Pith. sign in

REVIEW 4 major objections 5 minor 62 references

Supervising a pre-localization priority map, whether from boxes, real gaze, or transferred pseudo-gaze, improves infrared weak-small UAV detection over DINO-DETR.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 13:35 UTC pith:3OL6JBSS

load-bearing objection Solid empirical paper with a genuinely new paired gaze+IR benchmark; the central claim is plausible but PAQI's mechanism is under-controlled and the pseudo-gaze transfer is unvalidated. the 4 major comments →

arxiv 2607.19040 v1 pith:3OL6JBSS submitted 2026-07-21 cs.CV

Gaze-DETR: Top-Down Guidance Through Priority Maps for Infrared Weak-Small UAV Detection with DETR

Nian Liu , Yuxin Yang , Shubo Lin , Sikui Zhang , Liang Li , Boyu Cai , Yizheng Wang , Weiming Hu
show 1 more author
Jin Gao
This is my paper
classification cs.CV
keywords supervisionprioritymapsguidancebox-deriveddetectiongaze-detrlocations
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Gaze-DETR is an object detector for finding tiny drones in thermal infrared video. Its key idea is to make the detector learn where to look before it decides what is there. The authors add a small branch that outputs a 64x64 'priority map' — a heatmap saying which image locations are worth checking first. Two mechanisms use that map: one multiplies image features by the priority values to emphasize high-priority regions, and the other turns the map's strongest peaks into extra starting points for the decoder's query boxes. At test time, only the infrared image is used; the priority map is predicted, not taken from annotations.

The priority map is trained with one of three supervision signals: a Gaussian drawn around the ground-truth box, real eye-tracking fixations from people searching for UAVs in the new TIR-UAV120-Gaze dataset, or 'pseudo-gaze' produced by transferring gaze statistics from that dataset to the public Anti-UAV410 benchmark. Experiments show gains over the DINO-DETR baseline on both benchmarks, mostly in precision and F1, and the gains hold when the same priority-supervision idea is attached to RT-DETR and Deformable-DETR.

The largest improvements are a few mAP50 points, and real gaze only beats the cheap box-derived Gaussian by about 0.4 mAP50 on the new dataset — less than one might expect given the cost of collecting eye-tracking data. The paper is honest about this, noting that neither supervision source is uniformly best.

Core claim

Gaze-DETR's central claim is that an explicitly supervised spatial priority map provides pre-localization guidance complementary to bounding-box supervision: 'explicit spatial-priority learning provides effective pre-localization guidance that complements conventional bounding-box localization supervision' (abstract/conclusion). Empirically, this manifests as Gaze-DETR beating DINO-DETR by +1.84 mAP50/+2.10 F1 with real-gaze supervision on TIR-UAV120-Gaze and +3.71/+5.17 with box-derived supervision on Anti-UAV410, with gains across three detector families.

Load-bearing premise

The pseudo-gaze transfer scheme assumes the gaze–box offset and reliability distributions estimated from TIR-UAV120-Gaze training boxes (Eq. 35-36, Section IV-A3) generalize to Anti-UAV410 training boxes after target-area pooling and temporal smoothing (Eq. 37-39). If those statistics do not transfer — because of different cameras, target-scale distributions, or search behavior — the 'transferred pseudo-gaze' maps are essentially box-conditional noise, and the pseudo-gaze headline result (87.08 mAP50) does not validate the annotation-transfer claim. The box-derived and real-gaze results could still support a narrower claim.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. Gaze-DETR proposes a DETR-based infrared weak-small UAV detector that learns a normalized spatial priority map from image features, supervised by three schemes: box-derived Gaussian maps, real-gaze fixation-density maps, and transferred pseudo-gaze maps. The predicted map is consumed by two modules: RPFM performs residual priority-guided feature modulation, and PAQI injects additional decoder queries at high-priority peaks. On TIR-UAV120-Gaze (a new paired detection/gaze benchmark) and Anti-UAV410, the method improves over DINO-DETR by about 1.4–3.7 mAP50 across supervision schemes, while retaining image-only inference. The paper reports ablations, attribute-level analysis, priority-map visualizations, and generalization across DINO-DETR, RT-DETR-l, and Deformable-DETR.

