Pith. sign in

REVIEW 4 major objections 6 minor 36 references

Flow Guided Short-term Trackers with Cascade Detection for Long-term Tracking

T0 review · 4 major / 6 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 A long-term tracker built from MDNet, SiamRPN++, and a flow-guided cascade detector can re-catch targets after occlusion or out-of-view events, reporting 0.5405 F-score on VOT-2018 LTB35 and large gains over its MDNet baseline on…

desk verdict Competent integration of MDNet, SiamRPN++, and a flow-guided cascade, but the state-of-the-art claim outruns the evidence. read the letter →

arxiv 1909.00319 v1 pith:V76EBEOJ submitted 2019-09-01 cs.CV eess.IV

Han Wu , Xueyuan Yang , Yong Yang , Guizhong Liu This is my paper · ORCID
classification cs.CVeess.IV
keywords long-termobjecttrackingsingleMDNetSiamRPN++cascadedetectionopticalflowVOT-2018LTB35VisDrone-SOT2019
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

Long-term visual tracking requires knowing not only where the target is, but also when it has disappeared and how to catch it when it reappears. This paper claims that a standard short-term tracker can be turned into an effective long-term tracker by adding a judgement module and a cascade detection module. The judgement module decides, from two complementary scores, whether the reported box actually contains the target; the cascade detection module expands its search from a local neighborhood to the whole frame, guided by optical flow, and re-locates the target after occlusion or out-of-view events. On the VOT-2018 LTB35 benchmark the combined system reaches an F-score of 0.5405, and on VisDrone-SOT2019 it improves over the MDNet baseline by 10.4 success points and 17 precision points on the test set.

What carries the argument

The load-bearing object is the three-module pipeline: a short-term tracking module (MDNet plus SiamRPN++), a judgement module, and a cascade detection module. The judgement module works on two numbers: the MDNet classification score $S_c$ and the SiamRPN++ similarity score $S_s$, compared against thresholds $\theta_{\mathrm{mid}}=0.5$ and $\theta_{\mathrm{low}}=0.1$. Depending on which score is above or below threshold, the algorithm either accepts the box, resamples around it with Gaussian sampling, compensates camera motion with optical flow from PWC-Net, or regresses the box. When both scores fall below the low threshold, the cascade detector starts: it searches locally with flow-guided sampling, then with GA-RPN proposals in a $5^2$-area patch, then an $18^2$-area patch, then the whole frame, until both scores clear the threshold. This mechanism is what lets the tracker survive long out-of-view intervals.

What would settle it

Disable the cascade detection module while keeping the same judgement logic, then rerun on VisDrone-SOT2019 test; if the 10.4 success-point and 17 precision-point gains over MDNet persist, re-detection is not what carries the result. Alternatively, sweep the two fixed thresholds per sequence; if no single setting reproduces the reported F-score on LTB35, the fixed rule is not the mechanism.

Watch

Extended reading notes

Core claim

The central claim is that target disappearance can be handled by close cooperation between a short-term tracker, a failure judge, and a detector, rather than by building a one-stage long-term model. The paper assembles a short-term module from two complementary trackers: MDNet contributes an online-updated classifier with a classification score, and SiamRPN++ contributes a fixed Siamese similarity score. A rule-based judgement module combines these scores and, on failure, hands control to a cascade detector that first searches around the last position with Gaussian sampling and optical-flow compensation, then widens to a proposed-box search and finally to the global frame. The reported result is that this pipeline, called flow_MDNet_RPN, achieves the best F-score on VOT-2018 LTB35 among the compared trackers and large success and precision gains over its MDNet baseline on VisDrone-SOT2019 test sequences.

Load-bearing premise

The whole gain rests on two fixed numbers in the judgement module, 0.5 and 0.1, reliably telling apart frames where the target is visible from frames where it is gone; if they misfire, the detector starts too early or too late and the improvement disappears.

Editorial extensions

If this is right

  • The proposed flow_MDNet_RPN reports both a target box and a presence confidence, so it is suited to practical tasks that need to know when the target is absent.
  • The cascade detection module expands from local to global search, so the tracker can recover targets that reappear far from their last location.
  • Optical-flow compensation in the judgement and detection steps is intended to handle camera motion, a common failure source in drone videos.
  • On the reported benchmarks, the largest gaps over MDNet and SiamRPN++ come from sequences where the target leaves view, suggesting re-detection is the main source of improvement.

