CST Anti-UAV: Tiny Drone Thermal Benchmark

• CST Anti-UAV is a thermal infrared benchmark designed for single-object tracking of tiny UAVs amid cluttered urban and dynamic scenes.
• It addresses gaps in earlier datasets by providing rigorous frame-level annotations and detailed attribute challenges like occlusion and dynamic background clutter.
• Evaluation using OPE metrics reveals significant performance drops for current trackers, highlighting the need for more robust tracking methods.

CST Anti-UAV is a thermal infrared benchmark for single-object tracking of tiny unmanned aerial vehicles in complex scenes. The benchmark was introduced to address a gap in earlier anti-UAV datasets, which the paper characterizes as insufficiently representative of the conditions that make operational anti-drone perception difficult: tiny targets, complex urban and dynamic backgrounds, and fine-grained frame-level challenge annotation. Within the anti-UAV literature, it is positioned as a more demanding evaluation setting for thermal infrared UAV tracking, and the reported benchmark results show that current trackers remain far from robust performance under these conditions (Xie et al., 31 Jul 2025).

1. Problem definition and benchmark rationale

CST Anti-UAV is formulated as a thermal infrared single-object tracking benchmark for Single Object Tracking (SOT) in Complex Scenes with Tiny UAVs (CST). Its central premise is that existing anti-UAV tracking datasets often contain comparatively conspicuous targets and scenes that are cleaner than those encountered in practical counter-UAV surveillance. The benchmark therefore emphasizes two properties simultaneously: a significant concentration of tiny UAV targets and complex real-world scene structure (Xie et al., 31 Jul 2025).

The problem addressed by CST Anti-UAV is not generic object tracking, but anti-UAV tracking under conditions where the target may occupy very few pixels, appear against urban or operational clutter, become partially visible, disappear from the field of view, or be confounded by moving distractors such as pedestrians, vehicles, birds, and airplanes. This places the benchmark within the broader Anti-UAV benchmark lineage initiated by the multi-modal Anti-UAV dataset, which established UAV tracking as a state-aware task with visibility reasoning rather than pure box regression (Jiang et al., 2021).

A recurring misconception in anti-UAV benchmarking is that strong results on earlier datasets imply operational readiness. CST Anti-UAV is designed explicitly against that assumption. The benchmark shows that trackers that perform well on earlier benchmarks can degrade sharply when scene complexity and tiny-target prevalence are increased, suggesting that prior benchmark saturation did not translate into robust tiny-UAV tracking capability (Xie et al., 31 Jul 2025).

2. Dataset composition and scale regime

The dataset contains 220 video sequences with over 240k high-quality bounding box annotations. Sequence lengths range from 600 to 2,062 frames, with an average of 1,132 frames. The imagery is thermal infrared at 25 fps and 640 × 512 resolution. The split is 120 training sequences, 40 validation sequences, and 60 test sequences (Xie et al., 31 Jul 2025).

A defining characteristic of CST Anti-UAV is its object-scale distribution, which is based on bounding-box diagonal length. The benchmark defines tiny as $(0,10]$, small as $(10,30]$, normal as $(30,50]$, and large as $(50,\inf]$. Under this definition, the dataset contains 78,224 tiny objects, 155,202 small objects, 11,208 normal objects, and 837 large objects. The distribution is therefore strongly skewed toward tiny and small targets, which is a primary source of difficulty (Xie et al., 31 Jul 2025).

The benchmark is presented as differing from earlier anti-UAV datasets in both scale regime and scene complexity. In the paper’s comparison, Anti-UAV is described as having 318 sequences and 293k boxes, but with mostly relatively large objects and simple or clean scenes; Anti-UAV410 is described as having 410 sequences and 438k boxes, with some tiny objects and wild scenes but still limited tiny-object proportion and scenes that are not truly complex urban environments. CST Anti-UAV, despite having 220 sequences and 240k boxes, is organized around a much larger concentration of tiny UAVs and richer scene complexity (Xie et al., 31 Jul 2025).

This design extends the anti-UAV benchmark trajectory in a specific direction. The original Anti-UAV benchmark emphasized large-scale RGB-T tracking with state accuracy and unaligned modalities (Jiang et al., 2021), while the 3rd Anti-UAV Workshop & Challenge expanded thermal infrared tracking and introduced a detection-and-tracking track (Zhao et al., 2023). CST Anti-UAV narrows its focus to thermal infrared SOT, but increases the difficulty by concentrating on tiny targets in complex scenes rather than on benchmark breadth alone (Xie et al., 31 Jul 2025).

