Pith. sign in

REVIEW 5 major objections 5 minor 75 references

TrackingMiM: Efficient Mamba-in-Mamba Serialization for Real-time UAV Object Tracking

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

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

Pith's one-line read A Mamba-in-Mamba serialization is argued to deliver top UAV tracking accuracy at 268 FPS.

desk verdict A plausible lightweight Mamba tracker with a transferable retrieval module, but the internal numbers disagree where it matters most, so the SOTA claim is unverifiable. read the letter →

arxiv 2507.01535 v1 pith:XIHSIEQB submitted 2025-07-02 cs.CV

Bingxi Liu , Calvin Chen , Junhao Li , Guyang Yu , Haoqian Song , Xuchen Liu , Jinqiang Cui , Hong Zhang This is my paper
classification cs.CV
keywords UAVtrackingstatespacemodelsMambavisualobjecttemporalmodelingretrieval-augmentedreal-timeefficientserialization
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 aims to show that a state-space model can replace quadratic self-attention in UAV object tracking without losing accuracy, by serializing video frames in a nested spatial-temporal scan. It claims TrackingMiM reaches 86.3 precision and 66.1 success averaged over five UAV benchmarks while running at 268.3 FPS on a GPU and above 95 FPS on CPU. A second claim is that its retrieval attention module is a plug-and-play component that raises precision by at least one point in six existing trackers. These results matter because drone tracking must balance accuracy against tight power, memory, and latency budgets.

What carries the argument

The load-bearing object is the Mamba-in-Mamba block, a nested serialization of the selective state-space scan: an inner Mamba scans spatial patches inside each frame with a window-swing shifter, and an outer Mamba scans corresponding patches across time with forward and backward passes that begin from the template frame. Between the two scans sits the retrieval-augmented tracking module: a growing memory corpus of historical object features, a top-K cosine-similarity retriever, mean fusion, and cross-attention in which the fused feature acts as the query. The two-level design is what converts a single-sequence scanner into a spatiotemporal model at linear rather than quadratic cost.

What would settle it

Reimplement TrackingMiM from the text alone, replacing the temporal update in Eq. (4) with a simple per-frame spatial scan; if the +2.0 precision gain persists, the temporal mechanism is not the cause, and if it vanishes, the paper's central comparison depends on an operator that the manuscript never specifies.

Watch

Extended reading notes

Core claim

TrackingMiM's central claim is that temporally consistent Mamba scanning, not the quadratic attention of ViTs, is what UAV trackers need: it orders patches so that an inner Mamba learns within-frame spatial structure and an outer Mamba learns across-frame continuity at each spatial location, always starting from the template frame. On top of this, a retrieval-augmented tracking module stores historical object features in a memory corpus, retrieves the top-K most similar to the current crop, averages them, and uses the result as the query in a cross-attention layer inserted between the spatial and temporal scans. The authors report that this design reaches 86.3 precision and 66.1 success averaged over DTB70, UAVDT, VisDrone2018, UAV123, and UAV123@10fps, at 268.3 FPS on GPU and 97.2 FPS on CPU, surpassing all 25 compared lightweight trackers. Ablations attribute +2.0 precision to temporal scanning, +2.9 to retrieval attention, and +4.0 to their combination, and the same attention module improves six existing CNN and ViT trackers by at least +1.0 precision at roughly a 3 percent FPS cost.

Load-bearing premise

The temporal scan is never actually defined, and its update equation refers to itself, so the claimed accuracy jump from temporal scanning rests on an unstated implementation that a reimplementation might not reproduce.

Editorial extensions

If this is right

  • If the reported numbers hold, UAV trackers no longer need quadratic self-attention to reach top precision: a linear-cost Mamba architecture matches or beats the strongest lightweight ViT baselines on all five benchmarks.
  • The retrieval attention module can be inserted into existing CNN- and ViT-based trackers with an expected gain of at least +1.0 precision and +0.9 success, at roughly 3% lower FPS.
  • Temporal serialization scanning is worth about +2.0 precision and +1.5 success on its own, meaning frame order and scan continuity are substantive cues rather than a minor implementation detail.
  • The measured speed implies real-time operation on CPU-only hardware (97.2 FPS), which would enable deployment without a discrete GPU.
  • Increasing the temporal window beyond eight frames gives diminishing returns, so the temporal context needed for these benchmarks saturates quickly.

Reading between the lines

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

  • If the temporal scan is reproducible, the same nested serialization should transfer to other video tasks, such as action recognition or video object segmentation, where linear-cost cross-frame state is the bottleneck.
  • The plug-and-play gains suggest retrieval attention is a general conditioning mechanism; testing it on ground-vehicle or generic short-term trackers would reveal whether the improvement is specific to aerial imagery.
  • The text's tokenization details are inconsistent (patch size 36 versus a 16x16 token grid), so a code release resolving that discrepancy is the direct test of reproducibility.
  • The self-referential form of Eq. (4) means the temporal scan's exact behavior cannot be re-derived from the text alone; an explicit scan-operator definition would settle what mechanism deserves the +2.0 precision credit.
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

5 major / 5 minor

Summary. The paper proposes TrackingMiM, a Mamba-in-Mamba architecture for real-time UAV object tracking. The model performs nested spatial and temporal Mamba scans over patch tokens, uses template-first spatial scanning, and inserts a query-based retrieval-augmented attention module between spatial and temporal stages. The authors report state-of-the-art average precision of 86.3 and success of 66.1 on five UAV benchmarks at 268.3 FPS, and they further claim that the retrieval attention module is plug-and-play, improving six existing trackers by at least +1.0 precision and +0.9 success. Extensive ablations examine patch size, layer depth, temporal window, retrieval top-K, fusion strategy, and feature injection mechanism. The central claim is that TrackingMiM achieves the best accuracy-efficiency trade-off among lightweight UAV trackers.

