Pith. sign in

REVIEW 3 major objections 5 minor 63 references

An Angular-Temporal Interaction Network for Light Field Object Tracking in Low-Light Scenes

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read By replacing raw light field views with an epipolar-plane structure image and gating attention to geometric structure points, ATINet achieves state-of-the-art low-light object tracking on the new R8LUT benchmark.

desk verdict A useful new task and ESI representation, but the dataset description needs fixing before the SOTA numbers can be trusted. read the letter →

arxiv 2507.21460 v2 pith:UKOFWYYW submitted 2025-07-29 cs.CV

classification cs.CV
keywords lightfieldobjecttrackinglow-lightscenesepipolarplaneimageangular-temporalmodelingself-supervisedlearningsinglemultiple
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 is trying to establish that light field object tracking in low light should be built on geometric structure, not on intensity appearance. It proposes a 2D epipolar-plane structure image (ESI) that records where light-ray directions change abruptly along the epipolar plane, arguing these points mark object contours even when brightness cues are weak. On top of the ESI, the paper introduces ATINet, whose geometry adaptive selection (GAS) module lets the network attend only to embeddings at geometric structure points, across both frames and within a frame, and a self-supervised loss that trains temporal feature interaction by masked reconstruction. The paper also introduces R8LUT, a new low-light light field video dataset with SOT and MOT annotations. On this benchmark ATINet reports the best single-object tracking results among 24 compared methods (success 0.64, precision 0.79, normalized precision 0.81), and its MOT extension AMTrack reports the highest MOTA (87.5) and IDF1 (85.4).

What carries the argument

The machinery is the ESI representation: a 2D image $S = \sqrt{S_H^2 + S_T^2}$, where $S_H$ and $S_T$ are first-order angular gradients of horizontal and vertical epipolar-plane images computed at the central view $u=U/2$, $v=V/2$ with step $d=1$. This turns the 4D light field's sparse geometric cues into a dense image that an ordinary CNN-transformer tracker can ingest, while discarding appearance redundancy. The second piece is GAS, a plug-in attention gate: a Gumbel-Softmax prediction module partitions patch embeddings into geometric and non-geometric groups, and a relation matrix restricts self- and cross-frame attention to the geometric group. The third piece is a self-supervised loss $L_M = \|D_\phi(E \odot M) - E \odot (1-M)\|^2$ that masks part of the structure embeddings and trains a lightweight decoder to reconstruct them from the other frame's embeddings, with spatial-attention dropout to push the network toward using inter-frame rather than intra-frame cues.

What would settle it

Compute the ESI from off-center angular views (for example $u = U/2 \pm k$ with $k > 1$, or $v = V/2 \pm k$) and run the same ATINet pipeline on R8LUT; if any off-center ESI variant matches or exceeds the central-view ESI on success or precision, the claim that central-view first-order gradients capture the needed structure is wrong. A more controlled test uses synthetic light fields in which two similar objects share identical central-view angular gradient profiles but differ only in off-center views, checking whether the tracker can still distinguish them.

Watch

Extended reading notes

Core claim

The paper's central claim is that the angular structure of a light field carries the target-discrimination signal that low-light RGB images lack. Concretely, it claims that geometric structure points — locations in the epipolar plane where the first-order angular gradient is maximal — can be collapsed into a single 2D ESI without losing the information needed for tracking, and that a tracker trained on this representation outperforms trackers trained on RGB views or on raw light field representations. The ablation studies support the claim by showing that ESI input alone barely helps a standard tracker, that the GAS attention gate accounts for most of the gain, and that the self-supervised temporal loss adds a further improvement, with all three components together yielding the state-of-the-art numbers.

Load-bearing premise

The representation assumes that the object boundaries needed for low-light tracking appear as first-order brightness changes along the angular dimension at the central viewpoint; if the decisive information lives in off-center angular views, the ESI discards it before the network sees the data.

