What Limits Thermal Infrared Drone Detection? A Cross-Dataset Benchmark Analysis
Authors: Tamara R. Lenhard, Jan Kels, Tobias Koch
Organizations: Institute for the Protection of Terrestrial Infrastructures, German Aerospace Center (DLR), Sankt Augustin, Germany · ACIDA Lab, Technical University of Applied Sciences Würzburg-Schweinfurt, Schweinfurt, Germany · EUT+ Data Science Institute, European University of Technology, European Union
Thermal infrared (IR) imaging enables drone detection with reduced dependence on ambient illumination, supporting continuous surveillance beyond the limitations of visible-spectrum sensing. However, substantial and often undocumented variability in camera systems, processing pipelines, and annotation practices across public IR drone datasets complicates the assessment of detector performance and cross-domain generalization. To examine these sources of variability and their implications for drone detection, we systematically characterize existing IR datasets and establish an artifact taxonomy reflecting their origins across the acquisition and processing pipeline, encompassing optical and sensor effects, tone mapping and image enhancement, and recording-related artifacts. Focusing on long-wave IR data, we investigate the impact of training data composition and model characteristics on YOLO-based detection performance, generalization, and condition-dependent failure modes. Our results demonstrate that detection performance is governed more strongly by training data composition than by YOLO generation or model capacity, while cross-dataset generalization exhibits pronounced asymmetries. Even the best-performing configuration fails to detect over 40% of drones, with elevated miss rates for very small targets, cluttered urban backgrounds, and thermal crossover, highlighting the shortcomings of existing datasets.
Figures & tables
Dataset
Pub. Year
No. Seq.
No. Frames
Resolution
IR Band
OSD
Image Quality / Visual Artifacts
(k)
(px)
LRDDv3 ◆ [ 19 ]
2026
38
29.6
640 × 512
LWIR
✗
shaprness, blur, pseudocolor mapping
CST Anti-UAV [ 31 ]
2025
220
238
640 × 512
LWIR ★
✓
sharpening & compression artifacts
Anti-UAV410 [ 8 ]
2023
410
438 ▲
640 × 512
MWIR
✓
lens reflection, compression, motion blur
Hamburg2 (H2) [ 2 ]
2023
23
65
640 × 480
LWIR
✗
compression artifacts, watermark,
1024 × 768
vignetting, motion blur, defocus blur
Table 1: Overview of publicly available IR datasets for drone detection and tracking. Sequence and frame counts refer exclusively to IR data. Unreported IR-bands were inferred by visual inspection. ✓/✗ denote the presence/absence of on-screen display (OSD).
Figure 1: Selected examples of artifacts along the imaging pipeline in public IR drone detection and tracking datasets (left to right): Anti-UAV [ 10 ] , USC [ 4 ] , LRDDv3 [ 19 ] , InfraTec [ 2 ] , and CST [ 31 ] . Top: full-frame images with crop regions outlined in red (distinct from GT bounding boxes). Bottom: corresponding detail crops. Colors indicate stage of the imaging pipeline at which each artifact originates.
Model
Scale
Params (M)
FLOPs (B)
YOLOv8
n
3.2
8.7
s
11.2
28.6
m
25.9
78.9
l
43.7
165.2
YOLO11
s
9.4
21.5
m
20.1
68.0
Table 2: Capacity and computational complexity of evaluated YOLO configurations. Parameter counts and FLOPs refer to COCO-pretrained models at an input resolution of 640 × 640.
Dataset
Subsets
Split
train
val
test
H2 [ 2 ]
FLIR
26,368
–
14,251
InfraTec
23,373
–
2,446
Total
49,741
–
16,697
LRDDv3 [ 19 ]
–
18,460
3,379
7,623
CST [ 31 ]
–
124,301
43,527
70,214
Table 3: Frames per dataset and split. FLIR and InfraTec denote disjoint H2 camera subsets. Test frames are sequence-disjoint from training and validation sets. ”–” indicates not applicable.
Split
Background Category
building
mountain
sky
jungle
urban
water
train
53,648
20,230
8,470
4,908
26,177
10,868
val
13,974
9,659
2,231
1,197
11,670
4,796
test
8,458
10,687
6,611
13,651
26,004
4,803
Table 4: Frame counts per background category across the CST [ 31 ] splits. Categories are defined on a per-sequence basis, with each frame assigned to exactly one category.
Figure 2: Proportion of annotated drone instances exhibiting TC, DBC, and CDB across all splits of CST [ 31 ] .
Figure 3: Normalized object-size distributions for H2 [ 2 ] , LRDDv3 [ 19 ] , and CST [ 31 ] , showing annotation proportions by object-size category (Tab. I , Supp.) at 640 × 640 YOLO input resolution.
Table 5: Cross-dataset generalization of YOLOv8l models trained on individual source datasets (rows) and evaluated on all test splits (columns). Shading indicates in-domain evaluations. Per-metric bests for each dataset are highlighted in bold .
Size
mAP ↑
FNR ↓
Cat.
@0.25
@0.5
@0.5-0.95
XXS
0.675
0.623
0.320
0.705
XS
0.808
0.799
0.528
0.452
S
0.848
0.842
0.569
0.343
M
0.553
0.543
0.391
0.690
L
–
–
–
–
Table 6: Performance of YOLOv8l on the CST test split by object-size category (defined in Tab. I , Supp.). Per-metric bests are shown in bold , while second-best results are underlined .