Reading between the lines

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

  • Extension: an untested improvement implied by the design is learning the judgement thresholds per scene or adapting them online; the paper itself notes the fixed thresholds are closely related to the scene.
  • Extension: because the detector is activated only by the fixed thresholds, the same pipeline could be combined with a learned target-presence classifier to remove the manual threshold choice.
  • Extension: on sequences with frequent distractors, the SiamRPN++ similarity score may need a higher threshold to avoid false re-detection, so the reported 0.6095 precision on LTB35 could degrade on denser scenes.
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 / 6 minor

Summary. The paper proposes flow_MDNet_RPN, a long-term single-object tracker that couples two short-term trackers (MDNet and SiamRPN++) with a rule-based judgement module and a cascade detection module. The judgement module uses MDNet classification scores, SiamRPN++ similarity scores, and PWC-Net optical flow to decide whether the target is present, and the detection module expands the search area from local to global when the target is deemed absent. The method is evaluated on VOT-2018 LTB35 and VisDrone-SOT2019 validation and test sets, reporting an F-score of 0.5405 on LTB35 and success/precision improvements over MDNet and SiamRPN++ on VisDrone.

Significance. If the reported results are reproducible, the paper provides a useful engineering recipe for converting short-term trackers into long-term ones: the combination of two complementary short-term trackers with a flow-guided re-detection mechanism gives consistent gains over its two baselines on long-term benchmarks. The method is described in enough detail for reimplementation, and the evaluation uses public benchmarks with standard metrics. However, the state-of-the-art claim is not backed by comparison with strong long-term baselines in the LTB35 table, and the central decision thresholds are not validated; the significance is therefore conditional on additional evidence.

major comments (4)
  1. [3.3, 4.2] The judgement thresholds theta_mid=0.5 and theta_low=0.1 are hand-set, and no sensitivity analysis or ablation is reported. The authors themselves note in Section 4.2 that pre-defined failure-judgement thresholds are 'closely related to the scene' in their discussion of MBMD, which applies equally to their own thresholds. Because these thresholds control whether the cascade detector is triggered, the reported F-score and success/precision gains could be an artifact of the chosen operating point rather than of the flow-guided cascade. A threshold sweep and an ablation that turns off the judgement/detection modules are needed to support the central claim.
  2. [4.2, Table 1] Table 1 compares only SiamRPN++ and MDNet on VOT-2018 LTB35, yet Section 4.2 states that the proposed tracker 'achieves state-of-the-art performance' on this dataset. Without comparisons to other published long-term trackers on LTB35 (e.g., MBMD, DaSiamRPN_LT, TLD, LCT), the state-of-the-art claim is unsupported. The comparison with MBMD in Figure 5 is limited to VisDrone validation and is only described qualitatively in the text; numerical results for that figure or a supplementary table would strengthen the claim.
  3. [4.2] The sentence 'our proposed algorithm which has been trained on the VisDrone-SOT2019 training dataset is more suitable for video taken by drones' suggests that the method or its components were trained on VisDrone training data before evaluation on the validation and test splits. This is not described in the implementation details (Section 4.1), which list only ImageNet/COCO/YouTube-BB pretraining. If VisDrone training data was used, it should be disclosed explicitly, since it affects the fairness of comparisons against trackers not trained on VisDrone; if not, the sentence should be corrected.
  4. [4.2, Table 1] The reported numbers are single-run results despite the stochastic nature of MDNet's Gaussian sampling and online updates; no error bars, multiple-run statistics, or variance estimates are provided. Since several reported improvements (e.g., 5.2 success points over MDNet on VisDrone validation) are modest relative to likely run-to-run variance, the quantitative claims need at least a statement of variance or a deterministic protocol.
minor comments (6)
  1. [1] The sentence 'Finally, we summarize our work in Section 1' should refer to Section 5 (Conclusion).
  2. [4.2, Eq. (1)] Equation (1) is garbled by duplicated lines and likely a rendering issue; please fix the formula and define the threshold theta_F explicitly.
  3. [Figures 2-4] The flowcharts and framework figures are low-resolution, and the symbols for the scores are barely legible; please provide higher-quality versions.
  4. [3.3] The notation switches between 'classification score s_c' and 'similarity score s_s' and later uses 'confidence s_t' without a definition; please define all scores and their ranges (e.g., whether MDNet scores are probabilities or raw classifier outputs).
  5. [References] Several references have incomplete author lists (e.g., [9]) and inconsistent formatting; please check the bibliography.
  6. [4.2] In the VisDrone validation paragraph, the text says 'ours is much higher than the performance of the baseline algorithm MDNet and SiamRPN++' but does not provide numerical results for TLD, LCT, and MBMD in a table; adding the actual values would help the reader.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivation-level circularity: the reported gains are measured against external benchmark ground truth, and the method's learned components are independently published modules rather than self-citations.