Significance. If the mechanism holds, the paper offers a useful new supervision axis for ISTD and introduces a valuable paired eye-tracking/detection benchmark. Strengths include the clear formulation, the image-only inference protocol, the evaluation across three detector families, and a generally conservative discussion of limitations. However, the central mechanistic claim—that the spatial-priority content, rather than added query capacity or a relabeled box signal, drives the gains—is not fully isolated by the current experiments. The absence of a random-anchor control for PAQI, the circularity of the box-derived supervision signal, and the single-run results are load-bearing gaps that need to be addressed before the conclusions can be accepted at face value.

major comments (4)
  1. [§V-F, Table V, §III-D (Eqs. 15-24)] The PAQI component is confounded with added decoder-query capacity. PAQI adds Np = K·R = 6 learnable queries with reference boxes at selected peaks. The ablation G3 vs G1 (and M3 vs M1) therefore changes both the number of queries and their spatial initialization. Because the priority map is detached before peak selection (Eq. 15), PAQI's functional role is exactly to place extra learnable queries at selected locations. Without a control that injects the same number of queries at random or fixed-grid anchors, the reported +1.58/+1.23 mAP50 gains cannot be attributed to the spatial-priority content; they may simply reflect extra decoder capacity or anchor diversity. This control is necessary to validate the central claim that the predicted priority map, not the added queries, drives the improvement.
  2. [§IV-A1, Eqs. (27)-(28)] The box-derived priority-supervision map is a size-adaptive anisotropic Gaussian centered at the same ground-truth box center used by the conventional detection loss. The box-derived scheme thus does not provide an independent spatial signal; it is a blurred relabeling of the box center already present in Ldet. This circularity weakens the claim that the box-derived results demonstrate priority supervision 'complements' bounding-box localization. The real-gaze experiments provide a non-circular test, and the paper should explicitly acknowledge that the box-derived rows in Tables III and V support a narrower interpretation (denser spatial reweighting of box supervision). A control using, e.g., random-offset or edge-derived Gaussian targets would help establish that the dense spatial form—not the box-center position—is the operative ingredient.
  3. [§V-C, Tables III/IV] All experimental results are reported for a single random seed (seed 42). The improvements over DINO-DETR are modest on TIR-UAV120-Gaze (+0.84 to +1.84 mAP50), and several attribute-level deltas in Table IV are negative (e.g., CM mAP50 -0.02, OV F1 -0.23 for box-derived). Without multiple seeds and variance estimates or significance tests, the reader cannot determine whether the headline gains are statistically reliable. Please provide mean±std over at least three seeds, or a clear justification for the single-run protocol.
  4. [§IV-A3, Table III] The transferred pseudo-gaze scheme assumes that the gaze–box offset and reliability statistics estimated from TIR-UAV120-Gaze (Eqs. 35-36) generalize to Anti-UAV410 after pooling and temporal smoothing (Eqs. 37-39). The results show that the transferred pseudo-gaze scheme does not outperform box-derived supervision on Anti-UAV410 (87.08 vs 87.06 mAP50; F1 90.43 vs 90.90). Thus the transfer is not shown to provide gaze-specific information beyond box-derived maps. The paper acknowledges the domain gap in §V-I, but the contribution list still presents pseudo-gaze transfer as a contribution. Please either provide quantitative evidence that the transferred maps encode spatial structure distinct from box-derived Gaussians (e.g., offset distributions or quantitative gaze-like properties) or temper the claim to 'a feasible annotation-transfer procedure' that is currently on par with low-cost bo
minor comments (5)
  1. [§IV-A3, Eq. (34)] The gaze descriptor g_i is defined as (x_g, y_g, v_i, s_i), but the gaze strength s_i is never defined (e.g., pupil confidence, fixation duration, or tracker score). Please define the quantity and its range.
  2. [§III-D, §V-C] The peak count K and the number of size priors R are introduced in Eqs. (17)-(20), but the concrete values K=2 and R=3 appear only in the training details. State these defaults in the method section for clarity.
  3. [§IV-A3, Eq. (37)] The temporal-smoothing parameters ρ=0.45 and the five-frame propagation gap are hyperparameters but are not mentioned in the training-details section. Please list them in §V-C.
  4. [Table III, CHAL row] The CHAL row reports F1=82.9 with one decimal place, while all other F1 values use two decimals. Make the formatting consistent.
  5. [Figure 6(c)] The target-size distribution panel labels the largest bin as '>1600 px^2' in the caption text, but the text and surrounding context indicate the bin should be '>400 px^2'. Please correct this typo.

