, and T. Delbruck. A 128 ×128 120 db 15 µs latency asynchronous temporal contrast vision sensor.IEEE J. Solid-State Circuits, 43(2):566–576, 2008. 1, 3

    2008

  7. [15]

    Event-based vision: A survey.IEEE Trans

    Guillermo Gallego, Tobi Delbr¨uck, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, J ¨org Conradt, Kostas Daniilidis, et al. Event-based vision: A survey.IEEE Trans. Pattern Anal. Mach. Intell., 44(1):154–180, 2020. 1, 3, 4

    2020

  8. [16]

    Eva2: Event-assisted video frame interpolation via cross- modal alignment and aggregation.IEEE Trans

    Zeyu Xiao, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Eva2: Event-assisted video frame interpolation via cross- modal alignment and aggregation.IEEE Trans. Comput. Imaging, 8:1145–1158, 2022. 2, 3

    2022

  9. [17]

    Superfast: 200×video frame interpolation via event camera

    Yue Gao, Siqi Li, Yipeng Li, Yandong Guo, and Qionghai Dai. Superfast: 200×video frame interpolation via event camera. IEEE Trans. Pattern Anal. Mach. Intell., 45(6):7764–7780,

  10. [18]

    Video frame interpolation via direct synthesis with the event-based reference

    Yuhan Liu, Yongjian Deng, Hao Chen, and Zhen Yang. Video frame interpolation via direct synthesis with the event-based reference. InIEEE Conf. Comput. Vis. Pattern Recog., pages 8477–8487, 2024. 2, 3, 6

    2024

  11. [19]

    Time lens++: Event-based frame interpolation with parametric non- linear flow and multi-scale fusion

    Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stama- tios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time lens++: Event-based frame interpolation with parametric non- linear flow and multi-scale fusion. InIEEE Conf. Comput. Vis. Pattern Recog., pages 17755–17764, 202...

    2022

  12. [20]

    Training weakly su- pervised video frame interpolation with events

    Zhiyang Yu, Yu Zhang, Deyuan Liu, Dongqing Zou, Xijun Chen, Yebin Liu, and Jimmy S Ren. Training weakly su- pervised video frame interpolation with events. InInt. Conf. Comput. Vis., pages 14589–14598, 2021

    2021

  13. [21]

    Timereplayer: Unlocking the potential of event cameras for video interpolation

    Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. InIEEE Conf. Comput. Vis. Pattern Recog., pages 17804–17813, 2022

    2022

  14. [22]

    Video in- terpolation by event-driven anisotropic adjustment of optical flow

    Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video in- terpolation by event-driven anisotropic adjustment of optical flow. InEur. Conf. Comput. Vis., pages 267–283, 2022. 3

    2022

  15. [23]

    Event-based video frame interpolation with cross- modal asymmetric bidirectional motion fields

    Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross- modal asymmetric bidirectional motion fields. InIEEE Conf. Comput. Vis. Pattern Recog., pages 18032–18042, 2023. 2, 3, 6, 7, 8

    2023

  16. [24]

    Video frame in- terpolation via adaptive convolution

    Simon Niklaus, Long Mai, and Feng Liu. Video frame in- terpolation via adaptive convolution. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 670–679, 2017. 3

    2017

  17. [25]

    Video frame inter- polation via adaptive separable convolution

    Simon Niklaus, Long Mai, and Feng Liu. Video frame inter- polation via adaptive separable convolution. InProceedings of the IEEE international conference on computer vision, pages 261–270, 2017

    2017

  18. [26]

    Adacof: Adaptive collaboration of flows for video frame interpolation

    Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive collaboration of flows for video frame interpolation. InIEEE Conf. Comput. Vis. Pattern Recog., pages 5316–5325, 2020. 3

    2020

  19. [27]

    Tta-evf: Test-time adaptation for event-based video frame interpolation via reliable pixel and sample estimation

    Hoonhee Cho, Taewoo Kim, Yuhwan Jeong, and Kuk-Jin Yoon. Tta-evf: Test-time adaptation for event-based video frame interpolation via reliable pixel and sample estimation. InIEEE Conf. Comput. Vis. Pattern Recog., pages 25701– 25711, 2024. 3

