YOLOatr: MWIR ATR Detector

• YOLOatr is a modified YOLOv5s single-stage, anchor-based detector designed for automatic target recognition in MWIR imagery.
• It incorporates an additional small-scale detection head and replaces PANet with BiFPN, boosting detection of small, low-contrast targets with up to 99.6% mAP correlated and 37.7% mAP decorrelated.
• The model achieves real-time throughput (110 fps on a P100 GPU) on the DSIAC MWIR dataset, demonstrating its practical significance in defense and surveillance.

Searching arXiv for the primary YOLOatr paper and closely related usages of the term in recent YOLO literature. YOLOatr is a modified anchor-based single-stage detector for automatic target detection and recognition in thermal infrared imagery, specifically mid-wave infrared imagery in the defense and surveillance domain. It is architecturally based on a modified YOLOv5s and is designed for real-time detection and classification of small, low-contrast tactical and civilian vehicles at long ranges in cluttered environments, under both day and night conditions. Its defining modifications are an additional small-scale detection head, replacement of the default PANet neck with BiFPN, and a custom augmentation profile tuned for thermal infrared data. On the DSIAC MWIR dataset, it reports state-of-the-art ATR performance of up to 99.6% and about 110 fps at $640 \times 640$ on an NVIDIA P100 GPU (Safdar et al., 15 Jul 2025).

1. Domain and task formulation

YOLOatr is situated in the ATR problem setting for thermal infrared imaging. In this context, ATD denotes automatic target detection, i.e., target localization by bounding boxes, whereas ATR denotes detection plus class recognition. The relevant sensing modality is thermal IR, and specifically MWIR in the $3\text{–}5\,\mu\text{m}$ band, which is widely used for ground-based tactical platforms because it measures emitted rather than reflected radiation and therefore operates in both day and night and under poor visible-light conditions (Safdar et al., 15 Jul 2025).

The paper frames ATR in MWIR as substantially more difficult than standard RGB detection. The reported causes include limited datasets and labeling, lower native sensor resolution than modern RGB sensors, strong dependence of dynamic range and thermal contrast on weather and time of day, extreme scale variation from 1000 m to 5000 m, deliberate occlusion and camouflage, clutter and low signal-to-clutter noise ratio, and increased intra-class variability combined with higher inter-class similarity. A central consequence is that contemporary state-of-the-art deep learning architectures underperform in this domain, especially under decorrelated testing, where training and testing occur at different range regimes (Safdar et al., 15 Jul 2025).

Within that setting, YOLOatr is explicitly characterized as a single-stage, single-frame, anchor-based object detector derived from YOLOv5s. The operational objective is not only high correlated-range accuracy but improved generalization to long-range, small-target conditions, while preserving real-time throughput on constrained hardware (Safdar et al., 15 Jul 2025).

2. Architectural design

The base architecture is YOLOv5s. As summarized in the source description, YOLOv5s consists of a CSPDarknet-like backbone with Focus, CSP bottlenecks, and SPPF, a PANet-style neck for multi-scale fusion, and three anchor-based detection heads at progressively lower resolutions. YOLOatr retains the YOLOv5s backbone and modifies the neck and head components (Safdar et al., 15 Jul 2025).

The first major change is the introduction of an additional small-scale detection head $P2$. In the reported formulation, standard YOLOv5s uses heads at strides $8$, $16$, and $32$ for $P3/P4/P5$, whereas $P2$ corresponds to stride $4$. The motivation is explicit: DSIAC MWIR targets at long ranges may be only a few pixels wide or high, and a finer feature scale preserves more spatial detail. The head design remains YOLOv5-style, predicting $(x,y,w,h)$, objectness, and class probabilities, but now at an additional high-resolution level (Safdar et al., 15 Jul 2025).

The second major change is the replacement of the default PANet neck with BiFPN. The paper motivates this substitution by the need for improved multi-scale feature fusion under cluttered small-object conditions. BiFPN is used as a weighted bidirectional feature pyramid, so that features from different resolutions are fused in both top-down and bottom-up directions with learned layer weights. This is presented as a way to improve propagation of semantic and spatial information across scales and thereby improve small-target visibility (Safdar et al., 15 Jul 2025).

No explicit modification to the backbone is reported. The complexity increase is therefore attributed primarily to the added $3\text{–}5\,\mu\text{m}$0 head and the PANet-to-BiFPN replacement rather than to a larger feature extractor (Safdar et al., 15 Jul 2025).

The loss functions are described as the standard YOLOv5 losses. The paper does not detail the loss formulas explicitly, but the summary reports CIoU for box regression and BCE for objectness and classification, with focal-style weighting through a nonzero fl_gamma. In the reported notation,

$3\text{–}5\,\mu\text{m}$1

and

$3\text{–}5\,\mu\text{m}$2

The augmentation profile also sets fl_gamma = 0.3, described as a weak focal effect (Safdar et al., 15 Jul 2025).