Editorial extensions

If this is right

  • If the paper is right, low-light tracking no longer needs raw intensity across many viewpoints: a single 2D structure image carries enough angular-geometric information to localize and follow targets.
  • The GAS result implies that attention for light field tracking should be gated toward geometric structure points; dense self-attention over all patches actively hurts target-background discrimination.
  • The self-supervised loss shows that light field video features can be learned without dense annotation, which matters because light field video datasets are scarce.
  • Because the same ESI and GAS modules transfer to multiple object tracking through a re-identification branch, the recipe generalizes beyond single-target tracking to association.
  • The R8LUT benchmark gives future light field trackers a common low-light testbed with both SOT and MOT protocols.

Reading between the lines

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

  • Editorial inference: the ESI construction is not specific to tracking; the same angular-gradient collapse could be offered as an input representation for other low-light light field tasks such as depth estimation, saliency, or segmentation, but the paper does not test those.
  • Editorial inference: computing ESI from first-order gradients should make it inherently more invariant to global illumination shifts than raw intensity; a direct test would be to compare ATINet under synthetic brightness augmentation against a variant trained on intensity images.
  • Editorial inference: the central-view-only gradient ($u=U/2$, $v=V/2$, $d=1$) is the simplest possible choice; extending ESI to fuse gradients from multiple angular baselines might help on fast-moving or motion-blurred targets, at some cost in redundancy.
  • Editorial inference: since R8LUT was captured with a 5×5 angular-resolution light field camera, it is untested how the method's gains transfer to cameras with different angular grids; a natural extension is a cross-camera evaluation.
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

3 major / 5 minor

Summary. The paper proposes ATINet, a light field object tracking framework for low-light scenes, built on a new epipolar-plane structure image (ESI) representation that encodes geometric structure via angular gradients, plus a geometry adaptive selection (GAS) module and a self-supervised loss (SSL) for angular-temporal feature learning. The authors also introduce R8LUT, a new light field low-light video dataset for single object tracking (SOT) and multiple object tracking (MOT), and report state-of-the-art results on it: ATINet achieves 0.64 success, 0.79 precision, and 0.81 normalized precision on SOT, while AMTrack achieves 87.5 MOTA and 85.4 IDF1 on MOT. The paper includes ablations showing each proposed component contributes positively, and visualizations illustrating the claimed benefits.

Significance. If the results hold, the paper makes a useful contribution by introducing a compact angular representation (ESI) that reduces light field redundancy, and by demonstrating that explicit geometric-structure selection and self-supervised temporal learning can improve tracking in low-light conditions. The attempt to create a light field video tracking benchmark and to extend the method to MOT is also valuable for the community. The ablation studies are internally consistent and provide evidence that GAS and SSL each matter. However, the empirical claims rest entirely on a dataset whose description is internally inconsistent, and the ESI derivation contains mathematical errors that need correction. These issues currently prevent independent verification of the state-of-the-art claims.