Significance. If the reported results are correct and reproducible, the paper would offer a meaningful efficiency-accuracy contribution to real-time UAV tracking: the nested Mamba design is a plausible way to combine spatial and temporal modeling at linear complexity, and the retrieval-augmented attention module is a general idea worth testing across trackers. The paper provides a broad comparison against 25 trackers on five benchmarks, and the ablation tables are internally structured enough to show monotone contributions from the individual components. The plug-and-play experiments, if trustworthy, would be a useful practical result. However, the significance is currently undermined by the manuscript's inconsistent reporting of its own headline numbers, by an incoherent formal description of the temporal scanning mechanism, and by hyperparameter selection performed on the same test benchmarks used for the final evaluation. These issues must be resolved before the reported state-of-the-art claim can be verified or transferred to a reimplementation.

major comments (5)
  1. [§I, §IV-C, Table I, Table III, Fig. 3] The headline results are internally inconsistent, and this makes the central state-of-the-art claim unverifiable. Section IV-C and Table I report the final TrackingMiM as 86.3 precision / 66.1 success at 268.3 FPS, while the ablation tables in Table III(a)-(f), which identify the adopted final configuration (patch size 36, depth 24, window 8, K=7, K-retrieval mean fusion, query attention), report 85.8 precision / 65.1 success at the same 268.3 FPS. Section IV-C additionally reports an average Precision AUC of 0.850 and Success AUC of 0.666, a third summary. Furthermore, the per-dataset success values in Table I for the final model (67.8, 62.4, 66.2, 68.0, 67.1) average to 66.3, not 66.1, and the Fig. 1 caption gives a UAV123 precision of 86.6 while Table I lists 87.1 for that row. The authors must reconcile these numbers and provide a single, clearly labeled table of per-dataset results for the final configuration; until then the claimed SOTA margin over Aba-ViTrack (85.4/64.9) is not determinable.
  2. [§III-C, Eq. (4)] Equation (4) does not define a meaningful temporal sampling operation and is self-referential. As written, P_t^i appears on both sides of the equality, so the equation is tautological unless some other quantity is intended. The term dP_t/dt is also not a derivative: the text defines dP_t = P_t^i - P_{t}^{i-1} as a 'patch residual', and no continuous-time interpolation is given. The sum index i is reused for both the time summation and the layer/iteration index, and the scan operator is never formally defined in terms of the Mamba hidden-state recurrence. Because the temporal scan is one of the two main claimed contributions and the ablation attributes a +2.0 precision gain to it, the method cannot be reimplemented from this description. Please replace Eq. (4) with a precise definition of the temporal Mamba scan, specifying input/output shapes, indexing, and the recurrence used.
  3. [§III-C (Tokenisation) vs. §IV-A (Implementation Details)] The tokenization description contradicts the implementation details. Section III-C states that a 2D patchify convolution with kernel size K=36 is applied to each frame, producing patches P_{t,s} in R^{C×K×K}, while Section IV-A states that the input is tokenized with 'spatial token resolution of 16×16' and a temporal stride of 2 with temporal length 8. For the stated 256×256 search and template frames, a 36×36 kernel yields roughly a 7×7 spatial grid, not a 16×16 token grid. The term '3D Patch Tokenisation' is also not described as a 3D operation; only a per-frame 2D convolution is presented. Please clarify the actual patch size, stride, and number of tokens per frame, and align the notation in Section III-C with the configuration used in the experiments.
  4. [§IV-A, §IV-E, Table III] Hyperparameters are selected on the same five test benchmarks used for the final evaluation, which inflates the reported performance and makes the gains difficult to interpret. In Table III(a)-(f), the patch size, layer depth, temporal window, retrieval top-K, fusion strategy, and feature injection mechanism are all chosen by comparing average precision and success over the test sets (DTB70, UAVDT, VisDrone, UAV123, UAV123@10fps); the text then reports those chosen values as the final model. The similarity threshold τ=0.8 and the bbox enlargement factor 1.1 are also fixed without a validation-set analysis. Please either hold out a validation set for these choices or clearly state that the reported numbers are selected-on-test and therefore likely optimistic, and provide the corresponding selection procedure.
  5. [§IV-E, Table II] The plug-and-play experiments in Table II lack the training and evaluation protocol needed to support the +1.0 to +2.5 precision gains. For TCTrack, UDAT, DRCI, Aba-ViTrack, LiteTrack, and LightFC, the paper reports improved precision/success after adding the tracking-attention module, but it does not state whether the base trackers are frozen, how the inserted attention layers are initialized, what learning rate and number of epochs are used, or whether the models are retrained on the same data splits. Without this information the reader cannot distinguish a genuine architectural improvement from additional fine-tuning. Please specify the training protocol for the inserted modules, or, if the gains are obtained at inference time only, explain how the newly introduced projection matrices W_Q, W_K, W_V are obtained.
minor comments (5)
  1. [Abstract / §V] The phrase 'noticeable higher speed' is vague; please report the actual wall-clock speedup relative to the strongest comparable ViT-based tracker rather than an informal claim.
  2. [§III-A] The matrix dimensions in Eq. (1) are nonstandard: B ∈ R^{N×L}, C ∈ R^{L×N}, and D ∈ R^{L×L} do not align with the usual Mamba/SSM convention where B ∈ R^{N×1}, C ∈ R^{1×N}, and D is a scalar; please correct or clarify the notation.
  3. [References] Reference [22] and reference [55] are the same paper (Gu and Dao, 'Mamba: Linear-time sequence modeling with selective state spaces') and should be merged or cross-referenced consistently.
  4. [Throughout] There are several textual inconsistencies in cross-references to tables: Section IV-E refers to 'Tab. IVa' through 'Tab. IVf' while the actual sub-tables are in Table III. Please correct all such references.
  5. [§V] The conclusion repeats the exact numbers 86.3/66.1 and 268.3 FPS without acknowledging the discrepancies discussed in the major comments; the final summary should be updated to match the reconciled results.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SOTA and ablation claims are direct benchmark measurements, not derivations from fitted constants.