Circularity Check

0 steps flagged

No significant circularity; real-gaze supervision provides external grounding, while box-derived maps and PAQI priors are acknowledged training-set constructions rather than predictions.

full rationale

The paper's central claim is empirical: adding an explicit priority-supervision loss to a DINO-DETR baseline improves detection. The three supervision schemes are not defined in terms of the target result. The real-gaze scheme (Eqs. 30-32) uses independent eye-tracking annotations, so it is externally grounded and can support the complementarity claim. The transferred pseudo-gaze scheme (Eqs. 35-39) learns gaze-box statistics from TIR-UAV120-Gaze and applies them to Anti-UAV410 training boxes; this depends on a transfer assumption that could be weak, but it is not circular because the gaze-offset distribution comes from paired external gaze data. The box-derived scheme (Eqs. 27-28) constructs its priority map from the same ground-truth boxes used by the detection loss, so it is partly a re-rendering of existing supervision rather than an independent signal; however, the paper explicitly labels it a low-cost scheme and does not rest the conceptual claim on it. PAQI's size priors (Eq. 18) are fitted to training-box statistics, and the absence of a random-anchor control weakens the causal attribution of the PAQI gains, but this is an experimental confound rather than a definitional reduction or a fitted parameter renamed as a prediction. No load-bearing self-citations or imported uniqueness theorems were found. The paper's own limitations discussion acknowledges the domain-gap dependence of pseudo-gaze transfer and the fixed size priors. Overall, the derivation is not circular; the main empirical result has independent content via the real-gaze supervision.

Axiom & Free-Parameter Ledger

11 free parameters · 5 axioms · 2 invented entities

The method leans on standard DETR/Hungarian machinery plus hand-set weights and distributional assumptions. Most supervision hyperparameters are chosen without sensitivity analysis, and the two domain assumptions carrying the gaze story — fixation density equals task priority, and gaze-box statistics transfer across benchmarks — are plausible but unverified. The box-derived scheme is partly a relabeling of the ground-truth boxes. No physical entities are invented; the 'priority map' and 'pseudo-gaze center' are internal or synthetic constructs without independent external validation.

free parameters (11)
  • λp = 0.5
    Priority-loss weight in Eq. 25; set by hand, no sensitivity analysis.
  • λsim = 1
    Weight of histogram-intersection term in Eq. 8; chosen without reported ablation.
  • σp = 8
    Spatial spread of transferred pseudo-gaze Gaussian maps in Eq. 39; chosen by hand.
  • σg = not specified
    Spatial spread for real-gaze coordinate-rendered maps in Eq. 32; value not reported.
  • ρ = 0.45
    Temporal smoothing coefficient for pseudo-gaze offsets in Eq. 37; chosen by hand.
  • K = 2
    Number of PAQI priority peaks selected (Eq. 17); no ablation reported.
  • R = 3
    Number of size priors per peak in Eq. 18; chosen by hand.
  • suppression radius = 2 grid cells
    Local suppression radius in PAQI peak selector; chosen by hand.
  • PAQI warm-up = 4 epochs
    Warm-up period before PAQI activation; chosen by hand.
  • size priors S = from training-set box distribution
    Normalized width-height pairs in Eq. 18 fitted to training boxes; not externally specified.
  • target-area pooling intervals = unspecified
    Grouping of gaze-box observations for empirical pools in Section IV-A3; binning choice not reported.
axioms (5)
  • domain assumption Fixation-density maps from observers searching for UAVs in thermal video are a valid proxy for task-relevant spatial priority.
    Motivates real-gaze supervision in Sections II-B and IV-B2; if gaze reflects bottom-up saliency rather than target search, the 'top-down priority' interpretation is weakened.
  • domain assumption The gaze–box offset and reliability distributions learned from TIR-UAV120-Gaze transfer to Anti-UAV410 training boxes.
    Load-bearing for the pseudo-gaze scheme; Eqs. 35-39 and Section IV-A3; invalid if domains differ in target scale, camera, or search behavior.
  • ad hoc to paper A normalized priority map can be supervised only on target-present frames, while target-absent frames receive zero priority weight yet still use the map at inference.
    Section IV-A2 sets w=0 for target-absent frames; the Discussion acknowledges the no-target state remains unresolved, so PAQI may inject false-positive queries on such frames.
  • ad hoc to paper Box-centered anisotropic Gaussian maps (Eqs. 27-28) are a useful priority target.
    Box-derived supervision heuristic: priority should concentrate at the box center with spread proportional to box size; this is an annotation heuristic, not a measured property.
  • standard math DINO-DETR is a correct and appropriately reproduced baseline.
    The method relies on DINO-DETR (Ref. [46]) as its backbone and comparison point; the exact reproduction is not independently verified here.
invented entities (2)
  • Predicted priority map P̂_i no independent evidence
    purpose: Internal normalized spatial map consumed by RPFM and PAQI to guide pre-localization processing; trained by the priority loss.
    No external metric (e.g., gaze prediction accuracy on held-out fixations) validates the map as a general priority predictor; visualizations are selected examples.
  • Pseudo-gaze center g̃_t no independent evidence
    purpose: Sampled gaze coordinate for Anti-UAV410 boxes, produced by transferring gaze-box offsets from TIR-UAV120-Gaze; used to render pseudo-gaze priority maps.
    It is a synthetic quantity; no human gaze on Anti-UAV410 exists to confirm that sampled pseudo-gaze resembles real search behavior.

