HIT-UAV Dataset: Infrared Thermal Benchmark

Updated 6 July 2026

HIT-UAV dataset is a high-altitude infrared thermal imaging benchmark designed for UAV-based detection in diverse real-world scenarios.
It comprises 2,898 annotated images sampled from over 43,000 frames captured under varying altitudes, angles, and lighting conditions.
The dataset features both standard and oriented bounding box annotations along with comprehensive flight metadata to analyze detection performance.

HIT-UAV is a high-altitude infrared thermal dataset for Unmanned Aerial Vehicle-based object detection. It comprises 2,898 infrared thermal images extracted from 43,470 frames in hundreds of videos captured by UAVs in various scenarios including schools, parking lots, roads, and playgrounds. The dataset provides essential flight data for each image, such as flight altitude, camera perspective, date, and daylight intensity, and it supplies manually annotated object instances with both oriented and standard bounding boxes. It is presented as the first publicly available high-altitude UAV-based infrared thermal dataset for detecting persons and vehicles (Suo et al., 2022).

1. Motivation and problem setting

HIT-UAV was created to address several limitations identified in existing UAV object-detection datasets. The paper states that most UAV datasets are visual-light only, which limits their usefulness for night-time operations and raises privacy concerns because person appearance is visible. It also argues that many datasets do not record enough flight metadata, such as altitude and viewing angle, making it difficult to study how UAV operating conditions affect detection. A further limitation is narrow data distribution: existing datasets are often synthetic rather than real scenes, low-altitude only, single-scene, or single-category (Suo et al., 2022).

Within that framing, HIT-UAV is designed as a real, high-altitude, infrared thermal resource with metadata sufficient for studying how flight parameters affect detection performance. The paper explicitly connects this design to practical UAV missions such as night-time search-and-rescue, traffic monitoring, and other multi-scene detection tasks. The infrared thermal modality is therefore not merely a sensor choice; it is central to the dataset’s positioning around night-time applicability and privacy-preserving detection.

2. Data acquisition and corpus composition

The sensing modality is infrared thermal imagery captured using the DJI Zenmuse XT2 thermal camera, described as a FLIR longwave infrared thermal sensor, with thermal resolution $640 \times 512$ . The UAV platform is a DJI Matrice M210 V2. Although the device also includes a visual camera, the dataset itself is described as an infrared thermal dataset (Suo et al., 2022).

The videos were collected in public, desensitized environments. The reported scenes include schools, parking lots, roads, playgrounds, and more similar real-world settings. Flights were performed during both day and night, and the weather condition recorded in filenames indicates no-rain conditions only. The raw corpus contains 43,470 frames extracted from hundreds of videos. Because the raw videos had a refresh rate of about 7 FPS, the authors sampled one image every 15 frames to reduce near-duplicate adjacent frames, improve feature diversity, and avoid redundant training samples. After sampling and cleaning, the final dataset contains 2,898 infrared thermal images.

Attribute	Value
Modality	Infrared thermal
Sensor	DJI Zenmuse XT2
Thermal resolution	$640 \times 512$
UAV platform	DJI Matrice M210 V2
Source material	43,470 frames from hundreds of videos
Final image count	2,898 infrared thermal images
Altitude range	60 m to 130 m
Camera perspective range	$30^\circ$ to $90^\circ$
Temporal conditions	Day and night
Weather inclusion	No-rain only
Scene types	Schools, parking lots, roads, playgrounds

The dataset’s scene and flight variability is one of its defining properties. It spans high-altitude capture from 60 m to 130 m, camera perspectives from $30^\circ$ to $90^\circ$ , day and night imaging, and multiple common UAV application contexts. This design is intended to make the dataset more representative of real UAV deployment conditions than low-altitude or synthetic datasets.

3. Annotation scheme and geometric representation

HIT-UAV includes five labeled categories: Person, Car, Bicycle, OtherVehicle, and DontCare. The first four are the main detection categories, while DontCare is reserved for objects that annotators could not reliably categorize (Suo et al., 2022).

For every object, the dataset provides two annotation formats. The standard bounding box is represented as

$(x_c, y_c, w, h),$

and the oriented bounding box is represented as

$(x_c, y_c, w, h, \theta),$

where $(x_c, y_c)$ is the box center, $w$ and $640 \times 512$ 0 are width and height, and $640 \times 512$ 1 is the angle relative to the horizontal direction.

The paper motivates oriented boxes by the geometry of high-altitude aerial imagery. In such images, objects can be densely packed and rotated relative to the image axes, and standard axis-aligned boxes can overlap heavily, especially in crowded aerial scenes. The authors therefore argue that oriented boxes reduce overlap, better fit elongated and rotated objects, better capture aerial geometry, and improve annotation accuracy for dense scenes. At the same time, because many detection frameworks do not natively support rotated boxes, the dataset also provides converted standard boxes for practical use.

The conversion from oriented to standard box is defined through the axis-aligned extremes of the rotated box. After obtaining $640 \times 512$ 2, the converted standard box is

$640 \times 512$ 3

$640 \times 512$ 4

The annotation workflow used a modified LabelImg tool. Three people participated in annotation, annotations were cross-checked by others, and difficult and truncated objects were also labeled. The repository additionally provides a conversion tool from oriented boxes to standard boxes (Suo et al., 2022).