3. Dataset, partitioning, and training protocol

YOLOatr is evaluated on the DSIAC MWIR ATR Algorithm Development Image Database (NVESD). The dataset contains MWIR thermal imagery and corresponding visible imagery, though the reported experiments use MWIR. Acquisition occurs in a desert environment over ranges from 1000 m to 5000 m in 500 m increments, with 72 aspect angles per vehicle and both day and night sequences. The dataset consists of 186 one-minute videos at 30 fps, approximately 1800 frames per video, covering 10 vehicles and 2 human classes. The total sample count is reported as approximately 367,200 (Safdar et al., 15 Jul 2025).

For the YOLOatr experiments, four vehicle classes are selected: T72 tank, BTR70, Pickup, and SUV. For each vehicle type, 3600 images are used, with a 70%/20%/10% train/validation/test split. The source text also distinguishes two partitioning protocols. In DS1, training ranges are 1.0, 1.5, 2.0, and 2.5 km with correlated testing on the same ranges. In DS2, training ranges are 3.0, 3.5, 4.0, and 4.5 km with decorrelated testing at 5.0 km. The key distinction is that correlated testing preserves the same range regime between train and test, whereas decorrelated testing evaluates generalization across range-induced scale and clutter changes (Safdar et al., 15 Jul 2025).

Training uses the Ultralytics YOLOv5 implementation in PyTorch on Google Colab Pro with an NVIDIA Tesla P100-PCIE 16 GB GPU and 32 GB RAM. Images are resized to $3\text{–}5\,\mu\text{m}$3. The optimizer is SGD, with batch size 32 and 100 epochs for training from scratch. The paper also evaluates transfer learning from ImageNet-pretrained RGB weights using a two-stage schedule—30 epochs with frozen backbone followed by 10 epochs of fine-tuning—but reports that training from scratch yields slightly better performance for MWIR, attributing this to the mismatch between RGB textures and MWIR thermal patterns (Safdar et al., 15 Jul 2025).

A central element of the training recipe is the Custom Augmentation Profile (CAP). The reported settings are hsv_h = 0.015, hsv_s = 0.7, hsv_v = 0.4, translate = 0.1, scale = 0.3, shear = 0.0, perspective = 0.0005, flipud = 0.1, fliplr = 0.5, [mosaic](https://www.emergentmind.com/topics/mosaic-b2e93c3d-34d6-4f4a-9633-8f0432aa236b) = 0.1, mixup = 0.4, copy_paste = 0.5, and fl_gamma = 0.3. The rationale given is domain-specific: brightness and contrast jitter emulate thermal variation; shear is disabled because low-resolution thermal targets are already structurally weak; mosaic is kept low because further shrinking already tiny targets is harmful; and mixup and copy-paste are kept relatively high to enrich clutter and context (Safdar et al., 15 Jul 2025).

4. Evaluation methodology and empirical results

The primary reported metrics are precision, recall, and mAP@0.5. The source description defines

$3\text{–}5\,\mu\text{m}$4

and reports mAP as the mean of classwise average precision values at IoU $3\text{–}5\,\mu\text{m}$5 (Safdar et al., 15 Jul 2025).

The principal comparison is between vanilla YOLOv5s and YOLOatr under correlated and decorrelated testing:

Scenario	YOLOv5s [email protected]	YOLOatr [email protected]
Correlated testing (T1)	$3\text{–}5\,\mu\text{m}$6	$3\text{–}5\,\mu\text{m}$7
Decorrelated testing (T2)	$3\text{–}5\,\mu\text{m}$8	$3\text{–}5\,\mu\text{m}$9

The correlated regime is near saturation for both models, with YOLOatr reported at 99.6% [email protected] and YOLOv5s at approximately 99.4%. The more consequential result is the decorrelated regime, where YOLOatr reaches 37.7% [email protected] versus 27.1% for YOLOv5s; the paper reports this as a +11.4 percentage point improvement in the decorrelated setting (Safdar et al., 15 Jul 2025).

Target-wise decorrelated performance for YOLOatr is also reported. At $P2$0, T72 tank achieves 62.2%, BTR70 21.4%, SUV 39.3%, and Pickup 16.3% [email protected]. The discussion associates the stronger T72 result with larger size and more distinctive thermal signature, whereas commercial vehicles remain more difficult because they are smaller and more easily confused with clutter or each other (Safdar et al., 15 Jul 2025).

The paper also positions YOLOatr against prior DSIAC results. The summary states that correlated performance at 4.0–5.0 km is approximately 99.4–99.6%, competitive with or higher than earlier ATR reports, and that in decorrelated ATR at 5.0 km YOLOatr reaches 73.3% under the comparison setup cited from prior work. It is also compared with older GPU measurements, including YOLOv2 at 59 fps, Faster R-CNN VGG16 at 7 fps, and Faster R-CNN ResNet at 5 fps, whereas YOLOatr is reported at 110 fps on a P100 at $P2$1 (Safdar et al., 15 Jul 2025).

