Pith. sign in

REVIEW 4 major objections 5 minor 39 references

A Delay Metric for Video Object Detection: What Average Precision Fails to Tell

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Average precision hides detection delay in video object detection

desk verdict mAP really does hide detection delay, and AD is a credible metric to expose it; just don't take the exponential model literally. read the letter →

arxiv 1908.06368 v2 pith:65BYV6CT submitted 2019-08-18 cs.CV

Huizi Mao , Xiaodong Yang , William J. Dally This is my paper · ORCID
classification cs.CV
keywords averagedelayvideoobjectdetectionevaluationmetricprecisionImageNetVIDTfalsepositiveratiotemporaldynamics
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

This paper argues that average precision (AP), the standard accuracy metric for video object detection, is blind to when a detector first recognizes an object, and proposes a complementary metric called average delay (AD) that measures how many frames pass before an instance is detected. AD is defined over six false-positive ratios and averages the inferred per-frame detection probability, so it rewards early detection without rewarding a detector that simply emits many false positives. Experiments on a new subset of ImageNet VID, named VIDT, show that several video-specific detectors, including key-frame, feature-aggregation, and cascaded methods, keep mAP roughly unchanged while substantially increasing detection delay. For latency-critical uses such as autonomous driving, the paper concludes, accuracy metrics should be accompanied by a delay metric.

What carries the argument

The load-bearing object is the AD formula, $\mathrm{AD} = 1 / ((1/R) \sum_r 1/(\bar{D}^*_r + 1)) - 1$, where $\bar{D}^*_r$ is the clipped mean delay to first detection at false-positive ratio $r$. The argument treats each detection as a Bernoulli event with constant probability $p$ per frame, so delay follows a discrete exponential distribution and $p$ is estimated as $1/(\text{mean clipped delay} + 1)$. Averaging $p$ rather than raw delay makes the metric a harmonic-style mean that weights small delays more heavily, preventing a detector with many very late detections from looking good, and the false-positive ratio controls the false-alarm budget so a detector cannot reduce delay by emitting many false positives.

What would settle it

Compare each detector's empirical delay distribution against the fitted discrete exponential model on VIDT, for example by computing the Kaplan-Meier survival curve of first-detection times and testing whether detection probability per frame is constant. The paper's own Figure 7, showing FGFA confidence rising steadily over frames, already suggests the constant-p model fails there; a quantitative test would estimate p(t) and show whether re-ranking detectors by a nonparametric mean delay changes the AD ordering.

Watch

Extended reading notes

Core claim

The central claim is that mAP cannot distinguish detectors with different temporal behavior: a detector that recognizes objects only late in their trajectory can match a detector that recognizes them immediately, as long as the total number of true positives over frames is similar. The paper introduces AD as the delay counterpart to AP: for each of six false-positive ratios it measures the clipped mean delay to first detection, converts that mean to a per-frame detection probability under a discrete exponential model, averages those probabilities, and converts back to frames. This design makes AD sensitive to retardation (suppressing early detections) and nearly insensitive to tail boost (raising confidence of late detections), behaviors that mAP, NAB scores, and the CaTDet delay metric do not share. On the VIDT subset, DFF with sparse key frames, FGFA with feature aggregation, and CaTDet all preserve mAP while raising AD relative to frame-by-frame baselines, and the paper interprets this as evidence that temporal information is often used to refine late detections rather than to detect earlier.

Load-bearing premise

The metric's numbers rest on the assumption that on each frame a detector detects an object with the same fixed probability, so delay follows a single exponential distribution; if a detector's per-frame detection probability changes with time since appearance, the reported AD values and rankings may partly reflect that assumption rather than the true temporal behavior.