major comments (3)
  1. [§IV-A] The dataset statistics are internally contradictory and prevent reproduction. The text states that after filtering, 3,500 high-quality light fields remain, forming 26 light field video sequences of about 150 frames each. It then says the MOT training subset contains 26 videos with 1,800 samples and the testing subset includes an equivalent number of videos, totaling 1,500 samples. If there are only 26 videos total, there cannot be 26 training videos plus an equivalent number of test videos. Furthermore, the SOT split is described as 102 training videos with 3,500 samples and 71 test videos with 15,380 samples, giving 173 videos and 18,880 samples total, which far exceeds the 3,500 retained light fields and the 26-video MOT collection. Since the SOT set is assembled from MOT objects, the paper must clarify the exact relationship between the MOT and SOT splits and rule out train/test leakage between the two tasks. As written, the reported SOTA numbers (Tables I and II) cannot be checked against a consistent benchmark.
  2. [§III-A, Eqs. (1)-(3)] The derivation of the ESI gradient is mathematically unclear and appears to contain typos. Eq. (1) defines L'_u as a weighted sum of three neighboring values plus a noise term, which is not a gradient. Eq. (2) then states a separate rational expression that is said to follow from Taylor series expansion, but the expression contains duplicated terms (L(u+dl) appears twice) and inconsistent signs (the last term uses L(u-dr) rather than L(u+dr)), and it does not reduce to the central difference in Eq. (3) when dl=dr=d. Please rewrite the derivation carefully, define all symbols, and confirm that Eq. (3) is the actual operational definition used to compute ESI. This matters because the entire method is built on this representation.
  3. [Tables I, II, III-VI] No error bars, confidence intervals, or significance tests are reported, and several performance differences are small (e.g., SOT success 0.64 vs. 0.63 for GRM in Table I; MOTA 87.5 vs. 87.3 for Trades in Table II). With single runs and no variance estimates, the claimed improvements could be run-to-run noise. Since the dataset split is also inconsistent, the empirical foundation for the SOTA claim is currently fragile. Please provide multiple-run statistics or at least report the variance across seeds, and ideally release the dataset splits and evaluation code.
minor comments (5)
  1. [§IV-A, implementation details] The text says 'the mask rate in ALS is set at 0.5'; this should presumably read 'SSL' rather than 'ALS'.
  2. [Throughout] Several typographical errors need correction: 'low-ligtht' in §IV-B.1, 'obejct' in §III-B, and 'V ondrick' in §II-C.
  3. [§IV-A] The phrase 'as illustrated in Fig. 3' refers to Figure 3, which depicts the GAS method, not the dataset construction pipeline; please use the correct figure reference.
  4. [Table V] Table V is referred to as 'Table 5' in the text; please standardize the table numbering and citations.
  5. [§IV-C.5] The sentence 'In the proposed GAS, GAS is embedded into all layer of neurons in the baseline model' is grammatically awkward; consider rephrasing to 'all layers of the baseline model'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; ESI, GAS, and SSL are derived from light field geometry and trained with standard losses, not from the tracking metrics they aim to predict.

full rationale

The paper's derivation chain is self-contained and does not reduce to its inputs. The ESI representation is explicitly constructed from first-order angular gradients of EPIs (Eqs. 1-4) with a fixed central-view choice and d=1; no parameter is fitted to tracking results and then relabeled as a prediction. The GAS module is a learned, trainable selector whose relation matrix and Gumbel-Softmax partition depend on embeddings and training losses, not on the reported success/precision/MOTA numbers. The SSL objective (Eq. 13) is a masked-reconstruction loss inspired by DropMAE and is used only during training; it is not a renamed evaluation metric. The authors' self-citations (e.g., [11], [18]) provide prior light field processing context and a baseline competitor (LFtrack), but they are not invoked as a load-bearing uniqueness theorem or as a substitute for the proposed derivation. The dataset is self-created, but evaluating a method on a new dataset is not circular reasoning. The internally inconsistent train/test sample counts in Section IV-A are a serious reproducibility and verifiability concern, and the small performance margins over strong baselines deserve scrutiny, but those are correctness/benchmark-integrity issues, not circularity of the derivation. No equation or fitted parameter is equivalent by construction to the claimed tracking outcome.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The only new elements are computational constructs (ESI, GAS grouping, SSL decoder), which are not independent entities in the sense of new postulated physics. The axioms above list the assumptions those constructs depend on, and the free parameters are hand-set design choices that the paper does not justify with sensitivity studies.