Table 7: YOLOv8l performance on the CST test split by sequence-level background category (trained on the CST training split). Per-metric bests are shown in bold ; second bests are underlined .
Attr.
Incl.
mAP ↑
FNR ↓
@0.25
@0.5
@0.5-0.95
TC
✗
0.887
0.868
0.545
0.377
✓
0.666
0.643
0.397
0.653
DBC
✗
0.768
0.748
0.470
0.532
✓
0.707
0.670
0.388
0.604
CDB
✗
0.792
0.770
0.481
0.494
Table 8: YOLOv8l performance across CST test labels, stratified by the presence (✓) or absence (✗) of each annotated attribute.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Size Category
Object Area ( pixels2 )
Description
xx-small (XXS)
A<82
extremely tiny objects
x-small (XS)
82≤A<162
very small objects
small (S)
162≤A<322
small objects
medium (M)
322≤A<962
medium-sized objects
large (L)
A≥962
large objects
Appendix
Table I : Object size categories by bounding-box area A=w×h .
Table II : Cross-dataset generalization of YOLOv8l at [email protected]. Rows and columns correspond to training and evaluation datasets, respectively. Shading indicates in-domain evaluations. The highest score per evaluation dataset is highlighted in bold .
Model
mAP ↑
FDR ↓
FNR ↓
@0.25
@0.5
@0.5-0.95
YOLOv8n
0.657
0.637
0.386
0.084
0.538
YOLOv8s
0.648
0.619
0.363
0.367
0.419
YOLOv8m
0.711
0.691
0.420
0.062
0.532
YOLOv8l
0.747
0.726
0.452
0.030
0.537
YOLO11s
0.742
0.717
0.434
0.068
0.472
Appendix
Table III : Comparison of YOLO variants across model generations and architectural scales, trained on the CST training split and evaluated on the corresponding test split. Best results are highlighted in bold ; second-best results are underlined .
Figure I : Qualitative detection failures of YOLOv8l on the CST [ 31 ] test split (red: GT; green: predictions; orange: regions enlarged below). Left: an aircraft is falsely detected as a drone, while the actual target is missed. Middle: an OSD character occludes the drone, shifting both the annotation and prediction onto the overlay. Right: OSD characters produce false detections alongside the correctly detected drone.
Figure II : Consecutive frames from the H2 FLIR subset [ 2 ] , illustrating the loss of drone detection (green box) following an apparent AGC-related contrast shift, despite the target remaining visible.
YOLO
Size
mAP ↑
FNR ↓
Model
Cat.
@0.25
@0.5
@0.5-0.95
v8n
XXS
0.553
0.505
0.257
0.671
XS
0.728
0.720
0.468
0.463
S
0.801
0.789
0.495
0.407
M
0.461
0.411
0.260
0.785
L
–
–
–
–
Appendix
Table IV : Detection performance by object-size category across model configurations on the CST test split [ 31 ] . Size categories follow the definitions in Tab. I . Per-metric bests are shown in bold , while second-best results are underlined .
YOLO
Attr.
Incl.
mAP ↑
FNR ↓
Model
@0.25
@0.5
@0.5-0.95
v8n
TC
✗
0.864
0.843
0.519
0.314
✓
0.542
0.517
0.307
0.700
DBC
✗
0.674
0.655
0.402
0.536
✓
0.653
0.596
0.323
0.562
CDB
✗
0.712
0.689
0.418
0.498
Appendix
Table V : Attribute-specific detection performance across YOLOv8 (n/s/m) and YOLO11 (s/m/l) configurations on the CST test split [ 31 ] , stratified by the presence (✓) or absence (✗) of annotated instance attributes. For each attribute and metric, the superior result is highlighted in bold .
Figure III : Spatial distribution of YOLOv8l detection outcomes across selected CST test sequences [ 31 ] spanning diverse background categories. GT boxes are accumulated across frames and superimposed on a representative frame, with true positives (TP) shown in green and false negatives (FN) in red. Detection performance is higher against homogeneous sky and water backgrounds, whereas FNs are concentrated in areas with vegetation and complex background structures, indicating considerable variability even within the same background category ( mountain ).
Figure IV : Accumulated GT boxes of all building instances in the CST test split [ 31 ] , superimposed on a representative frame. Left: instances with (blue) and without (red) TC annotation. Right: detection outcomes across all detectors in Tab. III , distinguishing instances detected by all (green), some (yellow), or none (red). TC-annotated instances largely coincide with regions of frequent missed detections along the building facade. However, misses also occur in regions without TC annotation, suggesting a potential role of background structure beyond the annotated attributes and possible limitations in capturing low-contrast conditions through TC annotations.
Table VI : Detection performance by sequence-level background category across model configurations on the CST test split [ 31 ] . For each model and metric, the best and second-best results are highlighted in bold and underlined , respectively.
Institute for the Protection of Terrestrial Infrastructures, German Aerospace Center (DLR), Sankt Augustin, Germany · ACIDA Lab, Technical University of Applied Sciences Würzburg-Schweinfurt, Schweinfurt, Germany