Editorial extensions

If this is right

  • If mAP is retained as the headline accuracy number, video object detection papers should report AD alongside it; a method that improves or preserves mAP while worsening AD is not a strict improvement.
  • Latency-critical deployment, such as autonomous vehicle perception, should prefer detectors with low AD even when their mAP is slightly lower, because detections arriving after the decision deadline are effectively useless.
  • Methods that aggregate features over many frames (FGFA style) or use sparse key frames (DFF style) should be expected to trade delay for accuracy; the paper's measurements quantify that trade.
  • Anchor-scale choices affect delay as well as accuracy: adding small anchor scales improves AD for small instances, so delay evaluation can guide architecture choices beyond mAP.
  • AD's class-agnostic aggregation on VIDT means future evaluation datasets for delay should be built to contain instances entering at random times and locations, as VIDT does.

Reading between the lines

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

  • An inference beyond the paper: AD as defined is one summary statistic of the whole delay distribution; a natural extension would be a delay-recall curve or a time-aware precision-recall surface that exposes not just mean delay but the full distribution over instances.
  • The paper's exponential assumption is a convenience; a detector whose confidence ramps up gradually, as FGFA visibly does, has time-dependent detection probability, so future work could replace the single p with a time-varying hazard model and still use the same false-positive-ratio framework.
  • The paper evaluates established detectors but does not propose a training objective; a testable extension is to train with a delay-aware loss that penalizes late first detections, and check whether AD can be improved without sacrificing mAP.
  • Because AD is computed from tracklet-level first detections, it implicitly assumes clean track identities; extending the metric to tracking-by-detection systems with identity switches would require a policy for when a detection belongs to an existing instance.
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

4 major / 5 minor

Summary. The paper argues that average precision (AP) is insensitive to the temporal behavior of video object detectors and proposes a new metric, average delay (AD), to measure detection delay. AD is defined in Eq. (3) as a function of per-frame detection probabilities inferred from clipped mean delays, with a detection window W=30 and six false-positive ratios. To support the evaluation, the authors introduce VIDT, a subset of ImageNet VID with multiple tracklets. Experiments with DFF, FGFA, CaTDet, and standard image detectors show that methods can preserve or improve mAP while substantially worsening AD, and the authors conclude that video object detection should be evaluated with a delay metric, especially for latency-critical applications.

Significance. If the main empirical claim holds, the paper makes a useful methodological contribution: it identifies a blind spot in mAP-based video object detection evaluation, releases code and a curated subset (VIDT), and provides a concrete way to quantify algorithmic delay. The toy-case experiments in Table 2 are a nice sanity check of the metric's behavior under controlled manipulations. The central claim, however, rests on AD being a faithful summary of temporal detection behavior, and the paper's own evidence casts doubt on the stochastic model used to derive AD. The contribution is therefore significant but currently not fully validated.