free parameters (6)
  • Angular gradient step d = 1
    Set by hand in Eq. (3); controls the angular neighborhood for detecting abrupt ray changes, with no search or sensitivity analysis.
  • Central angular index u=U/2 = U/2
    Only the central horizontal angular view is used to build S_H in Eq. (4); this choice is not compared against learned or aggregated angular views.
  • Central angular index v=V/2 = V/2
    Only the central vertical angular view is used to build S_T in Eq. (4); same limitation as u=U/2.
  • SSL mask ratio = 0.5
    Mask ratio for the self-supervised loss is set to 0.5 in the implementation details, with no sensitivity study.
  • Loss weights lambda_1, lambda_2, lambda_3 = not reported
    Total losses in Eqs. (14)-(15) depend on weighting factors, but their values are never given, leaving the training objective underspecified.
  • Gumbel-Softmax temperature = not reported
    The binary partition P in Eq. (10) uses Gumbel-Softmax, but the temperature is not specified; it controls the hardness of the geometric/non-geometric split.
assumptions (4)
  • standard math Taylor series central difference approximates the angular derivative of EPI brightness.
    Used in Eqs. (1)-(3) to define geometric structure points as maxima of the first-order angular gradient.
  • domain assumption Abrupt changes in light ray directions at object boundaries are captured by angular gradient maxima and delineate target contours in low light.
    Section III-A defines the ESI on this premise; the paper provides qualitative figures but no quantitative validation that ESI corresponds to true object boundaries.
  • ad hoc to paper A learned binary partition of patch embeddings into geometric and non-geometric groups (GAS) selects the structure points relevant for tracking.
    Eqs. (10)-(12); the paper assumes an MLP with Gumbel-Softmax can learn a meaningful geometry split, but no analysis of the learned partitions is given.
  • domain assumption Masked reconstruction with inter-frame attention dropout encourages temporal motion cues.
    Section III-C, Eq. (13), borrowing from DropMAE [40]; the assumption is that reconstructing masked ESI tokens from the other frame transfers to tracking, without a dedicated analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Angular-Temporal Interaction Network for Light Field Object Tracking in Low-Light Scenes." pith.science (2026). https://pith.science/paper/UKOFWYYW

@misc{pith2026250721460,
  author       = {Pith},
  title        = {Pith review of: An Angular-Temporal Interaction Network for Light Field Object Tracking in Low-Light Scenes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UKOFWYYW}},
  note         = {Machine review of arXiv:2507.21460}
}
read the original abstract

High-quality 4D light field representation with efficient angular feature modeling is crucial for scene perception, as it can provide discriminative spatial-angular cues to identify moving targets. However, recent developments still struggle to deliver reliable angular modeling in the temporal domain, particularly in complex low-light scenes. In this paper, we propose a novel light field epipolar-plane structure image (ESI) representation that explicitly defines the geometric structure within the light field. By capitalizing on the abrupt changes in the angles of light rays within the epipolar plane, this representation can enhance visual expression in low-light scenes and reduce redundancy in high-dimensional light fields. We further propose an angular-temporal interaction network (ATINet) for light field object tracking that learns angular-aware representations from the geometric structural cues and angular-temporal interaction cues of light fields. Furthermore, ATINet can also be optimized in a self-supervised manner to enhance the geometric feature interaction across the temporal domain. Finally, we introduce a large-scale light field low-light dataset for object tracking. Extensive experimentation demonstrates that ATINet achieves state-of-the-art performance in single object tracking. Furthermore, we extend the proposed method to multiple object tracking, which also shows the effectiveness of high-quality light field angular-temporal modeling.

Figures

Figures reproduced from arXiv: 2507.21460 by the authors.

Figure 1
Figure 1. Illustrations of light field representations: (a) Light field imaging system. (b) Light field multi-view and EPI representation. (c) Changes in viewpoints [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An overview of our ATINet framework, which is a hybrid architecture tracker that employs Depthwise cross correlation to localize targets. It uses [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. An illustration of the geometry adaptive selection method. It is a [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Visual comparison between our proposed tracker and other methods in the low-light scenes. The black box indicates the ground truth, the red box [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visual tracking results of the proposed light field MOT algorithm on some scenes. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Visual results of different light field representations. (a) [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Visualization of feature maps following Siamese network. (a) and (d) [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Visualization of tracking score maps across various frames. (a) Search [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 49 canonical work pages

  1. [2]

    Light field reconstruction via deep adaptive fusion of hybrid lenses,

    J. Jin, M. Guo, J. Hou, H. Liu, and H. Xiong, “Light field reconstruction via deep adaptive fusion of hybrid lenses,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

  2. [3]

    Light-field depth estimation via epipolar plane image analysis and locally linear embedding,

    Y . Zhang, H. Lv, Y . Liu, H. Wang, X. Wang, Q. Huang, X. Xiang, and Q. Dai, “Light-field depth estimation via epipolar plane image analysis and locally linear embedding,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 27, no. 4, pp. 739–747, 2016

