Pith. sign in

REVIEW 4 major objections 5 minor 26 references

Multiple Object Tracking in Video SAR: A Benchmark and Tracking Baseline

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper introduces VSMB, the first public benchmark for multi-object tracking in video synthetic aperture radar, and a DETR-based tracker that uses Doppler line artifacts to outperform previous MOT methods on it.

desk verdict A useful first benchmark for video SAR MOT, with a solid baseline tracker; the main gap is an under-specified train/val split that should be verifiable from the released code. read the letter →

arxiv 2506.12105 v1 pith:27WSYZVY submitted 2025-06-13 cs.CV

Haoxiang Chen , Wei Zhao , Rufei Zhang , Nannan Li , Dongjin Li This is my paper · ORCID
classification cs.CV
keywords multi-objecttrackingvideoSARbenchmarkdatasetDopplermismatchlinefeatureenhancementmotion-awareassociationDETRshadowdetection
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 remove a bottleneck in video synthetic aperture radar (Video SAR): the absence of a public dataset for testing multi-object tracking (MOT). It introduces VSMB, a collection of 45 annotated sequences with 4,394 frames, 25,466 target instances, and 355 trajectories covering cars, ships, and airplanes. Alongside the dataset, it proposes VSMT, a detection-based tracker that treats Doppler-induced trailing and defocusing streaks as informative line features rather than noise, and that adaptively down-weights appearance similarity when a target's radar appearance changes. If the claims hold, VSMB provides a common evaluation ground for future Video SAR MOT research, and VSMT demonstrates that line-feature enhancement plus motion-aware association yields state-of-the-art results on that benchmark.

What carries the argument

Three modules carry the argument. Line Feature Focusing Module (LFFM) applies a discrete Radon transform to each scale's feature map, accumulating responses along line segments into a pointwise Radon-domain representation $Y(\hat\theta_F,\hat\rho_F)=\sum_{l\in L}X(l)$, then reconstructs a line-feature intensity map $A_{\text{soft}}$ via an inverse-projection accumulation, gating the original features with this map. Line Feature Assigner (LFA) selectively enhances the decoder's top-$K$ proposals by pooling $A_{\text{soft}}$ in an adaptive radius $R_i = R_{\min,i} + \hat{v}_i(R_{\max,i}-R_{\min,i})$, where $\hat{v}_i$ is a speed estimate derived from the Doppler-shift relation $f_d(x,y)=\frac{2}{\lambda}(\frac{dx_r}{dt}-\frac{dx_a}{dt})$, so that faster targets sample a wider line-feature neighborhood. Motion-awareness Association (MaA) adjusts the weight of appearance similarity in the ByteTrack-style association cost matrix according to the target's motion state, keeping IDs stable when a target transitions from stationary to moving and its radar signature changes.

What would settle it

Re-run VSMT with a split that holds out entire videos (no frames of any given sequence appear in both training and validation), then compare MOTA on the validation set; a substantial drop relative to the reported 78.2 would confirm that temporal leakage contributed to the state-of-the-art numbers.

Watch

Extended reading notes

Core claim

The paper's central claim is that VSMB is the first publicly available dataset built specifically for video SAR target tracking, enabling standardized comparison of MOT algorithms in this modality. On this dataset, the proposed VSMT model achieves state-of-the-art detection and tracking performance: with a Swin-L backbone it reports MOTA 78.2, IDF1 86.1, and HOTA 70.3, outperforming prior JDT, DBT, and IAT trackers. The paper further claims that Doppler-generated line artifacts, previously treated as interference, can be turned into a positive cue: a line feature focusing module amplifies these streaks to distinguish moving-target shadows from static-occlusion shadows, while a motion-aware association mechanism reduces trajectory fragmentation when target appearance changes.

Load-bearing premise

The evaluation protocol assumes the 7:3 training/validation split is leakage-free, but the paper does not specify whether the split is at the video level or the frame level; if frames from the same recording appear in both sets, temporal correlation would inflate every reported metric, including the state-of-the-art claim.

Editorial extensions

If this is right

  • If VSMB becomes the standard benchmark, future video SAR tracking methods can be compared directly on identical data, replacing the current practice of private or non-public sequences.
  • VSMT's results suggest that detection-based tracking (DBT) is more suitable than joint-detection-and-tracking (JDT) or implicit-association frameworks for small radar targets, where multi-task feature sharing amplifies task misalignment.
  • The line-feature enhancement approach indicates that Doppler-induced trailing and defocusing artifacts can be exploited to reduce false alarms from static man-made shadows, at least on the scenarios in VSMB.
  • If the reported performance is reproducible, a Swin-L VSMT model can serve as a strong baseline for subsequent research on video SAR MOT, with the released code and weights.