major comments (4)
  1. [Sec. 3.1, Eqs. (1)-(2), Fig. 7, Table 3] The constant-p, independent-per-frame detection model underlying Eqs. (1)-(2) is contradicted by the paper's own evidence. Table 3 shows off-window percentages roughly double the exponential expectation (e.g., 10.2% vs. 5.3% for R-FCN), and Fig. 7 shows FGFA confidence gradually building over frames rather than appearing with constant probability. Under a time-varying detection probability, the quantity 1/(D*+1) is not a latent detection rate, and Eq. (3) becomes an arbitrary nonlinear weighting of clipped delays. The authors should test whether the reported AD rankings survive a model-free aggregation, for instance mean/median clipped delay, hazard-rate estimates, or time to 50% detection, and report rank correlations. Without such a robustness check, the main conclusion that 'most methods drastically increase detection delay' may be an artifact of the exponential assumption.
  2. [Sec. 3.3, Eq. (3)] The two key design choices, W=30 and the FP ratio set {0.1, 0.2, 0.4, 0.8, 1.6, 3.2}, are introduced without justification, yet AD is a function of both. Because the paper's headline comparisons (e.g., R-FCN 9.0 vs. DFF 11.6 vs. FGFA 10.2 in Fig. 6) are point estimates of AD, it is important to show that rankings are stable over reasonable variations in W (e.g., 10, 20, 50, 100) and over alternative FP-ratio grids, including an unweighted average of per-ratio delays. If rankings change, the metric's comparisons are not robust.
  3. [Sec. 6.4, Table 5] What is called a 'test of significance' is not a significance test. The 3-fold validation reports point estimates only and does not provide variances, confidence intervals, or any p-value for the differences between AD of R-FCN, DFF, and FGFA; moreover, folds are constructed from the dataset and are not independent samples of the underlying video distribution. The alternative dataset VIDT-2017 is a helpful additional check, but it also reports point estimates without uncertainty. The authors should add per-instance bootstrap or paired-difference confidence intervals to establish that the observed AD gaps are unlikely to arise from the small number of instances.
  4. [Sec. 4.2] The rule that an instance disappearing for more than 10 frames is counted as a new instance is an arbitrary definition that directly affects the measured delay: a detector that recognizes re-appearance quickly will appear to have low delay, while one that does not re-detect will be penalized. Since this threshold is a free parameter of the dataset construction, a sensitivity analysis (e.g., thresholds of 5, 10, 20 frames) is needed to show that the main conclusions do not hinge on this choice.
minor comments (5)
  1. [Abstract] 'A verage' in the title is a typo; also 'descend' should be 'decent' in the conclusion.
  2. [Sec. 3.1, Fig. 2] The caption contains a typo: 'delay dstribution' should be 'delay distribution'.
  3. [Sec. 3.3] In the sentence 'we derive the corresponding AD fom the averaged p', 'fom' should be 'from'.
  4. [Sec. 3.3, Eq. (3)] The notation ̄D*_r implicitly depends on the window W and the FP ratio r; it would be clearer to write ̄D*_r(W) or define the dependence explicitly.
  5. [Sec. 2.1] The reference to 'CaTDET [25]' in the background section should be capitalized consistently as 'CaTDet' throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: AD is a direct transform of measured detection delays, and the AP-insensitivity conclusion is an independent empirical comparison.

full rationale

AD is defined in Eq. 3 from measured per-instance delays: for each false-positive ratio r, D*_r is the clipped mean of frame delays (Eq. 2), converted to p = 1/(D*_r + 1), averaged across r, and mapped back to a delay value. Every quantity entering the final AD is a deterministic function of observed detection times; the only inputs are the manually chosen window W = 30 and six FP ratios. No model parameter is fitted to produce the paper's conclusion, and the central claim (AP is insensitive to delay differences) is an empirical comparison of two independently computed metrics, mAP and AD, on VIDT. The exponential-distribution assumption motivates the reciprocal transform, but it is not load-bearing for the claim: even if the assumption were false, AD would still be a well-defined deterministic summary of clipped mean delays, and the reported rankings would remain computable quantities. The only self-citation of note is CaTDet [25], which is used as a prior delay metric to argue against precision-based false-alarm control and as one baseline detector; the central AP-insensitivity result does not depend on any unverified claim from [25]. No uniqueness theorem is imported, and no known result is renamed. The model-fidelity concerns about heavy tails and gradual confidence buildup are validity questions about whether AD is the right metric, not circularity in the derivation. The paper is therefore self-contained with respect to its main claim, and no circular step is present.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on three kinds of input: a statistical model of delay (constant-p exponential), a dataset construction convention (tracklet splitting at 10-frame gaps, selection of VIDT), and a choice of operating points (W=30, six FP ratios). The first is acknowledged to be approximate, and the latter two are hand-set; the empirical conclusion could change under different choices.

free parameters (3)
  • Detection window W = 30 frames
    Used in Eq. 2 to clip delays before computing the mean; chosen by hand. AD values and rankings depend on this choice.
  • False-positive ratio set = [0.1, 0.2, 0.4, 0.8, 1.6, 3.2]
    Six operating points at which delays are measured and averaged in Eq. 3; chosen by hand and not justified against alternatives.
  • Tracklet split threshold = 10 frames
    An instance disappearing for more than 10 consecutive frames is marked as a new instance in VIDT, changing instance count from 555 to 666; this affects the delay measurements.
assumptions (3)
  • domain assumption Per-frame detection is modeled as an independent Bernoulli trial with constant probability p for each instance, so delay follows a discrete exponential distribution.
    Used in Section 3.1 to infer p from the sample mean; the paper concedes the actual distribution has a heavy tail, making the assumption approximate.
  • ad hoc to paper The annotated track identities in VID are correct, and a gap of more than 10 frames defines a new instance for delay measurement.
    This convention is introduced in Section 4.2 to construct VIDT and directly affects which delays are measured.
  • domain assumption Comparing detectors at fixed false-positive ratios provides a fair false-alarm control across methods.
    Section 3.2 argues precision control is unfair and FP ratio is preferable; this is a modeling choice, not derived from data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Delay Metric for Video Object Detection: What Average Precision Fails to Tell." pith.science (2026). https://pith.science/paper/65BYV6CT