  3. [4]

    Autoregressive visual tracking,

    X. Wei, Y . Bai, Y . Zheng, D. Shi, and Y . Gong, “Autoregressive visual tracking,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9697–9706

  4. [5]

    Joint feature learning and relation modeling for tracking: A one-stream framework,

    B. Ye, H. Chang, B. Ma, S. Shan, and X. Chen, “Joint feature learning and relation modeling for tracking: A one-stream framework,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 341– 357

  5. [6]

    Atom: Accurate tracking by overlap maximization,

    M. Danelljan, G. Bhat, F. S. Khan, and M. Felsberg, “Atom: Accurate tracking by overlap maximization,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 4660– 4669

  6. [7]

    Masked autoencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked autoencoders are scalable vision learners,” in2022 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 15 979–15 988

  7. [8]

    Deep light field spatial super-resolution using heterogeneous imaging,

    Y . Chen, G. Jiang, M. Yu, H. Xu, and Y .-S. Ho, “Deep light field spatial super-resolution using heterogeneous imaging,”IEEE Transactions on Visualization and Computer Graphics, 2022

  8. [9]

    Disen- tangling light fields for super-resolution and disparity estimation,

    Y . Wang, L. Wang, G. Wu, J. Yang, W. An, J. Yu, and Y . Guo, “Disen- tangling light fields for super-resolution and disparity estimation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 1, pp. 425–443, 2022

Show all 63 references
  1. [10]

    Accurate depth and nor- mal maps from occlusion-aware focal stack symmetry,

    M. Strecke, A. Alperovich, and B. Goldluecke, “Accurate depth and nor- mal maps from occlusion-aware focal stack symmetry,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 2814–2822

  2. [11]

    Lfbcnet: Light field boundary-aware and cascaded interaction network for salient object detection,

    M. Wang, F. Shi, X. Cheng, M. Zhao, Y . Zhang, C. Jia, W. Tian, and S. Chen, “Lfbcnet: Light field boundary-aware and cascaded interaction network for salient object detection,” inProceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 3430–3439

  3. [12]

    Combining implicit-explicit view correlation for light field semantic segmentation,

    R. Cong, D. Yang, R. Chen, S. Wang, Z. Cui, and H. Sheng, “Combining implicit-explicit view correlation for light field semantic segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 9172–9181

  4. [13]

    Opal: Occlusion pattern aware loss for unsupervised light field disparity estimation,

    P. Li, J. Zhao, J. Wu, C. Deng, Y . Han, H. Wang, and T. Yu, “Opal: Occlusion pattern aware loss for unsupervised light field disparity estimation,”IEEE Transactions on Pattern Analysis and Machine In- telligence, 2023

  5. [14]

    Beyond photometric consistency: Geometry-based occlusion-aware unsupervised light field disparity estimation,

    W. Zhou, L. Lin, Y . Hong, Q. Li, X. Shen, and E. E. Kuruoglu, “Beyond photometric consistency: Geometry-based occlusion-aware unsupervised light field disparity estimation,”IEEE Transactions on Neural Networks and Learning Systems, 2023

  6. [15]

    A novel occlusion-aware vote cost for light field depth estimation,

    K. Han, W. Xiang, E. Wang, and T. Huang, “A novel occlusion-aware vote cost for light field depth estimation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 8022–8035, 2021

  7. [16]

    Shape estimation from shading, defocus, and correspondence using light-field angular coherence,

    M. W. T. P. P. Srinivasan, S. H. S. Rusinkiewicz, and J. M. R. Ramamoor- thi, “Shape estimation from shading, defocus, and correspondence using light-field angular coherence,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 3, 2017. JOURNAL OF LATEX...