Reading between the lines

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

  • My inference: the Radon-domain line-feature trick may transfer to other radar and sonar modalities where targets produce streak-like signatures, such as through-wall radar or automotive radar, where similar static-clutter confusion arises.
  • My inference: the paper's core comparison would be stronger if the 7:3 split were shown to be video-level; a hold-one-video-out evaluation would clarify whether the reported accuracy reflects generalization or temporal correlation.
  • My inference: the adaptive radius in LFA couples detection to a per-frame speed estimate, so a natural stress test is to run VSMT on scenes with rapidly accelerating or decelerating targets, where the speed estimate may lag the true motion.
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

4 major / 5 minor

Summary. The paper introduces VSMB, a video SAR multi-object tracking benchmark with 45 annotated sequences covering cars, ships, and airplanes, and proposes VSMT, a DETR-based tracker built on DINO and ByteTrack with three new components: LFFM (a Radon-domain line feature focusing module), LFA (a line feature assigner with motion-aware matching radius), and MaA (motion-aware association that modulates appearance similarity). The authors evaluate VSMT against several recent MOT methods and report state-of-the-art MOTA, IDF1, and HOTA on their validation split, with ablations showing each component contributes. The dataset and code are released publicly.

Significance. If the evaluation split and the tabular results are made verifiable, VSMB fills a real gap by providing a public common benchmark for video SAR tracking, and the proposed modules target a genuine domain-specific failure mode (Doppler trailing and defocusing) that generic MOT methods ignore. The release of data and code is a concrete asset for the community. However, the scientific value of the SOTA claim is currently conditional on the unverified split protocol and on corrected numeric tables, so the contribution is promising but not yet established at the level claimed.

major comments (4)
  1. [V-A] The description of the training/validation split is not sufficient to rule out temporal leakage: the sentence 'Under the premise of avoiding data leakage' does not state whether the 7:3 split is at the video, scene, or frame level. Because video SAR sequences have strong temporal correlation (targets persist, illumination and Doppler artifacts evolve smoothly), a frame-level split would allow the tracker to memorize scene-specific patterns and would inflate MOTA, IDF1, and HOTA relative to an honest split. This is load-bearing for the state-of-the-art claim. Please specify the split granularity, publish the exact video and sequence IDs in train and validation, and provide a script that reproduces the split.
  2. [Table III] Several cells in Table III are unreadable due to formatting/merging issues: for example, OC-SORT appears as '70.192.362.4', StrongSORT as '75.7113243 2483.6', and Dino-Byte as '74.953219 47 ... 76.893.365.8'. Without clean columns and correct decimal separators, the reader cannot verify the claimed margins or even the direction of the differences, and the central SOTA comparison is not checkable. Please regenerate the table with proper alignment and include all metric columns (FP and FN are referenced in the text but not shown separately as columns).
  3. [Table II] Table II is ambiguous and internally inconsistent as printed. The column header 'Trajector(<16 2) (<32 2) (<64 2)' is garbled, and the Small/Medium/Large percentages clearly do not correspond to mutually exclusive bins (e.g., Car: 8.67% + 35.09% + 99.24% exceeds 100%). The reader cannot determine what the size thresholds are, how trajectory lengths are counted, or what the total trajectory distribution is. Please rewrite the table with disjoint size categories, clearly labeled trajectory-length statistics, and percentages that sum to 100% within each row or an explicit note explaining overlapping categories.
  4. [V-B] All comparisons are single-run point estimates with no variance, repeated runs, or per-sequence breakdown. On a dataset of 45 videos and a single split, the 2.5–3.0 MOTA gap between VSMT and the next-best method could be within run-to-run or split-to-split variability, especially given the small validation set. Please report results over multiple runs with means and standard deviations (or state that a fixed seed and deterministic inference make runs reproducible), and consider reporting per-sequence metrics to show the improvement is consistent rather than driven by a few videos.
minor comments (5)
  1. [IV-B, Eq. (6)] The sentence after Eq. (6) states that the indicator function 'controls the filtering of noise' but never names the threshold τ that appears in the equation; please state explicitly that τ is the filtering threshold and describe how it is set.
  2. [IV-C, Eq. (11)] The quantity Δf in Eq. (11) is not defined, and CMC is only cited; please define Δf and briefly describe the CMC operation or remove the notation if it is not essential.
  3. [IV-C] In the LFA paragraph, 'SVMT employs the LFA' appears to be a typo and should read 'VSMT employs the LFA.'
  4. [III and IV-A] The detector is referred to as both 'Dino' and 'DETR-Base model,' while reference [22] is DINO; please use one consistent designation throughout.
  5. [Table I] The abbreviation 'SS' is expanded only as 'Smart Satellite' in the table header, which is easy to confuse with the text's 'SmartSat-X1'; please align the naming.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the SOTA claim is an empirical benchmark result, and the two self-citations are not load-bearing.