full rationale

TrackingMiM is an empirical architecture paper. Its load-bearing claims—SOTA precision/success on five UAV benchmarks and the plug-and-play gains in Table II—are reported measurements of trained models, not quantities derived from a fitted equation. The method section introduces update rules and scanning schemes (Eqs. 2–8), but none of these equations is calibrated against the benchmark averages, so there is no self-definitional or fitted-input-called-prediction step. Although Eq. (4) is written with Pt_i on both sides and the scan operator is under-specified, this is a formula-definition/implementation ambiguity, not a reduction of a claimed derivation to its own output. The only author-overlapping citation, MiM-ISTD [40], is presented as related work introducing nested Mamba scanning; it is not invoked as the sole justification for the design, and the paper supplies its own ablations. The selection of K=7, patch size 36, and window size 8 on the evaluation benchmarks (Table III) is a model-selection concern that can inflate reported numbers, but it does not reduce the central result to its inputs by construction and therefore falls below the circularity bar. Internal inconsistencies in the reported averages (86.3/66.1 vs. 85.8/65.1 vs. 85.0/66.6) are verifiability/correctness issues rather than circularity.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard Mamba background, a set of architecture hyperparameters chosen on the five evaluation benchmarks, and several unproven design assumptions about temporal scanning and memory retrieval. No new physical entities are introduced; the memory corpus, template token, and query feature are internal neural network components with no independent physical handle.

free parameters (8)
  • patch_size = 36 pixels
    Table IIIa selects 36 as the best speed-accuracy balance after testing 24, 36, and 48. This is a validation choice, not dictated by the task.
  • layer_depth = 24 blocks
    Table IIIb selects 24; depth 36 improves accuracy but drops FPS by 116.1, so the choice is a speed-accuracy tradeoff.
  • temporal_window = 8 frames
    Table IIIc selects 8; window 16 gives only +0.3 precision with a large FPS drop.
  • retrieval_top_k = 7
    Table IIId selects K=7 after testing 3, 5, 7, and 9 on the five benchmarks.
  • memory_similarity_threshold_tau = 0.8
    Equation (5) sets tau=0.8 for memory update; no ablation is reported for this value.
  • bbox_enlarge_factor = 1.1
    Section III-D enlarges the previous bounding box by a factor of 1.1 before cropping; hand-set without ablation.
  • temporal_stride = 2
    Section IV-A tokenizes with a temporal stride of 2 and temporal length 8; no ablation is reported for the stride.
  • hidden_state_dim = 384
    A 'medium-scale' Mamba configuration with hidden state dimension 384 is adopted; no ablation justifies this size.
assumptions (6)
  • standard math The Mamba/SSM discretization and convolutional kernel representation (Eq. 1 and surrounding text) are correct background.
    Section III-A imports the SSM equations from ref. [22]. This is standard background, not derived anew.
  • domain assumption Temporal inconsistency in the Mamba scanning mechanism is the key failure of existing Mamba-based trackers.
    The introduction states this as the motivating insight without a supporting study or quantitative evidence.
  • domain assumption Nested inner spatial and outer temporal Mamba scans improve tracking accuracy.
    Supported only by the paper's own ablations, not by an independent baseline or theory.
  • ad hoc to paper Equation (4) defines a meaningful temporal sampling operation.
    The equation is self-referential and dimensionally unclear; no derivation or pseudocode is provided, yet it is presented as the formal basis of the temporal scan.
  • domain assumption The training set and protocol 'aligned with Aba-ViTrack [8]' transfer to the five evaluation benchmarks.
    Section IV-A does not specify the training data or splits, so the benchmark results assume this alignment and transfer are valid.
  • domain assumption Top-K memory retrieval with cosine similarity is a valid mechanism for re-identification.
    Section III-D motivates retrieval by analogy to RAG; effectiveness is demonstrated only by the paper's ablations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TrackingMiM: Efficient Mamba-in-Mamba Serialization for Real-time UAV Object Tracking." pith.science (2026). https://pith.science/paper/XIHSIEQB