  8. [17]

    Light field saliency detection with deep convolutional networks,

    J. Zhang, Y . Liu, S. Zhang, R. Poppe, and M. Wang, “Light field saliency detection with deep convolutional networks,”IEEE Transactions on Image Processing, vol. 29, pp. 4421–4434, 2020

  9. [18]

    Visual object tracking based on light-field imaging in the presence of similar distractors,

    M. Wang, F. Shi, X. Cheng, M. Zhao, Y . Zhang, C. Jia, W. Tian, and S. Chen, “Visual object tracking based on light-field imaging in the presence of similar distractors,”IEEE Transactions on Industrial Informatics, vol. 19, no. 3, pp. 2705–2716, 2022

  10. [19]

    Tracking objects as points,

    X. Zhou, V . Koltun, and P. Kr ¨ahenb¨uhl, “Tracking objects as points,” in European conference on computer vision. Springer, 2020, pp. 474–490

  11. [20]

    Mixformer: End-to-end tracking with iterative mixed attention,

    Y . Cui, C. Jiang, L. Wang, and G. Wu, “Mixformer: End-to-end tracking with iterative mixed attention,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 13 608–13 618

  12. [21]

    Track to detect and segment: An online multi-object tracker,

    J. Wu, J. Cao, L. Song, Y . Wang, M. Yang, and J. Yuan, “Track to detect and segment: An online multi-object tracker,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 12 352–12 361

  13. [22]

    Transforming model prediction for tracking,

    C. Mayer, M. Danelljan, G. Bhat, M. Paul, D. P. Paudel, F. Yu, and L. Van Gool, “Transforming model prediction for tracking,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8731–8740

  14. [23]

    Bytetrack: Multi-object tracking by associating every detection box,

    Y . Zhang, P. Sun, Y . Jiang, D. Yu, F. Weng, Z. Yuan, P. Luo, W. Liu, and X. Wang, “Bytetrack: Multi-object tracking by associating every detection box,” inEuropean conference on computer vision. Springer, 2022, pp. 1–21

  15. [24]

    Geometry auxiliary salient object detection for light fields via graph neural networks,

    Q. Zhang, S. Wang, X. Wang, Z. Sun, S. Kwong, and J. Jiang, “Geometry auxiliary salient object detection for light fields via graph neural networks,”IEEE Transactions on Image Processing, vol. 30, pp. 7578–7592, 2021

  16. [25]

    Eco: Efficient convolution operators for tracking,

    M. Danelljan, G. Bhat, F. Shahbaz Khan, and M. Felsberg, “Eco: Efficient convolution operators for tracking,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 6638– 6646

  17. [26]

    Probabilistic regression for visual tracking,

    M. Danelljan, L. V . Gool, and R. Timofte, “Probabilistic regression for visual tracking,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 7183–7192

  18. [27]

    Learning discrim- inative model prediction for tracking,

    G. Bhat, M. Danelljan, L. V . Gool, and R. Timofte, “Learning discrim- inative model prediction for tracking,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 6182–6191

  19. [28]

    Fully-convolutional siamese networks for object tracking,

    L. Bertinetto, J. Valmadre, J. F. Henriques, A. Vedaldi, and P. H. Torr, “Fully-convolutional siamese networks for object tracking,” inComputer Vision–ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part II 14. Springer, 2016, pp. 850–865

  20. [30]

    Learning the model update for siamese trackers,

    L. Zhang, A. Gonzalez-Garcia, J. V . D. Weijer, M. Danelljan, and F. S. Khan, “Learning the model update for siamese trackers,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 4010–4019

  21. [31]

    Unsuper- vised deep representation learning for real-time tracking,

    N. Wang, W. Zhou, Y . Song, C. Ma, W. Liu, and H. Li, “Unsuper- vised deep representation learning for real-time tracking,”International Journal of Computer Vision, vol. 129, pp. 400–418, 2021

  22. [32]

    Swintrack: A simple and strong baseline for transformer tracking,

    L. Lin, H. Fan, Z. Zhang, Y . Xu, and H. Ling, “Swintrack: A simple and strong baseline for transformer tracking,”Advances in Neural Information Processing Systems, vol. 35, pp. 16 743–16 754, 2022

  23. [33]

    Self-supervised video representation learning by uncovering spatio-temporal statistics,