3. Annotation protocol and frame-level challenge model

The annotation protocol is one of the benchmark’s principal contributions. The paper reports over 2,000 hours spent on annotation. Bounding boxes were manually drawn for every frame, and a separate verification team performed frame-by-frame checking until issues were resolved. For attributes, one expert annotator labeled all frames to maintain consistency (Xie et al., 31 Jul 2025).

CST Anti-UAV provides 1440k frame-level attribute annotations and 220 × 6 sequence-level attribute annotations. The six annotated attributes are OC (Occlusion), OV (Out-of-View), SV (Scale Variation), TC (Thermal Crossover), DBC (Dynamic Background Clutter), and CDB (Complex Dynamic Background). The paper defines them as follows: OC indicates that the target is partially or fully occluded; OV indicates that the target leaves the view; SV indicates that the ratio of the first-frame box and current box is outside $[0.66, 1.5]$; TC indicates that the target has similar temperature to surrounding objects or background; DBC indicates dynamic changes in the background around the target; and CDB indicates multiple dynamic non-target objects in the background (Xie et al., 31 Jul 2025).

The benchmark is presented as the first UAV benchmark with full manual frame-level attribute annotation. That claim is methodologically important because prior datasets mainly used sequence-level labels. A sequence-level label can indicate that a challenge occurs somewhere in a video, but it does not localize which frames actually instantiate the challenge. CST Anti-UAV is designed to support frame-specific analysis of tracker behavior under changing conditions, particularly when a single sequence contains alternating periods of easy visibility, thermal crossover, out-of-view intervals, and cluttered distractor motion (Xie et al., 31 Jul 2025).

The introduction of CDB is especially notable. The paper argues that scenes with multiple dynamic non-target objects are common in real anti-UAV environments but insufficiently represented in prior datasets. This suggests a shift from simple background-clutter notions toward a stronger distractor model in which several independently moving entities compete with the tiny UAV for tracker attention (Xie et al., 31 Jul 2025).

4. Evaluation protocol and metric structure

CST Anti-UAV uses One-Pass Evaluation (OPE) and reports three standard metrics: precision plot, success plot, and state accuracy. The success and precision plots follow the standard OTB protocol. For state accuracy, the benchmark adopts the following expression:

$$SA = \sum_{t} \frac{I O U_{t} \times \delta\left(v_{t}>0\right) + p_{t} \times \left(1 - \delta\left(v_{t}>0\right)\right)}{T}.$$

Here, $IOU_t$ is the intersection-over-union between prediction and ground truth, $v_t$ is the visibility flag of the ground truth, $p_t$ is the tracker’s predicted state, and $T$ is the total number of frames. The benchmark reports mSA, P(AUC), and S(AUC) (Xie et al., 31 Jul 2025).

The use of state accuracy places CST Anti-UAV in continuity with earlier Anti-UAV benchmarks, where the task explicitly includes reasoning about whether the UAV is present or absent rather than only localizing it when visible (Jiang et al., 2021). In anti-UAV tracking, this distinction matters because out-of-view periods, temporary disappearance, and re-entry are operationally significant; a tracker that always predicts a box is not necessarily behaving correctly.

The paper emphasizes that frame-level evaluation is substantially more informative than sequence-level attribute evaluation. Sequence-level labels can hide the real difficulty of challenge-specific frames, whereas frame-level labels allow a tracker’s failure modes to be isolated more precisely. This is particularly relevant for OV, where a sequence may contain only a short but decisive disappearance interval, and for TC or CDB, where the challenge intensity may fluctuate sharply over time (Xie et al., 31 Jul 2025).

5. Empirical findings from tracker benchmarking

The authors retrained and evaluated 20 existing SOT methods on both CST Anti-UAV and Anti-UAV410. The headline result is that performance drops sharply on CST Anti-UAV. On Anti-UAV410, the best reported result is SiamDT* with 67.69% mSA, 87.63% P(AUC), and 66.34% S(AUC). On CST Anti-UAV, the best reported result is GlobalTrack with 35.92% mSA, 58.72% P(AUC), and 35.38% S(AUC), while SiamDT* is nearly identical at 35.84% mSA, 57.96% P(AUC), and 35.28% S(AUC) (Xie et al., 31 Jul 2025).