pith-pipeline@v1.3.0-alltime-deepseek · 26516 in / 15701 out tokens · 166239 ms · 2026-08-01T13:35:47.886396+00:00 · methodology

0 comments
read the original abstract

Infrared small target detection (ISTD) remains challenging because tiny, low-contrast targets are easily overwhelmed by clutter, noise, or occlusion. Conventional single-frame and multi-frame detectors rely on bounding-box supervision, which specifies final target locations but offers little explicit guidance for prioritizing candidate regions or preserving weak-target evidence before localization. Task-driven visual search offers such guidance: top-down goals and visual evidence jointly form a spatial priority map that ranks candidate locations. Building on this principle, we propose Gaze-DETR, a bio-inspired detector that learns an internal priority map before localization. First, a priority head predicts a normalized priority map from image features. Second, Residual Priority-Guided Feature Modulation (RPFM) enhances high-priority responses while retaining multi-scale features. Finally, Priority-Guided Anchor Query Injection (PAQI) converts high-priority locations into decoder anchor queries. We train the priority head using three supervision schemes: box-derived Gaussian maps; real-gaze maps constructed from fixation-density maps; and transferred pseudo-gaze maps learned from gaze--box relations in paired annotations and applied to Anti-UAV410 training boxes. To support the latter two schemes, we construct TIR-UAV120-Gaze with paired detection and task-driven eye-tracking annotations. On TIR-UAV120-Gaze, Gaze-DETR achieves 85.76 mAP$_{50}$ and 88.77 F1 with box-derived supervision, and 86.18 mAP$_{50}$ and 89.00 F1 with real-gaze supervision. On Anti-UAV410, it achieves 87.06 mAP$_{50}$ and 90.90 F1 with box-derived supervision, and 87.08 mAP$_{50}$ and 90.43 F1 with transferred pseudo-gaze supervision. These results show that explicit spatial-priority learning provides pre-localization guidance complementary to bounding-box supervision across annotation settings and costs.

Figures

Figures reproduced from arXiv: 2607.19040 by Boyu Cai, Jin Gao, Liang Li, Nian Liu, Shubo Lin, Sikui Zhang, Weiming Hu, Yizheng Wang, Yuxin Yang.

Figure 1
Figure 1. Figure 1: Motivation for the proposed priority-supervision formulation. (a) Infrared weak-small UAV detection faces occlusion, clutter, low contrast, and apparent disappearance. (b) Box-only supervision provides final localization labels but limited pre-localization priority guidance. (c) Priority supervision complements box supervision by adding a training-time priority loss; the real-gaze scheme is shown as an exa… view at source ↗
Figure 2
Figure 2. Figure 2: Effectiveness of the evaluated priority-supervision settings. Box-derived, [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview and main components of Gaze-DETR. (a) The backbone extracts multi-scale image features, from which the priority head produces the ˆ [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Construction of transferred pseudo-gaze priority-supervision maps for Anti-UAV410. Stage I estimates box-conditioned gaze-offset and reliability [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Representative TIR-UAV120-Gaze samples under seven challenge attributes: camera motion (CM), thermal crossover (TC), out-of-view (OV), occlusion [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Test-set statistics of TIR-UAV120-Gaze. (a) Occurrence frequency of the seven multi-label challenge attributes. (b) Occurrence frequency of the four [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison on challenging TIR-UAV120-Gaze scenes using the real-gaze-supervised Gaze-DETR. Rows show ground truth, Gaze-DETR, [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison on Anti-UAV410 using the pseudo-gaze-supervised Gaze-DETR. Each column shows one infrared frame, with ground truth [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Precision–recall curves for the evaluated priority-supervision settings. Panel (a) reports real-gaze supervision on TIR-UAV120-Gaze and transferred [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visualization of priority-head predictions on TIR-UAV120-Gaze and Anti-UAV410. The upper panel shows results from the real-gaze-supervised [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 1 canonical work pages

  1. [1]

    Dense nested attention network for infrared small target detection,

    B. Li, C. Xiao, L. Wang, Y . Wang, Z. Lin, M. Li, W. An, and Y . Guo, “Dense nested attention network for infrared small target detection,”IEEE Transactions on Image Processing, vol. 32, pp. 1745–1758, 2023

    2023

  2. [2]