    J. Wang, J. Jiao, L. Bao, S. He, W. Liu, and Y .-h. Liu, “Self-supervised video representation learning by uncovering spatio-temporal statistics,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 7, pp. 3791–3806, 2022

  24. [34]

    Self-supervised visual feature learning with deep neural networks: A survey,

    L. Jing and Y . Tian, “Self-supervised visual feature learning with deep neural networks: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 11, pp. 4037–4058, 2021

  25. [35]

    Tracking emerges by colorizing videos,

    C. V ondrick, A. Shrivastava, A. Fathi, S. Guadarrama, and K. Murphy, “Tracking emerges by colorizing videos,” inProceedings of the Euro- pean conference on computer vision (ECCV), 2018, pp. 391–408

  26. [36]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” inNorth American Chapter of the Association for Computational Linguistics,

  27. [37]

    Masked motion encoding for self-supervised video representation learning,

    X. Sun, P. Chen, L. Chen, C. Li, T. H. Li, M. Tan, and C. Gan, “Masked motion encoding for self-supervised video representation learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 2235–2245

  28. [38]

    Siamese masked autoen- coders,

    A. Gupta, J. Wu, J. Deng, and L. Fei-Fei, “Siamese masked autoen- coders,”arXiv preprint arXiv:2305.14344, 2023

  29. [39]

    Explicit visual prompting for low-level structure segmentations,

    W. Liu, X. Shen, C.-M. Pun, and X. Cun, “Explicit visual prompting for low-level structure segmentations,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 19 434–19 445

  30. [40]

    Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks,

    Q. Wu, T. Yang, Z. Liu, B. Wu, Y . Shan, and A. B. Chan, “Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 14 561–14 571

  31. [41]

    Raytrix gmbh

    “Raytrix gmbh.” https://raytrix.de/.Accessed:2019-02-03

  32. [42]

    Mot16: A benchmark for multi-object tracking,

    A. Milan, L. Leal-Taix ´e, I. Reid, S. Roth, and K. Schindler, “Mot16: A benchmark for multi-object tracking,”arXiv preprint arXiv:1603.00831, 2016

  33. [43]

    Online object tracking: A benchmark,

    Y . Wu, J. Lim, and M. Yang, “Online object tracking: A benchmark,” in2013 IEEE Conference on Computer Vision and Pattern Recognition, 2013, pp. 2411–2418

  34. [44]

    Evaluating multiple object tracking performance: the clear mot metrics,

    K. Bernardin and R. Stiefelhagen, “Evaluating multiple object tracking performance: the clear mot metrics,”EURASIP Journal on Image and Video Processing, vol. 2008, pp. 1–10, 2008

  35. [45]

    Performance measures and a data set for multi-target, multi-camera tracking,

    E. Ristani, F. Solera, R. Zou, R. Cucchiara, and C. Tomasi, “Performance measures and a data set for multi-target, multi-camera tracking,” in European conference on computer vision. Springer, 2016, pp. 17–35

  36. [46]

    High performance visual tracking with siamese region proposal network,

    B. Li, J. Yan, W. Wu, Z. Zhu, and X. Hu, “High performance visual tracking with siamese region proposal network,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 8971–8980

  37. [47]

    Siamrpn++: Evolution of siamese visual tracking with very deep networks,

    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/CVF conference on computer vision and pattern recognition, 2019, pp. 4282–4291

  38. [48]

    Deeper and wider siamese networks for real- time visual tracking,

    Z. Zhang and H. Peng, “Deeper and wider siamese networks for real- time visual tracking,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 4591–4600

  39. [49]

    Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines,

    Y . Xu, Z. Wang, Z. Li, Y . Yuan, and G. Yu, “Siamfc++: Towards robust and accurate visual tracking with target estimation guidelines,” in Proceedings of the AAAI conference on artificial intelligence, vol. 34, no. 07, 2020, pp. 12 549–12 556

  40. [50]

    Siamcar: Siamese fully convolutional classification and regression for visual tracking,

    D. Guo, J. Wang, Y . Cui, Z. Wang, and S. Chen, “Siamcar: Siamese fully convolutional classification and regression for visual tracking,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 6269–6277