The degradation is substantial across multiple trackers. The paper highlights GlobalTrack: 66.42 → 35.92 mSA, SiamDT*: 67.69 → 35.84 mSA, and MixFormer-V2: 61.68 → 29.39 mSA. Many other methods fall into the mid-20s on CST. This suggests that model capacity or architectural modernity alone is not sufficient when the target is both tiny and embedded in complex thermal clutter (Xie et al., 31 Jul 2025).

The benchmark also includes a training-transfer analysis. When trackers are trained on Anti-UAV410 and evaluated on CST, performance is much worse than when they are trained on CST itself. The paper gives examples including GlobalTrack: 30.96 mSA when trained on 410, vs. 35.92 when trained on CST, SeqTrack: 20.36 vs. 26.53, ROMTrack: 21.13 vs. 27.43, AQATrack: 23.12 vs. 27.60, and SiamDT*: 30.95 vs. 35.84. It further reports that 19 out of 22 trackers improved by more than 10% when trained on CST rather than Anti-UAV410, with the maximum relative gain reaching 30.3% (Xie et al., 31 Jul 2025).

Attribute-level analysis identifies the most difficult conditions as DBC, TC, CDB, OV, and tiny-sized objects. The paper reports a particularly strong failure mode under OV, where performance can peak at 65.77% but fall as low as 14.53%. This is interpreted as evidence that re-localizing the target after it leaves and re-enters the view is materially harder than handling ordinary occlusion (Xie et al., 31 Jul 2025).

The benchmark also differentiates tracker behavior by design choice. GlobalTrack and SiamDT are described as strong across many attributes because their global search helps re-localize lost targets, although global search can hurt performance on tiny targets due to extra background noise. KYS is reported as performing well on out-of-view and tiny-object cases because of dense state propagation and scene information. PrDiMP is described as especially strong on occlusion, out-of-view, complex dynamic backgrounds, and tiny objects due to probabilistic regression (Xie et al., 31 Jul 2025).

6. Significance, limitations, and benchmark position in anti-UAV research

CST Anti-UAV is significant chiefly because it exposes the gap between benchmark competence and operational robustness in anti-UAV tracking. The benchmark indicates that existing trackers generalize poorly to tiny UAVs in complex scenes, that performance is much lower than on previous benchmarks, and that frame-level annotation is essential for meaningful analysis of tracker failure (Xie et al., 31 Jul 2025).

Within the broader anti-UAV research landscape, CST Anti-UAV occupies the single-object tracking side of a larger benchmark ecosystem. The Anti-UAV line began with multi-modal RGB-T tracking and state-aware evaluation (Jiang et al., 2021); the Anti-UAV Workshop and Challenge series expanded thermal infrared evaluation and introduced distinct tracking and detection-and-tracking tracks (Zhao et al., 2023); AntiUAV600 then formalized continuous detection and tracking without first-frame target initialization, using evidential collaboration between detection and tracking (Zhu et al., 2023). CST Anti-UAV does not replace those settings. Rather, it intensifies one specific regime: thermal infrared SOT where tiny target size and complex dynamic scenes dominate the error profile.

The benchmark also clarifies a second misconception: frame-level annotation is not merely a more detailed version of sequence-level labeling. The paper’s argument is stronger than that. Frame-level labels change the type of analysis that becomes possible, because they allow evaluation on the exact frames where thermal crossover, dynamic distractors, or out-of-view transitions occur. A plausible implication is that future tracker development for anti-UAV tasks will increasingly depend on challenge-conditioned training and diagnosis, rather than on aggregate sequence scores alone (Xie et al., 31 Jul 2025).

The main limitations reported are also important. The benchmark focuses on single-object tracking rather than detection-and-tracking, multimodal fusion, or multi-UAV association. It demonstrates that current methods are not yet robust enough to handle the combination of tiny targets, complex dynamic backgrounds, occlusion, disappearance and reappearance, thermal crossover, and large scale change. The future directions implied by the paper are better long-term re-localization, tracking methods robust to tiny targets, stronger handling of dynamic backgrounds and multiple distractors, more effective use of temporal information, and continued use of fine-grained benchmarked evaluation using frame-level attributes (Xie et al., 31 Jul 2025).

In that sense, CST Anti-UAV functions less as an incremental dataset expansion than as a stress test for anti-UAV tracking methodology. Its principal encyclopedic importance lies in making the tiny-UAV, complex-scene regime measurable in a systematic way, and in showing that benchmark progress on earlier datasets did not resolve the underlying anti-UAV tracking problem (Xie et al., 31 Jul 2025).