@misc{pith2026250701535,
  author       = {Pith},
  title        = {Pith review of: TrackingMiM: Efficient Mamba-in-Mamba Serialization for Real-time UAV Object Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XIHSIEQB}},
  note         = {Machine review of arXiv:2507.01535}
}
read the original abstract

The Vision Transformer (ViT) model has long struggled with the challenge of quadratic complexity, a limitation that becomes especially critical in unmanned aerial vehicle (UAV) tracking systems, where data must be processed in real time. In this study, we explore the recently proposed State-Space Model, Mamba, leveraging its computational efficiency and capability for long-sequence modeling to effectively process dense image sequences in tracking tasks. First, we highlight the issue of temporal inconsistency in existing Mamba-based methods, specifically the failure to account for temporal continuity in the Mamba scanning mechanism. Secondly, building upon this insight,we propose TrackingMiM, a Mamba-in-Mamba architecture, a minimal-computation burden model for handling image sequence of tracking problem. In our framework, the mamba scan is performed in a nested way while independently process temporal and spatial coherent patch tokens. While the template frame is encoded as query token and utilized for tracking in every scan. Extensive experiments conducted on five UAV tracking benchmarks confirm that the proposed TrackingMiM achieves state-of-the-art precision while offering noticeable higher speed in UAV tracking.

Figures

Figures reproduced from arXiv: 2507.01535 by the authors.

Figure 1
Figure 1. Compared to state-of-the-art UAV tracking algorithms on the UAV123 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. TrackingMiM Architecture. The template frame is selected at t=0 with a bounding box (bbox) input. Subsequent frames are captured and processed within a fixed temporal window. Each frame is first tokenized into patches before being fed into the MiM blocks. In each MiM block, spatial bi-directional scanning is first applied to the template, followed by each frame to integrate template information. A temporal scan then… view at source ↗
Figure 3
Figure 3. Precision and Success rates versus overlap thresholds on three datasets: UAV123, UAVDT, and VisDrone2018. AUC-based rankings are shown on the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative evaluation on 6 video sequences from DTB70 ( [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results on 4 video sequences from UAV123 and VisDrone2018 ( [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: The center plot shows IoU trends over time for videos from DTB70, with each tracker represented by a distinct color. The leftmost column displays [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Class activation maps (CAM) for the ablation study. From left to [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

75 extracted references · 67 canonical work pages

  1. [1]

    Eye in the sky: Drone-based object tracking and 3d localization,

    H. Zhang, G. Wang, Z. Lei, and J.-N. Hwang, “Eye in the sky: Drone-based object tracking and 3d localization,” in Proceedings of the 27th ACM International Conference on Multimedia, ser. MM ’19. New York, NY, USA: Association for Computing Machinery, 2019, p. 899–907. [Online]. Available: https://doi.org/10.1145/3343031.3350933 1

    arXiv 2019

  2. [2]

    Strategies for optimized uav surveillance in various tasks and scenarios: A review,

    Z. Fang and A. V . Savkin, “Strategies for optimized uav surveillance in various tasks and scenarios: A review,” Drones, vol. 8, no. 5, 2024. [Online]. Available: https://www.mdpi.com/ 2504-446X/8/5/193 1

    work page 2024

  3. [3]

    A two-stage target search and tracking method for uav based on deep reinforcement learning,

    M. Liu, J. Wei, and K. Liu, “A two-stage target search and tracking method for uav based on deep reinforcement learning,” Drones, vol. 8, no. 10, 2024. [Online]. Available: https://www.mdpi.com/2504-446X/8/10/544 1

    work page 2024

  4. [4]

    Coarse-to-fine uav target tracking with deep reinforcement learning,

    W. Zhang, K. Song, X. Rong, and Y. Li, “Coarse-to-fine uav target tracking with deep reinforcement learning,” IEEE Transactions on Automation Science and Engineering , vol. 16, no. 4, pp. 1522–1530,

    work page

  5. [5]

    A strong uav vision tracker based on deep broad learning system and correlation filter,

    M. Wang, Q. Ge, B. Zhu, and C. Sun, “A strong uav vision tracker based on deep broad learning system and correlation filter,” IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 5714– 5728, 2025. 1

    work page 2025

  6. [6]

    Spatial-temporal initialization dilemma: towards realistic visual tracking,

    C. Liu, Y. Yuan, X. Chen, H. Lu, and D. Wang, “Spatial-temporal initialization dilemma: towards realistic visual tracking,” Visual Intelligence, vol. 2, no. 1, p. 35, 2024. 1

    work page 2024

  7. [7]