  41. [51]

    Siamese anchor proposal network for high-speed aerial tracking,

    C. Fu, Z. Cao, Y . Li, J. Ye, and C. Feng, “Siamese anchor proposal network for high-speed aerial tracking,” in2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 510–516

  42. [52]

    Stmtrack: Template-free visual tracking with space-time memory networks,

    Z. Fu, Q. Liu, Z. Fu, and Y . Wang, “Stmtrack: Template-free visual tracking with space-time memory networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 13 774–13 783

  43. [53]

    Trtr: Visual tracking with trans- former,

    M. Zhao, K. Okada, and M. Inaba, “Trtr: Visual tracking with trans- former,”arXiv preprint arXiv:2105.03817, 2021

  44. [54]

    Beyond sot: Tracking multiple generic objects at once,

    C. Mayer, M. Danelljan, M.-H. Yang, V . Ferrari, L. Van Gool, and A. Kuznetsova, “Beyond sot: Tracking multiple generic objects at once,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 6826–6836

  45. [55]

    Exploring enhanced contextual information for video-level object tracking,

    B. Kang, X. Chen, S. Lai, Y . Liu, Y . Liu, and D. Wang, “Exploring enhanced contextual information for video-level object tracking,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 4, 2025, pp. 4194–4202

  46. [56]

    Learning spatio-temporal transformer for visual tracking,

    B. Yan, H. Peng, J. Fu, D. Wang, and H. Lu, “Learning spatio-temporal transformer for visual tracking,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 448–10 457

  47. [57]

    Backbone is all your need: A simplified architecture for visual object tracking,

    B. Chen, P. Li, L. Bai, L. Qiao, Q. Shen, B. Li, W. Gan, W. Wu, and W. Ouyang, “Backbone is all your need: A simplified architecture for visual object tracking,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 375–392

  48. [58]

    Mixformerv2: Efficient fully transformer tracking,

    Y . Cui, T. Song, G. Wu, and L. Wang, “Mixformerv2: Efficient fully transformer tracking,”Advances in Neural Information Processing Sys- tems, vol. 36, 2024

  49. [59]

    Zoom- track: Target-aware non-uniform resizing for efficient visual tracking,

    Y . Kou, J. Gao, B. Li, G. Wang, W. Hu, Y . Wang, and L. Li, “Zoom- track: Target-aware non-uniform resizing for efficient visual tracking,” Advances in Neural Information Processing Systems, vol. 36, 2024

  50. [60]

    Generalized relation modeling for transformer tracking,

    S. Gao, C. Zhou, and J. Zhang, “Generalized relation modeling for transformer tracking,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 18 686–18 695. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 12

  51. [61]

    Odtrack: Online dense temporal token learning for visual tracking,

    Y . Zheng, B. Zhong, Q. Liang, Z. Mo, S. Zhang, and X. Li, “Odtrack: Online dense temporal token learning for visual tracking,” inAAAI, 2024

  52. [62]

    Su- track: Towards simple and unified single object tracking,

    X. Chen, B. Kang, W. Geng, J. Zhu, Y . Liu, D. Wang, and H. Lu, “Su- track: Towards simple and unified single object tracking,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 2, 2025, pp. 2239–2247

  53. [63]

    Hybrid-sort: Weak cues matter for online multi-object tracking,

    M. Yang, G. Han, B. Yan, W. Zhang, J. Qi, H. Lu, and D. Wang, “Hybrid-sort: Weak cues matter for online multi-object tracking,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 7, 2024, pp. 6504–6512

  54. [64]

    Sparsetrack: Multi- object tracking by performing scene decomposition based on pseudo- depth,

    Z. Liu, X. Wang, C. Wang, W. Liu, and X. Bai, “Sparsetrack: Multi- object tracking by performing scene decomposition based on pseudo- depth,”IEEE Transactions on Circuits and Systems for Video Technol- ogy, 2025

  55. [2019]

    Available: https://api.semanticscholar.org/CorpusID: 52967399

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 52967399

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.