    2024

  20. [28]

    Uniinr: Event-guided unified rolling shutter cor- rection, deblurring, and interpolation

    Yunfan Lu, Guoqiang Liang, Yusheng Wang, Lin Wang, and Hui Xiong. Uniinr: Event-guided unified rolling shutter cor- rection, deblurring, and interpolation. InEuropean Confer- ence on Computer Vision, pages 1–20. Springer, 2025. 3

    2025

  21. [29]

    Tap-vid: A benchmark for tracking any point in a video.Adv

    Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Re- casens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video.Adv. Neural Inform. Process. Syst., pages 13610–13626, 2022. 3, 6

    2022

  22. [30]

    Tapir: Tracking any point with per-frame initialization and temporal refinement

    Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. InInt. Conf. Comput. Vis., pages 10061– 10072, 2023

    2023

  23. [31]

    Tracking everything everywhere all at once

    Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. InInt. Conf. Comput. Vis., pages 19795–19806, 2023

    2023

  24. [32]

    Pointodyssey: A large-scale synthetic dataset for long-term point tracking

    Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. InInt. Conf. Comput. Vis., pages 19855–19865, 2023

    2023

  25. [33]

    Local all-pair correspondence for point tracking.arXiv preprint arXiv:2407.15420, 2024

    Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seungry- ong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking.arXiv preprint arXiv:2407.15420, 2024. 5

    2024 arXiv

  26. [34]

    Co- tracker: It is better to track together.arXiv preprint arXiv:2307.07635, 2023

    Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co- tracker: It is better to track together.arXiv preprint arXiv:2307.07635, 2023. 3, 5

    2023 arXiv

  27. [35]

    Eklt: Asynchronous photometric feature track- ing using events and frames.Int

    Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. Eklt: Asynchronous photometric feature track- ing using events and frames.Int. J. Comput. Vis., 128(3):601– 618, 2020. 3

    2020

  28. [36]

    Data-driven feature tracking for event cameras

    Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. InIEEE Conf. Comput. Vis. Pattern Recog., pages 5642–5651, 2023

    2023

  29. [37]

    Tracking any point with frame-event fusion network at high frame rate.arXiv preprint arXiv:2409.11953,

    Jiaxiong Liu, Bo Wang, Zhen Tan, Jinpu Zhang, Hui Shen, and Dewen Hu. Tracking any point with frame-event fusion network at high frame rate.arXiv preprint arXiv:2409.11953,

    arXiv

  30. [38]

    Dense continuous-time optical flow from event cameras

    Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event cameras. IEEE Trans. Pattern Anal. Mach. Intell., 2024. 3, 6

    2024

  31. [39]

    Motion- prior contrast maximization for dense continuous-time motion estimation.arXiv preprint arXiv:2407.10802, 2024

    Friedhelm Hamann, Ziyun Wang, Ioannis Asmanis, Kenneth Chaney, Guillermo Gallego, and Kostas Daniilidis. Motion- prior contrast maximization for dense continuous-time motion estimation.arXiv preprint arXiv:2407.10802, 2024. 3

    2024 arXiv

  32. [40]

    Optical flow with semantic segmentation and localized layers

    Laura Sevilla-Lara, Deqing Sun, Varun Jampani, and Michael J Black. Optical flow with semantic segmentation and localized layers. InIEEE Conf. Comput. Vis. Pattern Recog., pages 3889–3898, 2016. 4

    2016

  33. [41]

    Every frame counts: Joint learning of video segmentation and optical flow

    Mingyu Ding, Zhe Wang, Bolei Zhou, Jianping Shi, Zhiwu Lu, and Ping Luo. Every frame counts: Joint learning of video segmentation and optical flow. InAAAI Conf. Artif. Intell., pages 10713–10720, 2020

    2020

  34. [42]

    Semarflow: Injecting semantics into unsupervised optical flow estimation for autonomous driving