full rationale

The paper's central claim is that combining MDNet and SiamRPN++ with a flow-guided cascade detection and a judgement module improves long-term tracking. The evaluations on VOT-2018 LTB35 and VisDrone-SOT2019 use benchmark ground-truth labels that are not used to fit any parameter of the proposed algorithm; MDNet, SiamRPN++, PWC-Net and GA-RPN are cited external works, and no load-bearing argument reduces to a self-citation or to a uniqueness theorem imported from the authors' own prior work. The thresholds theta_mid = 0.5 and theta_low = 0.1 are hand-set design choices, and the authors themselves acknowledge in Sec. 4.2 that pre-defined failure-judgement thresholds are 'related to the scene closely,' but that is a generalization and calibration concern, not a circular derivation: the decision rule is not defined in terms of the benchmark output, and no reported quantity is a renamed fit of the data it is said to predict. The sentence claiming the algorithm 'has been trained on the VisDrone-SOT2019 training dataset' is not backed by the implementation details, yet even if true it would concern dataset-specific tuning of the validation comparison, not equivalence by construction. No equation in the paper reduces to its own input, so no circular step is identified.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The tracker is a software assembly of existing modules; the only new hand-set quantities are verification thresholds and search expansion factors. No new physical entities, forces, or conserved quantities are introduced.

free parameters (3)
  • Judgement threshold theta_mid = 0.5
    Hand-set threshold on the SiamRPN++ similarity score in Section 3.3. It controls when tracking is judged successful and when detection is triggered, so the central result depends on it.
  • Judgement threshold theta_low = 0.1
    Hand-set lower threshold in Section 3.3; if the similarity score falls below it, the short-term tracker is declared failed and the detection module starts.
  • Cascade search-area expansion factors = 5^2 and 18^2 times target area
    Chosen by hand in Section 3.4 to define the local, medium, and global search stages. They determine how quickly the detector can reacquire the target and are not justified by data.
assumptions (4)
  • domain assumption Pretrained networks (MDNet on ImageNet VID; SiamRPN++ on COCO, ImageNet DET, ImageNet VID, YouTube-BoundingBoxes; PWC-Net on FlyingChairs and FlyingThings3D) provide reliable features for arbitrary target tracking without further adaptation.
    Invoked throughout Sections 3.2 and 3.4; the method has no mechanism to verify that these features transfer to the benchmark videos beyond the reported results.
  • ad hoc to paper The similarity score from SiamRPN++ and the classification score from MDNet are comparable across frames, and fixed thresholds separate target-visible from target-absent states.
    Section 3.3 builds the judgement module on this rule; the authors themselves note in Section 4.2 that fixed pre-defined thresholds are scene-related, which weakens the assumption.
  • domain assumption Optical flow from PWC-Net estimates global motion well enough to shift the sampling center toward the true target position when the target disappears.
    Used in Sections 3.3 and 3.4 to guide Gaussian sampling and detection after tracking failure; no quantitative validation of the flow compensation is given.
  • domain assumption Benchmark annotations in VOT-2018 LTB35 and VisDrone-SOT2019 are correct and the evaluation protocol is applied consistently.
    The central numbers are taken from external benchmarks; the paper does not provide its own error analysis or annotation check.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Flow Guided Short-term Trackers with Cascade Detection for Long-term Tracking." pith.science (2026). https://pith.science/paper/V76EBEOJ

@misc{pith2026190900319,
  author       = {Pith},
  title        = {Pith review of: Flow Guided Short-term Trackers with Cascade Detection for Long-term Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V76EBEOJ}},
  note         = {Machine review of arXiv:1909.00319}
}
read the original abstract

Object tracking has been studied for decades, but most of the existing works are focused on the short-term tracking. For a long sequence, the object is often fully occluded or out of view for a long time, and existing short-term object tracking algorithms often lose the target, and it is difficult to re-catch the target even if it reappears again. In this paper a novel long-term object tracking algorithm flow_MDNet_RPN is proposed, in which a tracking result judgement module and a detection module are added to the short-term object tracking algorithm. Experiments show that the proposed long-term tracking algorithm is effective to the problem of target disappearance.