@misc{pith2026190806368,
  author       = {Pith},
  title        = {Pith review of: A Delay Metric for Video Object Detection: What Average Precision Fails to Tell},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/65BYV6CT}},
  note         = {Machine review of arXiv:1908.06368}
}
read the original abstract

Average precision (AP) is a widely used metric to evaluate detection accuracy of image and video object detectors. In this paper, we analyze object detection from videos and point out that AP alone is not sufficient to capture the temporal nature of video object detection. To tackle this problem, we propose a comprehensive metric, average delay (AD), to measure and compare detection delay. To facilitate delay evaluation, we carefully select a subset of ImageNet VID, which we name as ImageNet VIDT with an emphasis on complex trajectories. By extensively evaluating a wide range of detectors on VIDT, we show that most methods drastically increase the detection delay but still preserve AP well. In other words, AP is not sensitive enough to reflect the temporal characteristics of a video object detector. Our results suggest that video object detection methods should be additionally evaluated with a delay metric, particularly for latency-critical applications such as autonomous vehicle perception.

Figures

Figures reproduced from arXiv: 1908.06368 by the authors.

Figure 1
Figure 1. AD does not strongly correlate with mAP. Many algo [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. A delay histogram of R-FCN (ResNet-101) on VIDT at [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. A toy example to illustrate that using precision as the [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Snippets in the validation set of VID. Top: an ideal video snippet for delay evaluation with multiple instances emerging randomly [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Number of instances per class is highly imbalanced [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: How DFF and FGFA affect mAP and AD. Here [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 9
Figure 9. Figure 9: A zoomed-in plot of delay distribution of multiple detec [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 10
Figure 10. Figure 10: Examples of hard instances that have larger than 100- [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]
Figure 11
Figure 11. Figure 11: AD by class: we only demonstrate the six video ob [PITH_FULL_IMAGE:figures/full_fig_p008_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 34 canonical work pages

  1. [1]

    https://deepdrive.berkeley.edu/project/low- latency-deep-inference-self-driving-vehicles

    Berkeley DeepDrive: Low latency deep inference for self- driving vehicles. https://deepdrive.berkeley.edu/project/low- latency-deep-inference-self-driving-vehicles. 2

    work page

  2. [2]

    Object detection in video with spatiotemporal sampling networks

    Gedas Bertasius, Lorenzo Torresani, and Jianbo Shi. Object detection in video with spatiotemporal sampling networks. In ECCV, 2018. 1, 2, 6

    work page 2018

  3. [3]

    Adaptive neural networks for efficient inference

    Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. Adaptive neural networks for efficient inference. In ICML, 2017. 6

    work page 2017

  4. [4]

    Optimizing video object detection via a scale-time lattice

    Kai Chen, Jiaqi Wang, Shuo Yang, Xingcheng Zhang, Yuan- jun Xiong, Chen Change Loy, and Dahua Lin. Optimizing video object detection via a scale-time lattice. In CVPR,

    work page

  5. [5]

    R-FCN: Object detection via region-based fully convolutional networks

    Jifeng Dai, Yi Li, Kaiming He, and Jian Sun. R-FCN: Object detection via region-based fully convolutional networks. In NeurIPS, 2016. 1, 2

    work page 2016

  6. [6]

    ImageNet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 1, 2, 4

    2009

  7. [7]

    The pascal visual object classes (VOC) challenge

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (VOC) challenge. IJCV, 2010. 2, 5

    work page 2010

  8. [8]