    Isnet: Shape matters for infrared small target detection,

    M. Zhang, R. Zhang, Y . Yang, H. Bai, J. Zhang, and J. Guo, “Isnet: Shape matters for infrared small target detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 877–886

    2022

  3. [3]

    A real-time robust approach for tracking uavs in infrared videos,

    H. Wu, W. Li, W. Li, and G. Liu, “A real-time robust approach for tracking uavs in infrared videos,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2020

    2020

  4. [4]

    Anti-UA V: A large-scale benchmark for vision-based UA V tracking,

    N. Jiang, K. Wang, X. Peng, X. Yu, Q. Wang, J. Xing, G. Li, Q. Ye, J. Jiao, Z. Hanet al., “Anti-UA V: A large-scale benchmark for vision-based UA V tracking,”IEEE Transactions on Multimedia, 2021

    2021

  5. [5]

    Anti-UA V410: A thermal infrared benchmark and customized scheme for tracking drones in the wild,

    B. Huang, J. Li, J. Chen, G. Wang, J. Zhao, and T. Xu, “Anti-UA V410: A thermal infrared benchmark and customized scheme for tracking drones in the wild,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 5, pp. 2852–2865, May 2024. [Online]. Available: https://doi.ieeecomputersociety.org/10.1109/TPAMI.2023.3335338

    arXiv 2024

  6. [6]

    Evidential detection and tracking collaboration: New problem, benchmark and algorithm for robust anti-UA V system,

    X. Zhu, T. Xu, J. Zhao, J. Liu, K. Wang, G. Wang, J. Li, Z. Zhang, Q. Wang, L. Jinet al., “Evidential detection and tracking collaboration: New problem, benchmark and algorithm for robust anti-UA V system,” arXiv preprint arXiv:2306.15767, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Mor- phological clutter removal for detection of dim targets,

    S. D. Deshpande, M. H. Er, R. Venkateswarlu, and P. Chan, “Mor- phological clutter removal for detection of dim targets,”Digital Signal Processing, vol. 9, no. 2, pp. 84–96, 1999

    1999

  8. [8]

    Max-mean and max-median filters for detection of small targets,

    ——, “Max-mean and max-median filters for detection of small targets,” Signal and Data Processing of Small Targets 1999, vol. 3809, pp. 74–83, 1999

    1999

  9. [9]

    A local contrast method for small infrared target detection,

    C. P. Chen, H. Li, Y . Wei, T. Xia, and Y . Y . Tang, “A local contrast method for small infrared target detection,”IEEE Transactions on Geoscience and Remote Sensing, vol. 52, no. 1, pp. 574–581, 2013

    2013

  10. [10]

    Infrared patch-image model for small target detection in a single image,

    C. Gao, D. Meng, Y . Yang, Y . Wang, X. Zhou, and A. G. Hauptmann, “Infrared patch-image model for small target detection in a single image,” IEEE transactions on image processing, vol. 22, no. 12, pp. 4996–5009, 2013

    2013

  11. [11]

    Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection,

    Y . Dai and Y . Wu, “Reweighted infrared patch-tensor model with both nonlocal and local priors for single-frame small target detection,”IEEE journal of selected topics in applied earth observations and remote sensing, vol. 10, no. 8, pp. 3752–3767, 2017

    2017

  12. [12]

    Small target detection in infrared videos based on spatio-temporal tensor model,

    H.-K. Liu, L. Zhang, and H. Huang, “Small target detection in infrared videos based on spatio-temporal tensor model,”IEEE Transactions on Geoscience and Remote Sensing, vol. 58, no. 12, pp. 8689–8700, 2020

    2020

  13. [13]

    Asymmetric contextual modulation for infrared small target detection,

    Y . Dai, Y . Wu, F. Zhou, and K. Barnard, “Asymmetric contextual modulation for infrared small target detection,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2021, pp. 950–959

    2021

  14. [14]

    Attentional local contrast networks for infrared small target detection,

    ——, “Attentional local contrast networks for infrared small target detection,”IEEE Transactions on Geoscience and Remote Sensing, vol. 59, no. 11, pp. 9813–9824, 2021

    2021

  15. [15]

    Uiu-net: U-net in u-net for infrared small object detection,

    X. Wu, D. Hong, and J. Chanussot, “Uiu-net: U-net in u-net for infrared small object detection,”IEEE Transactions on Image Processing, vol. 32, pp. 364–376, 2023

    2023

  16. [16]

    Infrared small target detection with scale and location sensitivity,

    Q. Liu, R. Liu, B. Zheng, H. Wang, and Y . Fu, “Infrared small target detection with scale and location sensitivity,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 17 490–17 499