Figures

Figures reproduced from arXiv: 1909.00319 by the authors.

Figure 1
Figure 1. The overall framework of the proposed long-term [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. We make different decisions according to [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The framework of proposed cascade detection module, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Success plot and precision plot of VisDrone-SOT2019 [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Success plot and precision plot of VisDrone-SOT2019 [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Visual results of our tracker, along with SiamRPN++ [ [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 31 canonical work pages

  1. [1]

    Gladh, M

    S. Gladh, M. Danelljan, F. S. Khan, and M. Felsberg. Deep motion features for visual tracking. International Conference on Pattern Recognition, 1243-1248, 2016

    work page 2016

  2. [2]

    J. F. Henriques, R. Caseiro, P. Martins, and J. Batista. High- Speed Tracking with Kerneli zed Correlation Filters. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(3), 583-596, 2015

    work page 2015

  3. [3]

    Danelljan, G

    M. Danelljan, G. Bhat, F. S. Khan, and M. Felsberg. ECO: Efficient Convolution Operators for Tracking. Computer Vision and Pattern Recognition, 6931-6939, 2017

    work page 2017

  4. [4]

    Nam, and B

    H. Nam, and B. Han. Learning Multi-domain Convolutional Neural Networks for Visual Tracking. Computer Vision and Pattern Recognition, 4293-4302, 2016

    work page 2016

  5. [5]

    Lukezic, L

    A. Lukezic, L. C. Zajc, T. Vojir, J. Matas, and M. Kristan. Now you see me: evaluating performance in long-term visual tracking. arXiv preprint arXiv:1804.07056, 2018

    arXiv 2018

  6. [6]

    Mueller, N

    M. Mueller, N. Smith, and B. Ghanem. A Benchmark and Simulator for UAV Tracking. European Conference on Computer Vision, 2(2): 445-461, 2016

    work page 2016

  7. [7]

    Long-Term Visual Object Tracking Benchmark

    A. Moudgil, and V. Gandhi. Long-Term Visual Object Tracking Benchmark. arXiv preprint arXiv:1712.01358, 2017

    work page Pith review arXiv 2017

  8. [8]

    Valmadre, L

    J. Valmadre, L. Bertinetto, J. F. Henriques, R. Tao, A. Vedaldi, A. W. Smeulders, and E. Gavves. Long-Term Tracking in the Wild: A Benchmark. European Conference on Computer Vision, 692-707, 2018

    work page 2018

Show all 36 references
  1. [9]

    Kristan, A

    M. Kristan, A. Leonardis, J. Matas, M. Felsberg, R. Pflugfelder, L. C. Zajc, and Z. He. The Sixth Visual Object Tracking VOT2018 Challenge Results. European Conference on Computer Vision, 3-53, 2018

    2018

  2. [10]

    P. Zhu, L. Wen, X. Bian, H. Ling, and Q. Hu. Vision meets drones: A challenge. arXiv preprint arXiv:1804.07437, 2018

    2018 arXiv

  3. [11]

    Y. Wu, J. Lim, and M. Yang. Online Object Tracking: A Benchmark. Computer Vision and Pattern Recognition, 2411-2418, 2013

    2013

  4. [12]

    Kristan, A

    M. Kristan, A. Leonardis, J. Matas, M. Felsberg, R. P. Pflugfelder, L. C. Zajc, and Z. He. The Visual Object Tracking VOT2017 Challenge Results. International Conference on Computer Vision, 1949-1972, 2017

    1949

  5. [13]

    D. S. Bolme, J. R. Beveridge, B. A. Draper, and Y. M. Lui. Visual Object Tracking using Adaptive Correlation Filters. Computer Vision and Pattern Recognition, 2544-2550, 2010

    2010

  6. [14]

    Danelljan, G

    M. Danelljan, G. Hager, F. S. Khan, and M. Felsberg. Accurate Scale Estimation for Robust Visual Tracking. British Machine Vision Conference, Nottingham, September 1-5, 2014. BMVA Press, 2014

    2014

  7. [15]

    Danelljan, G

    M. Danelljan, G. Hager, F. S. Khan, and M. Felsberg. Learning Spatially Regularized Correlation Filters for Visual Tracking. International Conf erence on Computer Vision, 4310-4318, 2015

    2015

  8. [16]

    Bertinetto, J

    L. Bertinetto, J. Valmadre, J. F. Henriques, A. Vedaldi, and P. H. Torr. Fully-Convolutional Siamese Networks for Object Tracking. European Conference on Computer Vision, 850-865, 2016