    Learning target-aware vision transformers for real-time uav tracking,

    S. Li, X. Yang, X. Wang, D. Zeng, H. Ye, and Q. Zhao, “Learning target-aware vision transformers for real-time uav tracking,” IEEE Transactions on Geoscience and Remote Sensing, 2024. 1, 3

    work page 2024

  8. [8]

    Adaptive and background- aware vision transformer for real-time uav tracking,

    S. Li, Y. Yang, D. Zeng, and X. Wang, “Adaptive and background- aware vision transformer for real-time uav tracking,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 13 989–14 000. 1, 3, 6, 8

    work page 2023

Show all 75 references
  1. [9]

    Learning disentangled repre- sentation with mutual information maximization for real-time uav tracking,

    X. Wang, X. Yang, H. Ye, and S. Li, “Learning disentangled repre- sentation with mutual information maximization for real-time uav tracking,” in 2023 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2023, pp. 1331–1336. 1

    2023

  2. [10]

    High-speed tracking with kernelized correlation filters,

    J. F. Henriques, R. Caseiro, P . Martins, and J. Batista, “High-speed tracking with kernelized correlation filters,” IEEE transactions on pattern analysis and machine intelligence , vol. 37, no. 3, pp. 583–596,

  3. [11]

    Learning background- aware correlation filters for visual tracking,

    H. Kiani Galoogahi, A. Fagg, and S. Lucey, “Learning background- aware correlation filters for visual tracking,” in Proceedings of the IEEE international conference on computer vision, 2017, pp. 1135–1143. 2, 8

    2017

  4. [12]

    HiFT: Hierarchical Feature Transformer for Aerial Tracking,

    Z. Cao, C. Fu, J. Ye, B. Li, and Y. Li, “HiFT: Hierarchical Feature Transformer for Aerial Tracking,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 15 457– 15 466. 2, 8

    2021

  5. [13]

    SiamAPN++: Siamese Attentional Aggregation Network for Real-Time UAV Tracking,

    ——, “SiamAPN++: Siamese Attentional Aggregation Network for Real-Time UAV Tracking,” in Proceedings of the IEEE/RSJ Interna- tional Conference on Intelligent Robots and Systems (IROS) , 2021, pp. 3086–3092. 2, 8

    2021

  6. [14]

    Lighttrack: Finding lightweight neural networks for object tracking via one- shot architecture search,

    B. Yan, H. Peng, K. Wu, D. Wang, J. Fu, and H. Lu, “Lighttrack: Finding lightweight neural networks for object tracking via one- shot architecture search,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 15 180–15 189. 2, 8

    2021

  7. [15]

    Rank-based filter pruning for real-time uav tracking,

    X. Wang, D. Zeng, Q. Zhao, and S. Li, “Rank-based filter pruning for real-time uav tracking,” in IEEE International Conference on Multimedia and Expo (ICME), 2022. 2

    2022

  8. [16]

    Fisher pruning for real-time uav tracking,

    W. Wu, P . Zhong, and S. Li, “Fisher pruning for real-time uav tracking,” in 2022 International Joint Conference on Neural Networks (IJCNN). IEEE, 2022, pp. 1–7. 2

    2022

  9. [17]

    Fisher pruning for developing real-time uav trackers,

    P . Zhong, W. Wu, X. Dai, Q. Zhao, and S. Li, “Fisher pruning for developing real-time uav trackers,” Journal of Real-Time Image Processing, 2023. 2

    2023

  10. [18]

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

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

    2022

  11. [19]

    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,” in European Conference on Computer Vision. Springer, 2022, pp. 375–392. 2

    2022

  12. [20]

    Zoomtrack: target-aware non-uniform resizing for efficient visual tracking,

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

  13. [21]

    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,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 13 608–13 618. 2, 3

    2022

  14. [23]

    Vmamba: Visual state space model,

    Y. Liu, Y. Tian, Y. Zhao, H. Yu, L. Xie, Y. Wang, Q. Ye, and Y. Liu, “Vmamba: Visual state space model,” arXiv preprint arXiv:2401.10166, 2024. 2, 3

    2024 arXiv

  15. [24]

    Ye and T

    Z. Ye and T. Chen, “P-mamba: Marrying perona malik diffusion SUBMITTED TO IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. XX, NO. X, JULY 9999 12 with mamba for efficient pediatric echocardiographic left ventricu- lar segmentation,” arXiv preprint arXiv:2402.0850...

    2024 arXiv

  16. [25]

    Fast and incremental method for loop-closure detection using bags of visual words,

    A. Angeli, D. Filliat, S. Doncieux, and J.-A. Meyer, “Fast and incremental method for loop-closure detection using bags of visual words,” IEEE transactions on robotics , vol. 24, no. 5, pp. 1027–1037,

  17. [26]

    Learning residue-aware correlation filters and refining scale estimates with the grabcut for real-time uav tracking,

    S. Li, Y. Liu, Q. Zhao, and Z. Feng, “Learning residue-aware correlation filters and refining scale estimates with the grabcut for real-time uav tracking,” in 2021 International Conference on 3D Vision (3DV). IEEE, 2021, pp. 1238–1248. 2

    2021

  18. [27]