full rationale

The derivation chain is self-contained. Eq. (11) defines the velocity supervision target from ground-truth box displacements; this is a training label, not a post-hoc fitted quantity, and Eq. (13) uses the predicted velocity to set an association radius as an empirical design choice rather than a reduction. LFFM/LFA use Radon transforms and Doppler relation (10) as motivation; their effectiveness is established by the ablation study in Table IV, not by citing that relation. There are two self-citations: [15] as an example DBT tracker in Section III, and [25] as one of two cited inspirations for LFFM in Section IV-B. Neither is load-bearing: [15] is a category illustration, and [25] is accompanied by external inspiration [24] for a module whose contribution is validated by Table IV. No uniqueness theorem is imported, and no ansatz is smuggled through self-citation. The 'state-of-the-art on VSMB' claim is self-referential because the benchmark is new, but it is not circular: the comparison is against external baselines on the same held-out split, and the method's components are independently ablated. The one flagged weakness is the unsupported premise in Section V-A: 'Under the premise of avoiding data leakage, the training and validation sets are divided in a ratio of 7:3.' This does not specify whether the split is at video, scene, or frame level. If frames from the same video appear in both sets, temporal correlation could inflate all reported metrics; however, that is a data-integrity risk, not a circular reduction. Overall, no equation-level or definitional circularity is present; the score of 2 reflects only the presence of minor non-load-bearing self-citations.

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

The central claim rests mostly on dataset construction and an empirical tracker, so the ledger contains hyperparameters and domain assumptions rather than invented physical entities. The two fitted constants are the LFA radius scaling and the Radon threshold. The key domain assumptions are that the Doppler model explains trailing artifacts and that Radon-domain line features separate moving targets from static shadows. The evaluation additionally assumes a leakage-free split and accurate annotations.

free parameters (3)
  • lambda_max (LFA maximum radius scaling) = not reported; tuned in [0.2, 0.6] per Fig. 4
    Controls the upper bound of the matching radius in Eq. (13); the paper selects its value on the validation set, so it is a fitted hyperparameter that affects association performance.
  • tau (Radon threshold in Eq. 6) = not reported
    Filters weak line responses in the LFFM; no value or selection procedure is given, leaving a hidden tuning knob.
  • Radon bin resolutions Theta_F and P_F = not reported
    Set the angular and radial discretization in Eqs. (2)-(3); the paper does not specify these values, and they affect line feature quality.
assumptions (4)
  • domain assumption The Doppler shift model fd = (2/lambda)(dx_r/dt - dx_a/dt) from [2] correctly describes trailing and defocusing artifacts in video SAR.
    Invoked in Section IV-C to justify the LFA radius adjustment; if the physical model does not capture the observed artifacts, the radius formula loses its motivation.
  • domain assumption Line features in the Radon domain are discriminative for separating moving-target shadows from static occlusion shadows.
    The core premise of LFFM/LFA, motivated by [10] but not independently proved; the paper provides only empirical validation on its own benchmark.
  • ad hoc to paper The 7:3 train/validation split avoids data leakage.
    Stated in Section V-A but without detail on whether the split is at the video or frame level; this is a paper-specific assumption required for the SOTA claim.
  • domain assumption DarkLabel annotations are accurate and consistent.
    The dataset quality rests on frame-by-frame labeling with no inter-annotator agreement or quality metrics reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multiple Object Tracking in Video SAR: A Benchmark and Tracking Baseline." pith.science (2026). https://pith.science/paper/27WSYZVY

@misc{pith2026250612105,
  author       = {Pith},
  title        = {Pith review of: Multiple Object Tracking in Video SAR: A Benchmark and Tracking Baseline},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/27WSYZVY}},
  note         = {Machine review of arXiv:2506.12105}
}
read the original abstract