    2024

  17. [17]

    SCTransNet: Spatial- channel cross transformer network for infrared small target detection,

    S. Yuan, H. Qin, X. Yan, N. Akhtar, and A. Mian, “SCTransNet: Spatial- channel cross transformer network for infrared small target detection,” IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–15, 2024

    2024

  18. [18]

    Saliency at the helm: Steering infrared small target detection with learnable kernels,

    F. Wu, A. Liu, T. Zhang, L. Zhang, J. Luo, and Z. Peng, “Saliency at the helm: Steering infrared small target detection with learnable kernels,” IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1–14, 2025

    2025

  19. [19]

    Stdmanet: Spatio-temporal differential multiscale attention network for small moving infrared target detection,

    P. Yan, R. Hou, X. Duan, C. Yue, X. Wang, and X. Cao, “Stdmanet: Spatio-temporal differential multiscale attention network for small moving infrared target detection,”IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–16, 2023

    2023

  20. [20]

    Direction- coded temporal u-shape module for multiframe infrared small target detection,

    R. Li, W. An, C. Xiao, B. Li, Y . Wang, M. Li, and Y . Guo, “Direction- coded temporal u-shape module for multiframe infrared small target detection,”IEEE Transactions on Neural Networks and Learning Systems, pp. 1–14, 2023

    2023

  21. [21]

    SSTNet: Sliced spatio- temporal network with cross-slice convlstm for moving infrared dim- small target detection,

    S. Chen, L. Ji, J. Zhu, M. Ye, and X. Yao, “SSTNet: Sliced spatio- temporal network with cross-slice convlstm for moving infrared dim- small target detection,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–12, 2024

    2024

  22. [22]

    Triple-domain feature learning with frequency-aware memory enhancement for moving infrared small target detection,

    W. Duan, L. Ji, S. Chen, S. Zhu, and M. Ye, “Triple-domain feature learning with frequency-aware memory enhancement for moving infrared small target detection,”IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–14, 2024. [Online]. Available: http://dx.doi.org/10.1109/TGRS.2024.3452175

    arXiv 2024

  23. [23]

    Probing deep into temporal profile makes the infrared small target detector much better,

    R. Li, W. An, X. Ying, Y . Wang, Y . Dai, L. Wang, M. Li, Y . Guo, and L. Liu, “Probing deep into temporal profile makes the infrared small target detector much better,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 48, no. 8, pp. 10 157–10 175, 2026

    2026

  24. [24]

    Motion prior knowledge learning with homogeneous language descriptions for moving infrared IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 19 small target detection,

    S. Chen, L. Ji, W. Duan, S. Peng, and M. Ye, “Motion prior knowledge learning with homogeneous language descriptions for moving infrared IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 19 small target detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 2, 2025, pp. 2186–2194

    2025

  25. [25]

    Saist: Segment any infrared small target model guided by contrastive language-image pretraining,

    M. Zhang, X. Li, F. Gao, J. Guo, X. Gao, and J. Zhang, “Saist: Segment any infrared small target model guided by contrastive language-image pretraining,” in2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025, pp. 9549–9558

    2025

  26. [26]

    Sevil: Semi-supervised vision- language learning with text prompt guiding for moving infrared small target detection,

    W. Duan, L. Ji, J. Huang, and S. Zhu, “Sevil: Semi-supervised vision- language learning with text prompt guiding for moving infrared small target detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 5, 2026, pp. 3723–3731

    2026

  27. [27]

    Optimal eye movement strategies in visual search,

    J. Najemnik and W. S. Geisler, “Optimal eye movement strategies in visual search,”Nature, vol. 434, no. 7031, pp. 387–391, 2005

    2005

  28. [28]

    Contex- tual guidance of eye movements and attention in real-world scenes: The role of global features in object search,

    A. Torralba, A. Oliva, M. S. Castelhano, and J. M. Henderson, “Contex- tual guidance of eye movements and attention in real-world scenes: The role of global features in object search,”Psychological Review, vol. 113, no. 4, pp. 766–786, 2006

    2006

  29. [29]

    Salience, relevance, and firing: A priority map for target selection,

    J. H. Fecteau and D. P. Munoz, “Salience, relevance, and firing: A priority map for target selection,”Trends in Cognitive Sciences, vol. 10, no. 8, pp. 382–390, 2006

    2006

  30. [30]

    Guided search 6.0: An updated model of visual search,

    J. M. Wolfe, “Guided search 6.0: An updated model of visual search,” Psychonomic Bulletin & Review, vol. 28, no. 4, pp. 1060–1092, 2021

    2021

  31. [31]