    2016

  9. [17]

    D. Held, S. Thrun, and S. Savarese. Learning to track at 100 fps with deep regression networks. European Conference on Computer Vision, 749-765. Springer, Cham, 2016

    2016

  10. [18]

    Valmadre, L

    J. Valmadre, L. Bertinetto, J. F. Henriques, A. Vedaldi, and P. H. Torr. End-to-End Representation Learning for Correlation Filter Based Tracking. Computer Vision and Pattern Recognition, 5000-5008, 2017

    2017

  11. [19]

    R. Tao, E. Gavves, and A. W. Smeulders. Siamese Instance Search for Tracking. Computer Vision and Pattern Recognition, 1420-1429, 2016

    2016

  12. [20]

    B. Li, J. Yan, W. Wu, Z. Zhu, and X. Hu. High Performance Visual Tracking with Siamese Region Proposal Network. Computer Vision and Pattern Recognition, 8971-8980, 2018

    2018

  13. [21]

    Z. Zhu, Q. Wang, B. Li, W. Wu, J. Yan, and W. Hu. Distractor-aware Siamese Networks for Visual Object Tracking. European Conference on Computer Vision, 103- 119, 2018

    2018

  14. [22]

    B. Li, W. Wu, Q. Wang, F. Zhang, J. Xing, and J. Yan. Siamrpn++: Evolution of siamese visual tracking with very deep networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 4282-4291, 2019

    2019

  15. [23]

    L. Wen, P. Zhu, D. Du, X. Bian, H. Ling, Q. Hu, and Q. Nie. VisDrone-SOT2018: The Vision Meets Drone Single-Object Tracking Challenge Results. In Proceedings of the European Conference on Computer Vision, 469-495, 2018

    2018

  16. [24]

    Redmon, A

    J. Redmon, A. Farhadi. YOLOv3: An Incremental Improvement. arXiv preprint arXiv:1804.02767, 2018

    2018 arXiv

  17. [25]

    Kalal, K

    Z. Kalal, K. Mikolajczyk, and J . Matas. Tracking-Learning- Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(7), 1409-1422, 2012

    2012

  18. [26]

    C. Ma, X. Yang, C. Zhang, and M. H. Yang. Long-term correlation tracking. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 5388-5396, 2015

    2015

  19. [27]

    Zhang, D

    Y. Zhang, D. Wang, L. Wang, J . Qi, and H. Lu. Learning regression and verification networks for long-term visual tracking. arXiv preprint arXiv:1809.04320, 2018

    2018 arXiv

  20. [28]

    Russakovsky, J

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, and A. C. Berg. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3), 211-252, 2015

    2015

  21. [29]

    E. Real, J. Shlens, S. Mazzocchi, X. Pan, and V. Vanhoucke. YouTube-BoundingBoxes: A Large High-Precision Human- Annotated Data Set for Object Detection in Video. Computer Vision and Pattern Recognition, 7464-7473, 2017

    2017

  22. [30]

    K. He, X. Zhang, S. Ren, and J. Sun. Deep Residual Learning for Image Recognition. Computer Vision and Pattern Recognition, 770-778, 2016

    2016

  23. [31]

    Dosovitskiy, P

    A. Dosovitskiy, P. Fischery, E. Ilg, P. Hausser, C. Hazirbas, V. Golkov, and T. Brox. FlowNet: Learning Optical Flow with Convolutional Networks. International Conference on Computer Vision, 2758-2766. 2015

    2015

  24. [32]

    D. Sun, X. Yang, M. Liu, and J. Kautz. PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume. Computer Vision and Pattern Recognition, 8934-8943, 2018

    2018

  25. [33]

    J. Wang, K. Chen, S. Yang, C. C. Loy, and D. Lin. Region Proposal by Guided Anchoring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2965-2974, 2019

    2019

  26. [34]

    T. T. Lin, M. Maire, S. J. Belongie, J. Hays, P. Perona, D. Ramanan, and C. L. Zitnick. Microsoft COCO: Common Objects in Context. European Conference on Computer Vision, 740-755, 2014

    2014

  27. [35]

    Simonyan, and A

    K. Simonyan, and A. Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv preprint arXiv:1409.1556, 2014

    2014 arXiv

  28. [36]

    Mayer, E

    N. Mayer, E. Ilg, P. Hausser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox. A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation. Computer Vision and Pattern Recognition, 4040-4048. 2016

    2016

Pith tools

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