    Learning aberrance re- pressed correlation filters for real-time uav tracking,

    Z. Huang, C. Fu, Y. Li, F. Lin, and P . Lu, “Learning aberrance re- pressed correlation filters for real-time uav tracking,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , 2019, pp. 2891–2900. 2, 8

    2019

  19. [28]

    AutoTrack: To- wards High-Performance Visual Tracking for UAV with Automatic Spatio-Temporal Regularization,

    Y. Li, C. Fu, F. Ding, Z. Huang, and G. Lu, “AutoTrack: To- wards High-Performance Visual Tracking for UAV with Automatic Spatio-Temporal Regularization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 11 923–11 932. 2, 8

    2020

  20. [29]

    Global filter pruning with self-attention for real-time uav tracking,

    M. Liu, Y. Wang, Q. Sun, and S. Li, “Global filter pruning with self-attention for real-time uav tracking,” in British Machine Vision Conference (BMVC), 2022. 2

    2022

  21. [30]

    Tctrack: Temporal contexts for aerial tracking,

    Z. Cao, Z. Huang, L. Pan, S. Zhang, Z. Liu, and C. Fu, “Tctrack: Temporal contexts for aerial tracking,” in IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR), 2022. 2, 8

    2022

  22. [31]

    Autore- gressive queries for adaptive tracking with spatio-temporal trans- formers,

    J. Xie, B. Zhong, Z. Mo, S. Zhang, L. Shi, S. Song, and R. Ji, “Autore- gressive queries for adaptive tracking with spatio-temporal trans- formers,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 19 300–19 309. 3

    2024

  23. [32]

    Explicit visual prompts for visual object tracking,

    L. Shi, B. Zhong, Q. Liang, N. Li, S. Zhang, and X. Li, “Explicit visual prompts for visual object tracking,” in AAAI, 2024. 3

    2024

  24. [33]

    Learning track- ing representations via dual-branch fully transformer networks,

    F. Xie, C. Wang, G. Wang, W. Yang, and W. Zeng, “Learning track- ing representations via dual-branch fully transformer networks,” in IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), 2021. 3

    2021

  25. [34]

    Minivit: Compressing vision transformers with weight multiplex- ing,

    J. Zhang, H. Peng, K. Wu, M. Liu, B. Xiao, J. Fu, and L. Yuan, “Minivit: Compressing vision transformers with weight multiplex- ing,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3

    2022

  26. [35]

    Tprune: Efficient transformer pruning for mobile devices,

    J. Mao, H. Yang, A. Li, H. Li, and Y. Chen, “Tprune: Efficient transformer pruning for mobile devices,” ACM Transactions on Cyber-Physical Systems (TCPS), 2021. 3

    2021

  27. [36]

    Efficientformer: Vision transformers at mo- bilenet speed,

    Y. Li, G. Yuan, Y. Wen, E. Hu, G. Evangelidis, S. Tulyakov, Y. Wang, and J. Ren, “Efficientformer: Vision transformers at mo- bilenet speed,” in Advances in Neural Information Processing Systems (NeurIPS), 2022. 3

    2022

  28. [37]

    Mobile-former: Bridging mobilenet and transformer,

    Y. Chen, X. Dai, D. Chen, M. Liu, X. Dong, L. Yuan, and Z. Liu, “Mobile-former: Bridging mobilenet and transformer,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3

    2021

  29. [38]

    Dynamicvit: Efficient vision transformers with dynamic token sparsification,

    Y. Rao, W. Zhao, B. Liu, J. Lu, J. Zhou, and C.-J. Hsieh, “Dynamicvit: Efficient vision transformers with dynamic token sparsification,” in Advances in Neural Information Processing Systems (NeurIPS), 2021. 3

    2021

  30. [39]

    A-vit: Adaptive tokens for efficient vision trans- former,

    H. Yin, A. Vahdat, J. M. Alvarez, A. Mallya, J. Kautz, and P . Molchanov, “A-vit: Adaptive tokens for efficient vision trans- former,” in IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR), 2022. 3

    2022

  31. [40]

    Mim-istd: Mamba-in-mamba for efficient infrared small tar- get detection,

    T. Chen, Z. Ye, Z. Tan, T. Gong, Y. Wu, Q. Chu, B. Liu, N. Yu, and J. Ye, “Mim-istd: Mamba-in-mamba for efficient infrared small tar- get detection,” IEEE Transactions on Geoscience and Remote Sensing ,

  32. [41]

    Mamba-fetrack: Frame-event tracking via state space model,

    J. Huang, S. Wang, S. Wang, Z. Wu, X. Wang, and B. Jiang, “Mamba-fetrack: Frame-event tracking via state space model,” in Chinese Conference on Pattern Recognition and Computer Vision (PRCV). Springer, 2024, pp. 3–18. 3

    2024

  33. [42]

    Combining recurrent, convolutional, and continuous-time models with linear state-space layers,

    A. Gu, I. Johnson, K. Goel, K. K. Saab, T. Dao, A. Rudra, and C. R’e, “Combining recurrent, convolutional, and continuous-time models with linear state-space layers,” in Neural Information Processing Systems, 2021. 3

    2021

  34. [43]

    Vision mamba: Efficient visual representation learning with bidirectional state space model,