    Shuai Yuan, Shuzhi Yu, Hannah Kim, and Carlo Tomasi. Semarflow: Injecting semantics into unsupervised optical flow estimation for autonomous driving. InInt. Conf. Comput. Vis., pages 9566–9577, 2023. 4

    2023

  35. [43]

    Slic superpixels compared to state-of-the-art superpixel methods.IEEE Trans

    Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine S¨usstrunk. Slic superpixels compared to state-of-the-art superpixel methods.IEEE Trans. Pattern Anal. Mach. Intell., pages 2274–2282, 2012. 4

    2012

  36. [44]

    A. Zhu, L. Yuan, K. Chaney, and K. Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. InIEEE Conf. Comput. Vis. Pattern Recog., pages 989–997,

  37. [45]

    E-raft: Dense optical flow from event cameras

    Mathias Gehrig, Mario Millh¨ausler, Daniel Gehrig, and Da- vide Scaramuzza. E-raft: Dense optical flow from event cameras. InInternational Conference on 3D Vision, pages 197–206, 2021. 5

    2021

  38. [46]

    Object recognition from local scale-invariant features

    David G Lowe. Object recognition from local scale-invariant features. InInt. Conf. Comput. Vis., pages 1150–1157, 1999. 5

    1999

  39. [47]

    Attention is all you need.Adv

    A Vaswani. Attention is all you need.Adv. Neural Inform. Process. Syst., 2017. 5

    2017

  40. [48]

    Flowformer: A transformer architecture for optical flow

    Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Flowformer: A transformer architecture for optical flow. InEur. Conf. Comput. Vis., pages 668–685, 2022. 6

    2022

  41. [49]

    Unflow: Unsu- pervised learning of optical flow with a bidirectional census loss

    Simon Meister, Junhwa Hur, and Stefan Roth. Unflow: Unsu- pervised learning of optical flow with a bidirectional census loss. InAAAI Conf. Artif. Intell., 2018. 6

    2018

  42. [50]

    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. InIEEE Conf. Comput. Vis. Pattern Recog., pages 19352–19362, 2024. 6

    2024

  43. [51]

    Video to events: Recycling video datasets for event cameras

    Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carri´o, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. InIEEE Conf. Comput. Vis. Pat- tern Recog., pages 3586–3595, 2020. 6

    2020

  44. [52]

    Kingma and J

    D. Kingma and J. Ba. Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980, 2014. 6

    2014 arXiv

  45. [53]

    Paszke, S

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. De- Vito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer. Auto- matic differentiation in pytorch. 2017. 6

    2017

  46. [54]

    Deep multi-scale convolutional neural network for dynamic scene deblurring

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. InIEEE Conf. Comput. Vis. Pattern Recog., pages 3883–3891, 2017. 6, 7, 8

    2017

  47. [55]

    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. InProceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 10663–10671,

  48. [56]

    Perception-oriented video frame interpolation via asymmetric blending

    Guangyang Wu, Xin Tao, Changlin Li, Wenyi Wang, Xiao- hong Liu, and Qingqing Zheng. Perception-oriented video frame interpolation via asymmetric blending. InIEEE Conf. Comput. Vis. Pattern Recog., pages 2753–2762, 2024. 6, 7, 8

    2024

  49. [57]

    Video frame interpolation transformer

    Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming- Hsuan Yang. Video frame interpolation transformer. InIEEE Conf. Comput. Vis. Pattern Recog., pages 17482–17491, 2022. 6, 7

    2022

  50. [58]

    Event-based background-oriented schlieren.IEEE Trans

    Shintaro Shiba, Friedhelm Hamann, Yoshimitsu Aoki, and Guillermo Gallego. Event-based background-oriented schlieren.IEEE Trans. Pattern Anal. Mach. Intell., 2023. 8

    2023

  51. [59]

    Event-based imag- ing velocimetry: an assessment of event-based cameras for the measurement of fluid flows.Exp

    Christian E Willert and Joachim Klinner. Event-based imag- ing velocimetry: an assessment of event-based cameras for the measurement of fluid flows.Exp. Fluids, 63(6):101, 2022. 8

    2022

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.