    A model of saliency-based visual attention for rapid scene analysis,

    L. Itti, C. Koch, and E. Niebur, “A model of saliency-based visual attention for rapid scene analysis,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 20, no. 11, pp. 1254–1259, 1998

    1998

  32. [32]

    Irsam: Advancing segment anything model for infrared small target detection,

    M. Zhang, Y . Wang, J. Guo, Y . Li, X. Gao, and J. Zhang, “Irsam: Advancing segment anything model for infrared small target detection,” inComputer Vision – ECCV 2024, 2024

    2024

  33. [33]

    A benchmark of computational models of saliency to predict human fixations,

    T. Judd, F. Durand, and A. Torralba, “A benchmark of computational models of saliency to predict human fixations,” Massachusetts Institute of Technology, Tech. Rep., 2012

    2012

  34. [34]

    SALICON: Reducing the semantic gap in saliency prediction by adapting deep neural networks,

    X. Huang, C. Shen, X. Boix, and Q. Zhao, “SALICON: Reducing the semantic gap in saliency prediction by adapting deep neural networks,” inProceedings of the IEEE International Conference on Computer Vision (ICCV), 2015, pp. 262–270

    2015

  35. [35]

    Revisiting video saliency: A large-scale benchmark and a new model,

    W. Wang, J. Shen, F. Guo, M.-M. Cheng, and A. Borji, “Revisiting video saliency: A large-scale benchmark and a new model,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 4894–4903

    2018

  36. [36]

    Deepgaze iii: Modeling free-viewing human scanpaths with deep learning,

    M. Kümmerer, M. Bethge, and T. S. A. Wallis, “Deepgaze iii: Modeling free-viewing human scanpaths with deep learning,”Journal of Vision, vol. 22, no. 5, p. 7, 2022

    2022

  37. [37]

    A. L. Yarbus,Eye Movements and Vision. New York: Plenum Press, 1967

    1967

  38. [38]

    Holmqvist, M

    K. Holmqvist, M. Nyström, R. Andersson, R. Dewhurst, H. Jarodzka, and J. Van de Weijer,Eye Tracking: A Comprehensive Guide to Methods and Measures. Oxford University Press, 2011

    2011

  39. [39]

    COCO-Search18 fixation dataset for predicting goal-directed attention control,

    Y . Chen, Z. Yang, S. Ahn, D. Samaras, M. Hoai, and G. J. Zelinsky, “COCO-Search18 fixation dataset for predicting goal-directed attention control,”Scientific Reports, vol. 11, no. 1, pp. 1–11, 2021

    2021

  40. [40]

    Observational supervision for medical image classification using gaze data,

    K. Saab, J. Dunnmon, R. Goldman, C. Lee-Messer, D. Rubin, and C. Re, “Observational supervision for medical image classification using gaze data,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention, 2021, pp. 603–614

    2021

  41. [41]

    Learning from observer gaze: Zero-shot attention prediction oriented by human-object interaction recognition,

    Y . Zhou, L. Liu, and C. Gou, “Learning from observer gaze: Zero-shot attention prediction oriented by human-object interaction recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 28 402–28 412

    2024

  42. [42]

    A simultaneous EEG and eye-tracking dataset for remote sensing object detection,

    B. He, H. Zhang, T. Qin, B. Shi, Q. Wang, and W. Dong, “A simultaneous EEG and eye-tracking dataset for remote sensing object detection,” Scientific Data, vol. 12, no. 1, p. 651, 2025

    2025

  43. [43]

    EyeTrackUA V2: A large-scale binocular eye-tracking dataset for UA V videos,

    A.-F. Perrin, V . Krassanakis, L. Zhang, V . Ricordel, M. Perreira Da Silva, and O. Le Meur, “EyeTrackUA V2: A large-scale binocular eye-tracking dataset for UA V videos,”Drones, vol. 4, no. 1, p. 2, 2020. [Online]. Available: https://doi.org/10.3390/drones4010002

    work page doi:10.3390/drones4010002 2020

  44. [44]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in European Conference on Computer Vision (ECCV). Springer, 2020, pp. 213–229

    2020

  45. [45]

    Deformable DETR: Deformable transformers for end-to-end object detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable DETR: Deformable transformers for end-to-end object detection,” inInternational Conference on Learning Representations (ICLR), 2021

    2021

  46. [46]

    DINO: DETR with improved denoising anchor boxes for end-to-end object detection,

    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,” inInternational Conference on Learning Representations (ICLR), 2023

    2023

  47. [47]

    DETRs beat YOLOs on real-time object detection,

    Y . Zhao, W. Lv, S. Xu, J. Wei, G. Wang, Q. Dang, Y . Liu, and J. Chen, “DETRs beat YOLOs on real-time object detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 16 965–16 974