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: Efficient visual representation learning with bidirectional state space model,” ICML, 2024. 3

    2024

  35. [44]

    S4nd: Modeling images and videos as multidimen- sional signals with state spaces,

    E. Nguyen, K. Goel, A. Gu, G. Downs, P . Shah, T. Dao, S. Baccus, and C. Ré, “S4nd: Modeling images and videos as multidimen- sional signals with state spaces,” Advances in neural information processing systems, vol. 35, pp. 2846–2861, 2022. 3

    2022

  36. [45]

    Mamba-nd: Selective state space modeling for multi-dimensional data,

    S. Li, H. Singh, and A. Grover, “Mamba-nd: Selective state space modeling for multi-dimensional data,” arXiv, 2024. 3

    2024

  37. [46]

    Pan- mamba: Effective pan-sharpening with state space model,

    X. He, K. Cao, K. R. Yan, R. Li, C. Xie, J. Zhang, and M. Zhou, “Pan- mamba: Effective pan-sharpening with state space model,” ArXiv, vol. abs/2402.12192, 2024. 3

    2024 arXiv

  38. [47]

    Retrieval- augmented generation for knowledge-intensive nlp tasks,

    P . Lewis, E. Perez, A. Piktus, F. Petroni, V . Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel et al. , “Retrieval- augmented generation for knowledge-intensive nlp tasks,” in NeurIPS, 2020. 3

    2020

  39. [48]

    Retrieval augmented classification for long-tail visual recognition,

    A. Long, W. Yin, T. Ajanthan, V . Nguyen, P . Purkait, R. Garg, A. Blair, C. Shen, and A. van den Hengel, “Retrieval augmented classification for long-tail visual recognition,” in CVPR, 2022. 3

    2022

  40. [49]

    Texture memory-augmented deep patch-based image inpainting,

    R. Xu, M. Guo, J. Wang, X. Li, B. Zhou, and C. C. Loy, “Texture memory-augmented deep patch-based image inpainting,” TIP,

  41. [50]

    Re- trievegan: Image synthesis via differentiable patch retrieval,

    H.-Y. Tseng, H.-Y. Lee, L. Jiang, M.-H. Yang, and W. Yang, “Re- trievegan: Image synthesis via differentiable patch retrieval,” in ECCV, 2020. 3

    2020

  42. [51]

    Retrieval- augmented few-shot medical image segmentation with foundation models,

    L. Zhao, X. Chen, E. Z. Chen, Y. Liu, T. Chen, and S. Sun, “Retrieval- augmented few-shot medical image segmentation with foundation models,” arXiv preprint arXiv:2408.08813, 2024. 3

    2024 arXiv

  43. [52]

    Retrieval- augmented diffusion models,

    A. Blattmann, R. Rombach, K. Oktay, and B. Ommer, “Retrieval- augmented diffusion models,” in ARXIV, 2022. 3

    2022

  44. [53]

    Retrieval-augmented open- vocabulary object detection,

    J. Kim, E. Cho, S. Kim, and H. J. Kim, “Retrieval-augmented open- vocabulary object detection,” in Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2024, pp. 17 427– 17 436. 3

    2024

  45. [54]

    Rtagrasp: Learning task-oriented grasping from human videos via retrieval, transfer, and alignment,

    W. Dong, D. Huang, J. Liu, C. Tang, and H. Zhang, “Rtagrasp: Learning task-oriented grasping from human videos via retrieval, transfer, and alignment,” arXiv preprint arXiv:2409.16033, 2024. 3

    2024 arXiv

  46. [55]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” ArXiv, vol. abs/2312.00752, 2023. 4

    2023 arXiv

  47. [56]

    State space model for new- generation network alternative to transformers: A survey,

    X. Wang, S. Wang, Y. Ding, Y. Li, W. Wu, Y. Rong, W. Kong, J. Huang, S. Li, H. Yang et al. , “State space model for new- generation network alternative to transformers: A survey,” arXiv preprint arXiv:2404.09516, 2024. 4

    2024 arXiv

  48. [57]

    Discrimina- tive scale space tracking,

    M. Danelljan, G. Hager, F. S. Khan, and M. Felsberg, “Discrimina- tive scale space tracking,” IEEE transactions on pattern analysis and machine intelligence (TP AMI), 2017. 8

    2017

  49. [58]

    Eco: Efficient convolution operators for tracking,

    M. Danelljan, G. Bhat, F. Shahbaz Khan, and M. Felsberg, “Eco: Efficient convolution operators for tracking,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 6, 8

    2017

  50. [59]

    Multi- Cue Correlation Filters for Robust Visual Tracking,

    N. Wang, W. Zhou, Q. Tian, R. Hong, M. Wang, and H. Li, “Multi- Cue Correlation Filters for Robust Visual Tracking,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 4844–4853. 8

    2018

  51. [60]

    Learning spatial-temporal regularized correlation filters for visual tracking,

    F. Li, C. Tian, W. Zuo, L. Zhang, and M.-H. Yang, “Learning spatial-temporal regularized correlation filters for visual tracking,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 4904–4913. 8

    2018

  52. [61]