    Detect to track and track to detect

    Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Detect to track and track to detect. In ICCV, 2017. 1, 2

    work page 2017

Show all 39 references
  1. [9]

    Vision meets robotics: The KITTI dataset

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. Interna- tional Journal of Robotics Research, 2013. 1, 4

    2013

  2. [10]

    Impression network for video object detection

    Congrui Hetang, Hongwei Qin, Shaohui Liu, and Jun- jie Yan. Impression network for video object detection. arXiv:1712.05896, 2017. 5

    2017 arXiv

  3. [11]

    MobileNets: Efficient con- volutional neural networks for mobile vision applications

    Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco An- dreetto, and Hartwig Adam. MobileNets: Efficient con- volutional neural networks for mobile vision applications. arXiv:1704.04861, 2017. 2

    2017 arXiv

  4. [12]

    T-CNN: Tubelets with convolutional neural networks for object detection from videos

    Kai Kang, Hongsheng Li, Junjie Yan, Xingyu Zeng, Bin Yang, Tong Xiao, Cong Zhang, Zhe Wang, Ruohui Wang, Xiaogang Wang, et al. T-CNN: Tubelets with convolutional neural networks for object detection from videos. TCSVT,

  5. [13]

    A discriminative model with multiple temporal scales for action prediction

    Yu Kong, Dmitry Kit, and Yun Fu. A discriminative model with multiple temporal scales for action prediction. In ECCV, 2014. 2

    2014

  6. [14]

    Quickest moving object detection

    Dong Lao and Ganesh Sundaramoorthi. Quickest moving object detection. arXiv:1605.07369, 2016. 2

    2016 arXiv

  7. [15]

    Minimum delay ob- ject detection from video

    Dong Lao and Ganesh Sundaramoorthi. Minimum delay ob- ject detection from video. In Proceedings of the IEEE Inter- national Conference on Computer Vision, pages 5097–5106,

  8. [16]

    Evaluating real-time anomaly detection algorithms–the numenta anomaly bench- mark

    Alexander Lavin and Subutai Ahmad. Evaluating real-time anomaly detection algorithms–the numenta anomaly bench- mark. In ICMLA, 2015. 2, 5

    2015

  9. [17]

    The ar- chitectural implications of autonomous driving: Constraints and acceleration

    Shih-Chieh Lin, Yunqi Zhang, Chang-Hong Hsu, Matt Skach, Md E Haque, Lingjia Tang, and Jason Mars. The ar- chitectural implications of autonomous driving: Constraints and acceleration. In ASPLOS, 2018. 2

    2018

  10. [18]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll´ar. Focal loss for dense object detection. In ICCV,

  11. [19]

    Microsoft CoCo: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft CoCo: Common objects in context. In ECCV, 2014. 1, 2

    2014

  12. [20]

    Mobile video object detec- tion with temporally-aware feature maps

    Mason Liu and Menglong Zhu. Mobile video object detec- tion with temporally-aware feature maps. In CVPR, 2018. 1, 2, 6

    2018

  13. [21]

    SSD: Single shot multibox detector

    Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. SSD: Single shot multibox detector. In ECCV, 2016. 1, 2

    2016

  14. [22]

    Detect or track: Towards cost-effective video object detec- tion/tracking

    Hao Luo, Wenxuan Xie, Xinggang Wang, and Wenjun Zeng. Detect or track: Towards cost-effective video object detec- tion/tracking. In AAAI, 2019. 2, 5

    2019

  15. [23]

    Learning ac- tivity progression in LSTMs for activity detection and early detection

    Shugao Ma, Leonid Sigal, and Stan Sclaroff. Learning ac- tivity progression in LSTMs for activity detection and early detection. In CVPR, 2016. 2

    2016

  16. [24]

    Budget-aware activity detection with a recurrent policy network

    Behrooz Mahasseni, Xiaodong Yang, Pavlo Molchanov, and Jan Kautz. Budget-aware activity detection with a recurrent policy network. In BMVC, 2018. 2

    2018

  17. [25]