4. Metadata design and benchmark protocol

A distinctive feature of HIT-UAV is that each image filename encodes flight metadata in the format

$640 \times 512$ 5

In this encoding, $640 \times 512$ 6 denotes shooting time, with $640 \times 512$ 7 for day and $640 \times 512$ 8 for night; $640 \times 512$ 9 denotes flight altitude, ranging from 60 to 130 m; $30^\circ$ 0 denotes camera perspective, ranging from $30^\circ$ 1 to $30^\circ$ 2; $30^\circ$ 3 denotes weather condition, with only no-rain images included; and $30^\circ$ 4 denotes image serial number (Suo et al., 2022).

The paper also states that the following metadata are recorded for each video or image: flight altitude, camera perspective, flight date, and daylight intensity. These variables are intended to support analysis of how altitude affects object size and detection accuracy, how camera angle changes visible object area and overlap, how day versus night changes thermal contrast, and how lighting intensity and time of capture affect performance. A plausible implication is that HIT-UAV can be used not only for static benchmarking but also for conditional analyses of detector robustness under controlled flight-context variation.

The dataset split follows a hold-out method with 70% training, 10% validation, and 20% test. The exact counts are 2,029 training images, 290 validation images, and 579 test images. The main evaluation metric reported is AP per class and mAP@0.50 overall. The paper also reports true positives, false positives, and false negatives alongside AP and mAP for some models.

5. Baseline detectors and reported performance

The paper trains and evaluates four well-established object detectors on HIT-UAV: YOLOv4, YOLOv4-tiny, Faster R-CNN, and SSD-512 (Suo et al., 2022). YOLOv4 and YOLOv4-tiny were implemented in Darknet, while Faster R-CNN and SSD were implemented in MMDetection. The reported hardware was an RTX 2080 Ti.

For YOLO training, the setup used up to 10,000 steps, batch size 64, subdivision 16, learning rate 0.0013, and learning-rate drops at steps 8000 and 9000. For Faster R-CNN and SSD, the setup used a maximum of 32 epochs, batch size 16, learning rate 0.02, warm-up ratio 0.001, and warm-up iterations 500.

On the HIT-UAV test set, the reported results are as follows.

YOLOv4: Person 89.88, Car 92.64, Bicycle 86.48, OtherVehicle 69.99, with [email protected] of 84.75.
YOLOv4-tiny: Person 16.86, Car 83.61, Bicycle 51.90, OtherVehicle 49.17, with [email protected] of 50.38.
Faster R-CNN: Person 75.5, Car 95.6, Bicycle 86.4, OtherVehicle 46.8, with [email protected] of 76.8.
SSD-512: Person 85.6, Car 96.3, Bicycle 86.0, OtherVehicle 74.4, with [email protected] of 85.6.

The paper compares these numbers with several visual-light benchmarks, reporting YOLOv4 on COCO at 65.7 mAP, YOLOv4-tiny on COCO at 40.2 mAP, and RRNet on VisDrone-2019 at 55.82 mAP. The authors emphasize that detection performance on HIT-UAV is exceptionally high and interpret this as evidence that infrared thermal images remove irrelevant appearance details and simplify detection. That interpretation is the authors’ own; a cautious reading is that the dataset supports strong performance for standard detectors under the specific modality, scene types, and annotation protocol used here.

6. Applications, limitations, and relation to adjacent UAV datasets

The paper highlights HIT-UAV as useful for UAV-based object detection, night-time search and rescue, studying the effect of altitude on detection accuracy, studying the effect of camera angle on detection accuracy, exploring infrared thermal sensing for UAV mission planning, and privacy-preserving detection of persons and vehicles (Suo et al., 2022). The dataset is freely available on GitHub and also archived on Zenodo, with repository tools for converting oriented boxes to standard boxes, visualization, and converting labels to YOLO format.

Several limitations are either explicit or directly implied in the dataset description. Only no-rain conditions are included. The thermal data come from specific scenes and environments, so broader generalization still requires testing elsewhere. The OtherVehicle category has fewer instances, causing class imbalance. Oriented box detection is not universally supported by standard detectors, which is why standard-box conversion is provided. The dataset also emphasizes persons and vehicles rather than serving as a generic all-object benchmark. These limitations are important for interpretation: strong benchmark results do not, by themselves, establish robustness across all environmental conditions or object taxonomies.

In relation to adjacent UAV benchmarks, HIT-UAV occupies a different task and modality regime from UAVid. UAVid is an urban semantic segmentation dataset with 30 video sequences, 300 densely labeled 4K RGB images in slanted views, and 8 semantic classes, whereas HIT-UAV is an infrared thermal object-detection dataset with box annotations, flight metadata, and a focus on persons and vehicles (Lyu et al., 2018). A separate and later line of work introduces a UAV-based VNIR hyperspectral benchmark dataset for landmine and UXO detection and situates it within a broader HIT-UAV / UAV demining resource context; that resource is distinct in sensing modality, task definition, target type, and site design from the infrared person-and-vehicle dataset ordinarily denoted by HIT-UAV (Lekhak et al., 3 Oct 2025).

Taken together, HIT-UAV is best understood as a specialized but influential benchmark at the intersection of UAV perception, infrared thermal sensing, and metadata-aware aerial detection. Its defining contribution lies in the combination of high-altitude capture, day/night infrared imagery, dual box geometry, and per-image flight context, which makes it suitable not only for detector comparison but also for controlled analysis of how UAV operating parameters affect detection behavior.