    2024

  48. [48]

    Generalized intersection over union: A metric and a loss for bounding box regression,

    H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese, “Generalized intersection over union: A metric and a loss for bounding box regression,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 658–666

    2019

  49. [49]

    MIT/Tuebingen Saliency Benchmark: Datasets,

    “MIT/Tuebingen Saliency Benchmark: Datasets,” https://saliency. tuebingen.ai/datasets.html, accessed: 2026-06-05

    2026

  50. [50]

    CAT2000: A large scale fixation dataset for boosting saliency research,

    A. Borji and L. Itti, “CAT2000: A large scale fixation dataset for boosting saliency research,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2015

    2015

  51. [51]

    Unifying top-down and bottom-up scanpath prediction using transformers,

    Z. Yang, S. Mondal, S. Ahn, R. Xue, G. J. Zelinsky, M. Hoai, and D. Samaras, “Unifying top-down and bottom-up scanpath prediction using transformers,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 1683–1693

    2024

  52. [52]

    YOLOv6: A single-stage object detection framework for industrial applications,

    C. Li, L. Li, H. Jiang, K. Weng, Y . Geng, L. Li, Z. Ke, Q. Li, M. Cheng, W. Nie, Y . Li, B. Zhang, Y . Liang, L. Zhou, X. Xu, X. Chu, X. Wei, and X. Wei, “YOLOv6: A single-stage object detection framework for industrial applications,”arXiv preprint arXiv:2209.02976, 2022

    Pith/arXiv arXiv 2022

  53. [53]

    YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,

    C.-Y . Wang, A. Bochkovskiy, and H.-Y . M. Liao, “YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors,” arXiv preprint arXiv:2207.02696, 2022

    Pith/arXiv arXiv 2022

  54. [54]

    Ultralytics YOLOv8,

    G. Jocher, A. Chaurasia, and J. Qiu, “Ultralytics YOLOv8,” https://github. com/ultralytics/ultralytics, 2023

    2023

  55. [55]

    YOLOv9: Learning what you want to learn using programmable gradient information,

    C.-Y . Wang, I.-H. Yeh, and H.-Y . M. Liao, “YOLOv9: Learning what you want to learn using programmable gradient information,”arXiv preprint arXiv:2402.13616, 2024

    Pith/arXiv arXiv 2024

  56. [56]

    YOLOv10: Real-time end-to-end object detection,

    A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han, and G. Ding, “YOLOv10: Real-time end-to-end object detection,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  57. [57]

    Ultralytics YOLO11,

    Ultralytics, “Ultralytics YOLO11,” https://docs.ultralytics.com/models/ yolo11/, 2024

    2024

  58. [58]

    YOLOv12: Attention-centric real-time object detectors,

    Y . Tian, Q. Ye, and D. Doermann, “YOLOv12: Attention-centric real-time object detectors,”arXiv preprint arXiv:2502.12524, 2025

    Pith/arXiv arXiv 2025

  59. [59]

    YOLOv13: Real-time object detection with hypergraph- enhanced adaptive visual perception,

    M. Lei, S. Li, Y . Wu, H. Hu, Y . Zhou, X. Zheng, G. Ding, S. Du, Z. Wu, and Y . Gao, “YOLOv13: Real-time object detection with hypergraph- enhanced adaptive visual perception,”arXiv preprint arXiv:2506.17733, 2025

    Pith/arXiv arXiv 2025

  60. [60]

    Pinwheel- shaped convolution and scale-based dynamic loss for infrared small target detection,

    J. Yang, S. Liu, J. Wu, X. Su, N. Hai, and X. Huang, “Pinwheel- shaped convolution and scale-based dynamic loss for infrared small target detection,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 9, 2025, pp. 9202–9210

    2025

  61. [61]

    Ultralytics YOLO26: Unified real-time end-to-end vision models,

    G. Jocher, J. Qiu, M. Liu, S. Lyu, F. C. Akyon, and M. E. Kalfaoglu, “Ultralytics YOLO26: Unified real-time end-to-end vision models,” 2026. [Online]. Available: https://arxiv.org/abs/2606.03748

    Pith/arXiv arXiv 2026

  62. [62]

    CHAL: Causal- guided hierarchical anomaly-aware learning for moving infrared small target detection,

    W. Duan, L. Ji, S. Lei, S. Zhu, J. Huang, and M. Ye, “CHAL: Causal- guided hierarchical anomaly-aware learning for moving infrared small target detection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2026, pp. 21 357– 21 366

    2026