    Catdet: Cascaded tracked detector for efficient object detection from video

    Huizi Mao, Taeyoung Kong, and William J Dally. Catdet: Cascaded tracked detector for efficient object detection from video. arXiv preprint arXiv:1810.00434, 2018. 1, 2, 3, 5, 6

    2018 arXiv

  18. [26]

    Towards real-time object detection on embedded systems

    Huizi Mao, Song Yao, Tianqi Tang, Boxun Li, Jun Yao, and Yu Wang. Towards real-time object detection on embedded systems. TETC, 2018. 2

    2018

  19. [27]

    Moped: A scalable and low latency object recognition and pose estimation system

    Manuel Martinez, Alvaro Collet, and Siddhartha S Srinivasa. Moped: A scalable and low latency object recognition and pose estimation system. In ICRA, 2010. 2

    2010

  20. [28]

    A large- scale benchmark dataset for event recognition in surveillance video

    Sangmin Oh, Anthony Hoogs, Amitha Perera, Naresh Cun- toor, Chia-Chih Chen, Jong Taek Lee, Saurajit Mukherjee, JK Aggarwal, Hyungtae Lee, and Larry Davis. A large- scale benchmark dataset for event recognition in surveillance video. In CVPR, 2011. 4

    2011

  21. [29]

    Quickest detection

    Vincent Poor and Olympia Hadjiliadis. Quickest detection

  22. [30]

    YouTube-BoundingBoxes: A large high-precision human-annotated data set for object detection in video

    Esteban Real, Jonathon Shlens, Stefano Mazzocchi, Xin Pan, and Vincent Vanhoucke. YouTube-BoundingBoxes: A large high-precision human-annotated data set for object detection in video. In CVPR, 2017. 4

    2017

  23. [31]

    Faster R-CNN: Towards real-time object detection with re- gion proposal networks

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with re- gion proposal networks. In NeurIPS, 2015. 1, 2

    2015

  24. [32]

    Encouraging LSTMs to anticipate actions very early

    Mohammad Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Basura Fernando, Lars Petersson, and Lars Andersson. Encouraging LSTMs to anticipate actions very early. In ICCV, 2017. 2

    2017

  25. [33]

    Mul- tilayer and multimodal fusion of deep neural networks for video classification

    Xiaodong Yang, Pavlo Molchanov, and Jan Kautz. Mul- tilayer and multimodal fusion of deep neural networks for video classification. In ACM Multimedia, 2016. 2

    2016

  26. [34]

    Making convolutional networks recurrent for visual sequence learn- ing

    Xiaodong Yang, Pavlo Molchanov, and Jan Kautz. Making convolutional networks recurrent for visual sequence learn- ing. In CVPR, 2018. 1

    2018

  27. [35]

    STEP: Spatio-temporal pro- gressive learning for video action detection

    Xitong Yang, Xiaodong Yang, Ming-Yu Liu, Fanyi Xiao, Larry Davis, and Jan Kautz. STEP: Spatio-temporal pro- gressive learning for video action detection. In CVPR, 2019. 2

    2019

  28. [36]

    BDD100K: A diverse driving video database with scalable annotation tool- ing

    Fisher Yu, Wenqi Xian, Yingying Chen, Fangchen Liu, Mike Liao, Vashisht Madhavan, and Trevor Darrell. BDD100K: A diverse driving video database with scalable annotation tool- ing. arXiv:1805.04687, 2018. 4

    2018 arXiv

  29. [37]

    ShuffleNet: An extremely efficient convolutional neural net- work for mobile devices

    Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. ShuffleNet: An extremely efficient convolutional neural net- work for mobile devices. In CVPR, 2018. 2

    2018

  30. [38]

    Flow-guided feature aggregation for video object de- tection

    Xizhou Zhu, Yujie Wang, Jifeng Dai, Lu Yuan, and Yichen Wei. Flow-guided feature aggregation for video object de- tection. In ICCV, 2017. 1, 6

    2017

  31. [39]

    Deep feature flow for video recognition

    Xizhou Zhu, Yuwen Xiong, Jifeng Dai, Lu Yuan, and Yichen Wei. Deep feature flow for video recognition. In CVPR,

Pith tools

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