Ablation evidence is summarized qualitatively rather than as a complete factorized table. The reported experiments vary architecture, learning strategy, and augmentation profile. CAP improves training and validation accuracy over lower or default augmentation settings; from-scratch training slightly exceeds transfer learning; and the combined architectural modifications and CAP yield the reported decorrelated gain over vanilla YOLOv5s (Safdar et al., 15 Jul 2025).

5. Operational significance

YOLOatr is explicitly oriented toward defense and surveillance systems, including ground-based tactical vehicles, perimeter security with IR cameras, and forward-looking sensors on mobile platforms. Several properties are operationally salient. First, the architecture remains lean because it is based on YOLOv5s rather than a heavy backbone. Second, it is single-frame rather than temporal, so it does not rely on background subtraction, motion cues, or multi-frame aggregation, which may fail under sensor movement, camera shake, or multiple moving sources. Third, it is designed for day/night, multi-range, and cluttered environments, with qualitative examples reportedly showing detections in scenes where targets are difficult for human observers to notice (Safdar et al., 15 Jul 2025).

The deployment argument is therefore not merely accuracy-based. The paper presents YOLOatr as a detector with a favorable throughput–accuracy profile for constrained hardware. The reported 110 fps on a P100 suggests feasibility for embedded deployment after further optimization, and the single-stage formulation is described as offering high throughput per watt compared with heavier two-stage alternatives. A plausible implication is that the method is intended as a practical baseline for onboard MWIR perception where latency and power are first-order constraints (Safdar et al., 15 Jul 2025).

6. Terminological and research context

Although YOLOatr is the explicit title of the MWIR ATR detector in "YOLOatr : Deep Learning Based Automatic Target Detection and Localization in Thermal Infrared Imagery" (Safdar et al., 15 Jul 2025), closely related literature uses similar shorthand to denote several different YOLO-derived design trajectories.

One neighboring usage treats the term as referring to YOLO-based all-tasks-at-once perception for autonomous driving. In that sense, A-YOLOM or YOLOv8-multi-task is a single YOLO-style network that jointly performs object detection, drivable area segmentation, and lane line segmentation in real time, with a unified segmentation head and adaptive concatenation between backbone and neck features (Wang et al., 2023).

A second nearby usage interprets it as YOLO with adaptive routing. The Mixture-of-Experts detector built on YOLOv9-T uses multiple experts and per-scale routers that produce expert weights from expert features and a reweighted Hadamard fusion, improving mAP and average recall relative to a single YOLOv9-T under multi-domain training on COCO and VisDrone (Meiraz et al., 17 Nov 2025).

A third interpretation associates it with YOLO + transformer hybrids. In that line, YotoR combines a Swin Transformer backbone with a YoloR P6 neck and head, retaining YOLO-style multi-scale detection while improving over corresponding Swin detector baselines in both speed and COCO accuracy (Villa et al., 2024).

A fourth nearby interpretation reads the string as YOLO + ability to remember, i.e., continual object detection. In that context, YOLO LwF introduces self-distillation for YOLOv8n, with regression distillation on DFL outputs, IoU-weighted classification distillation, masking of old classes, and replay memory, improving continual-detection mAP over earlier baselines on VOC and COCO (Monte et al., 6 Mar 2025).

Taken together, these papers indicate that the label has acquired a broader informal resonance around YOLO variants that emphasize adaptation—across tasks, experts, modalities, or temporal training sequences—but the actual titled model YOLOatr remains the MWIR-specific ATR detector based on modified YOLOv5s (Safdar et al., 15 Jul 2025).

7. Limitations and future directions

The reported limitations are concrete. The authors do not use YOLOv5’s genetic hyperparameter optimizer, so further tuning of loss weights and augmentation probabilities remains open. They also do not explicitly analyze the false alarm rate trade-off, despite the relevance of FAR in ATR practice. Architectural exploration is limited to adding $P2$2 and replacing PANet with BiFPN; more extensive redesigns such as anchor-free formulations, attention mechanisms, or specialized small-target modules are not pursued (Safdar et al., 15 Jul 2025).

The paper also emphasizes that decorrelated performance, while substantially improved over YOLOv5s, is still far from saturated. The reported 37.7% [email protected] under decorrelated testing indicates that long-range ATR remains difficult, and Pickup and SUV remain especially challenging at 5 km because of extreme small size and low contrast. This suggests that the principal unresolved problem is not correlated recognition but robust generalization across range and clutter regimes (Safdar et al., 15 Jul 2025).

The stated future direction is to explore YOLOv7, YOLOv8, and YOLO-NAS variants. Given the current design, a plausible implication is that later work may preserve the same guiding principles—enhanced small-object sensitivity, stronger cross-scale fusion, and thermal-specific augmentation—while transplanting them into newer detector families. Within the present literature, however, YOLOatr is best understood as a domain-specialized YOLOv5s derivative that demonstrates how relatively limited but carefully targeted changes to head design, neck fusion, and augmentation can materially improve real-time MWIR ATR (Safdar et al., 15 Jul 2025).