    Learning residue-aware correlation filters and refining scale for real-time uav tracking,

    S. Li, Y. Liu, Q. Zhao, and Z. Feng, “Learning residue-aware correlation filters and refining scale for real-time uav tracking,” Pattern Recognition, vol. 127, p. 108614, 2022. 6, 8

    2022

  53. [62]

    Efficient visual tracking via hierarchical cross-attention transformer,

    X. Chen, B. Kang, D. Wang, D. Li, and H. Lu, “Efficient visual tracking via hierarchical cross-attention transformer,” in European conference on computer vision. Springer, 2022, pp. 461–477. 6, 8

    2022

  54. [63]

    Unsupervised domain adaptation for nighttime aerial tracking,

    J. Ye, C. Fu, G. Zheng, D. P . Paudel, and G. Chen, “Unsupervised domain adaptation for nighttime aerial tracking,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 8896–8905. 6, 8

    2022

  55. [64]

    Adversarial blur-deblur network for robust uav tracking,

    H. Zuo, C. Fu, S. Li, K. Lu, Y. Li, and C. Feng, “Adversarial blur-deblur network for robust uav tracking,” IEEE Robotics and Automation Letters (RAL), 2023. 8

    2023

  56. [65]

    Towards discriminative rep- resentations with contrastive instances for real-time uav tracking,

    D. Zeng, M. Zou, X. Wang, and S. Li, “Towards discriminative rep- resentations with contrastive instances for real-time uav tracking,” in 2023 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 2023, pp. 1349–1354. 6, 8

    2023

  57. [66]

    Exploring lightweight hierarchical vision transformers for efficient visual tracking,

    B. Kang, X. Chen, D. Wang, H. Peng, and H. Lu, “Exploring lightweight hierarchical vision transformers for efficient visual tracking,” in Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 9612–9621. 8

    2023

  58. [67]

    Litetrack: Layer pruning with asynchronous feature extraction for lightweight and efficient visual tracking,

    Q. Wei, B. Zeng, J. Liu, L. He, and G. Zeng, “Litetrack: Layer pruning with asynchronous feature extraction for lightweight and efficient visual tracking,” in 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 4968–4975. 8 SUBMITTED TO IEEE T...

    2024

  59. [68]

    Sgdvit: Saliency-guided dynamic vision transformer for uav tracking,

    L. Yao, C. Fu, S. Li, G. Zheng, and J. Ye, “Sgdvit: Saliency-guided dynamic vision transformer for uav tracking,” in 2023 IEEE Inter- national Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 3353–3359. 8

    2023

  60. [69]

    Mixformerv2: Efficient fully transformer tracking,

    Y. Cui, T. Song, G. Wu, and L. Wang, “Mixformerv2: Efficient fully transformer tracking,” Advances in neural information processing systems, vol. 36, pp. 58 736–58 751, 2023. 8

    2023

  61. [70]

    Separable self and mixed attention transformers for efficient object tracking,

    G. Y. Gopal and M. A. Amer, “Separable self and mixed attention transformers for efficient object tracking,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 6708–6717. 8

    2024

  62. [71]

    Lightweight full- convolutional siamese tracker,

    Y. Li, B. Wang, X. Wu, Z. Liu, and Y. Li, “Lightweight full- convolutional siamese tracker,” Knowledge-Based Systems, vol. 286, p. 111439, 2024. 8

    2024

  63. [72]

    Learning motion blur robust vision transformers with dynamic early exit for real-time uav tracking,

    Y. Wu, X. Wang, D. Zeng, H. Ye, X. Xie, Q. Zhao, and S. Li, “Learning motion blur robust vision transformers with dynamic early exit for real-time uav tracking,” 2024. [Online]. Available: https://arxiv.org/abs/2407.05383 6, 8

    2024 arXiv

  64. [73]

    A benchmark and simula- tor for uav tracking,

    M. Mueller, N. Smith, and B. Ghanem, “A benchmark and simula- tor for uav tracking,” inComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceed- ings, Part I 14. Springer, 2016, pp. 445–461. 6

    2016

  65. [74]

    Visdrone-sot2018: The vision meets drone single-object tracking challenge results,

    L. Wen, P . Zhu, D. Du, X. Bian, H. Ling, Q. Hu, C. Liu, H. Cheng, X. Liu, W. Ma et al. , “Visdrone-sot2018: The vision meets drone single-object tracking challenge results,” in Proceedings of the Eu- ropean conference on computer vision (ECCV) workshops, 2018, pp. 0–0. 6

    2018

  66. [75]

    The unmanned aerial vehicle benchmark: Object detection and tracking,

    D. Du, Y. Qi, H. Yu, Y. Yang, K. Duan, G. Li, W. Zhang, Q. Huang, and Q. Tian, “The unmanned aerial vehicle benchmark: Object detection and tracking,” in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 370–386. 6

    2018

  67. [76]

    Visual object tracking for unmanned aerial vehicles: A benchmark and new motion models,

    S. Li and D. Y. Yeung, “Visual object tracking for unmanned aerial vehicles: A benchmark and new motion models,” in AAAI Conference on Artificial Intelligence (AAAI), 2017. 6

    2017

Pith tools

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