In the context of multi-object tracking using video synthetic aperture radar (Video SAR), Doppler shifts induced by target motion result in artifacts that are easily mistaken for shadows caused by static occlusions. Moreover, appearance changes of the target caused by Doppler mismatch may lead to association failures and disrupt trajectory continuity. A major limitation in this field is the lack of public benchmark datasets for standardized algorithm evaluation. To address the above challenges, we collected and annotated 45 video SAR sequences containing moving targets, and named the Video SAR MOT Benchmark (VSMB). Specifically, to mitigate the effects of trailing and defocusing in moving targets, we introduce a line feature enhancement mechanism that emphasizes the positive role of motion shadows and reduces false alarms induced by static occlusions. In addition, to mitigate the adverse effects of target appearance variations, we propose a motion-aware clue discarding mechanism that substantially improves tracking robustness in Video SAR. The proposed model achieves state-of-the-art performance on the VSMB, and the dataset and model are released at https://github.com/softwarePupil/VSMB.

Figures

Figures reproduced from arXiv: 2506.12105 by the authors.

Figure 1
Figure 1. Challenges of MOT in Video SAR. (a) Doppler mismatch of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Schematic diagram of VSMT. (a) Detection stage of VSMT with Dino as a baseline model, LFFM and LFA are highlighted in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. LFFM transforms X in the spatial domain to Y in the Radom domain. Although the backbone performs well in extracting point￾like targets [23], it is less effective for elongated linear struc￾tures. Inspired by [24] and [25], LFFM converts linear features from trailing and defocusing into pointwise representations that are more easily perceived by the network. As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Experimental validation of LFA scaling factor [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Examples of detections and trajectories. The [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 23 canonical work pages

  1. [1]

    J. Ding, L. Wen, C. Zhong, and O. Loffeld, ”Video SAR moving target indication using deep neural network,”IEEE Trans. Geosci. Remote Sens., vol. 58, no. 10, pp. 7194–7204, 2020

    work page 2020

  2. [2]

    Chen and H

    C.-C. Chen and H. C. Andrews, ”Target-motion-induced radar imaging,” IEEE Trans. Aerosp. Electron. Syst., no. 1, pp. 2–14, 1980

    work page 1980

  3. [3]

    J. Hong, T. Wang, Y . Han, W. Di, and T. Dong, ”A multi-vehicle tracking method for video-SAR with reliable foreground-background motion feature compensation,”IEEE Trans. Geosci. Remote Sens., 2025

    work page 2025

  4. [4]

    J. Bao, X. Zhang, T. Zhang, T. Zeng, Z. Yang, X. Zhan, J. Shi, and S. Wei, ”Shadow-enhanced self-attention and anchor-adaptive network for video SAR moving target tracking,”IEEE Trans. Geosci. Remote Sens., vol. 61, pp. 1–13, 2023

    work page 2023

  5. [5]

    H. Fang, G. Liao, Y . Liu, C. Zeng, X. He, and Q. Meng, ”A dual-mode framework for robust long-term tracking in video SAR,”IEEE Sensors J., 2024

    work page 2024

  6. [6]

    Zhang, X

    W. Zhang, X. Zhang, X. Xu, Y . Xu, Z. Shao, J. Shi, S. Wei, and T. Zeng, ”GNN-JFL: Graph neural network for video SAR shadow tracking with joint motion-appearance feature learning,”IEEE Trans. Geosci. Remote Sens., 2024

    work page 2024

  7. [7]

    M. Su, P. Ni, H. Pei, X. Kou, and G. Xu, ”Graph feature representation for shadow assisted moving target tracking in video SAR,”IEEE Geosci. Remote Sens. Lett., 2025

    work page 2025

  8. [8]

    Zhong, J

    C. Zhong, J. Ding, and Y . Zhang, ”Joint tracking of moving target in single-channel video SAR,”IEEE Trans. Geosci. Remote Sens., vol. 60, pp. 1–18, 2021

    work page 2021

Show all 26 references
  1. [9]

    J. Luan, L. Wen, and J. Ding, ”Multifeature joint detection of moving target in video SAR,”IEEE Geosci. Remote Sens. Lett., vol. 19, pp. 1–5, 2022

    2022

  2. [10]

    B. Zhao, Y . Han, H. Wang, L. Tang, X. Liu, and T. Wang, ”Robust shadow tracking for video SAR,”IEEE Geosci. Remote Sens. Lett., vol. 18, no. 5, pp. 821–825, 2020

    2020

  3. [11]

    G.-S. Xia, X. Bai, J. Ding, Z. Zhu, S. Belongie, J. Luo, M. Datcu, M. Pelillo, and L. Zhang, ”DOTA: A large-scale dataset for object detection in aerial images,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 3974–3983

    2018

  4. [12]

    Wells, K

    L. Wells, K. Sorensen, A. Doerry, et al., ”Developments in SAR and IFSAR systems and technologies at Sandia National Laboratories,”IEEE Aerosp. Conf., Big Sky, MT, pp. 1085–1095, 2003

    2003

  5. [13]

    Zhang, C

    Y . Zhang, C. Wang, X. Wang, W. Zeng, and W. Liu, ”FairMOT: On the fairness of detection and re-identification in multiple object tracking,”Int. J. Comput. Vis., vol. 129, pp. 3069–3087, 2021

    2021

  6. [14]

    X. Zhou, V . Koltun, and P. Kr ¨ahenb¨uhl, ”Tracking objects as points,” in Proc. Eur . Conf. Comput. Vis. (ECCV), 2020, pp. 474–490

    2020

  7. [15]

    Y . Wang, W. Zhao, R. Zhang, N. Li, D. Li, J. Lv, and J. Xu, ”Multi- object tracking in UA Vs with feature fusion distribution and occlusion awareness,”Signal Image Video Process., vol. 19, no. 1, pp. 1–11, 2025

    2025

  8. [16]

    Zhang, P

    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,” inProc. Eur . Conf. Comput. Vis. (ECCV), 2022, pp. 1–21

    2022

  9. [17]

    Aharon, R

    N. Aharon, R. Orfaig, and B. Z. Bobrovsky, ”BoT-SORT: Robust associations multi-pedestrian tracking,”arXiv preprint arXiv:2206.14651, 2022

    2022 arXiv

  10. [18]

    Maggiolino, A

    G. Maggiolino, A. Ahmad, J. Cao, and K. Kitani, ”Deep OC-SORT: Multi-pedestrian tracking by adaptive re-identification,” inProc. IEEE Int. Conf. Image Process. (ICIP), 2023, pp. 3025–3029

    2023

  11. [19]

    Y . Du, Z. Zhao, Y . Song, Y . Zhao, F. Su, T. Gong, and H. Meng, ”StrongSORT: Make DeepSORT great again,”IEEE Trans. Multimedia, vol. 25, pp. 8725–8737, 2023

    2023

  12. [20]

    Zhang, T

    Y . Zhang, T. Wang, and X. Zhang, ”MOTRv2: Bootstrapping end-to-end multi-object tracking by pretrained object detectors,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2023, pp. 22056–22065

    2023

  13. [21]

    R. Gao, J. Qi, and L. Wang, ”Multiple object tracking as ID prediction,” arXiv preprint arXiv:2403.16848, 2024

    2024 arXiv

  14. [22]

    Zhang, F

    H. Zhang, F. Li, S. Liu, L. Zhang, H. Su, J. Zhu, L. M. Ni, and H.-Y . Shum, ”DINO: DETR with improved denoising anchor boxes for end-to- end object detection,”arXiv preprint arXiv:2203.03605, 2022

    2022 arXiv

  15. [23]

    Liang, X

    T. Liang, X. Chu, Y . Liu, Y . Wang, Z. Tang, W. Chu, J. Chen, and H. Ling, ”CBNet: A composite backbone network architecture for object detection,”IEEE Trans. Image Process., vol. 31, pp. 6893–6906, 2022

    2022

  16. [24]

    K. Zhao, Q. Han, C.-B. Zhang, J. Xu, and M.-M. Cheng, ”Deep Hough transform for semantic line detection,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 9, pp. 4793–4806, 2021

    2021

  17. [25]

    Shang, N

    X. Shang, N. Li, D. Li, J. Lv, W. Zhao, R. Zhang, and J. Xu, ”CCLDet: A cross-modality and cross-domain low-light detector,”IEEE Trans. Intell. Transp. Syst., 2025

    2025

  18. [26]

    Irani and P

    M. Irani and P. Anandan, ”A unified approach to moving object detection in 2D and 3D scenes,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 20, no. 6, pp. 577–589, 1998

    1998

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.