MONET Dataset: Drone Thermal Imaging

• MONET is a drone-borne thermal dataset comprising 53K images and 162K manually annotated bounding boxes, designed for robust object localization and tracking.
• It integrates synchronized drone metadata—including GPS, attitude, and speed—which enables multimodal analysis and transfer learning across distinct rural environments.
• The dataset tackles challenges such as low thermal resolution, moving viewpoints, and cross-site transfer, establishing a benchmark for aerial perception research.

Searching arXiv for the dataset paper and closely related work to support a precise article. MONET, short for Multimodal drOne thErmal daTaset, is a drone-borne long-wave infrared (LWIR) thermal dataset recorded in rural environments for studying object localisation and behaviour understanding of humans and vehicles under moving-viewpoint aerial thermal imaging. It was introduced as a dataset of approximately 53K images with 162K manually annotated bounding boxes, where each image is timestamp-aligned with drone metadata including attitudes, speed, altitude, and GPS coordinates. MONET is designed around two rural sites with distinct structures and cluttered backgrounds, and emphasizes the combined difficulty of thermal sensing, moving cameras, large scale variation, and cross-site transfer (Riz et al., 2023).

1. Recording context and intended problem setting

MONET was recorded in a rural area near the city of Nicosia, Cyprus, in mid-December, during the afternoon, evening, and night. The dataset contains two main recording sites: runway and dirt-road. The runway was recorded near a runway on property of The Cyprus Institute, while the dirt-road scenario was recorded on agricultural land. The two sites were chosen because they have different scene structures and different cluttered backgrounds, making them suitable for transfer-learning and domain-shift studies (Riz et al., 2023).

The dataset targets aerial thermal perception under conditions that are explicitly difficult for standard detectors. The paper identifies several sources of difficulty: thermal cameras usually have lower resolution than RGB cameras; target appearance varies with temperature, emissivity, ambient temperature, and humidity; backgrounds can emit similar heat to targets; images can be affected by noise and ghost effects; and objects can be small, occluded, or scale-varying because of altitude and viewpoint changes. MONET further compounds these issues through a moving drone and different and moving viewpoints, rather than a static top-down platform. This makes it simultaneously a detection dataset, a tracking dataset, and a testbed for multimodal aerial reasoning.

The target categories are person and vehicle, with an additional ignore category used for a specific runway region. The dataset was explicitly constructed to support object detection / localisation, multi-object tracking, trajectory analysis, behaviour/activity understanding, transfer learning / domain adaptation across sites, and multimodal learning that exploits aligned flight metadata.

2. Sensor platform, image modality, and metadata alignment

MONET was collected using a fully customised multirotor drone developed for automated surveillance and archaeological looting detection. The platform has eight motors and eight electronic speed controllers, mounted on four arms, with payload capacity of up to 1.5 kg. Onboard sensing for flight control includes a high-frequency IMU, barometric altimeter, external GPS, and compass module. The platform supports manual and autonomous flight modes, including hovering, automated waypoint navigation, return-to-home, auto take-off, and auto-landing (Riz et al., 2023).

The acquisition system uses the Workswell WIRIS Security camera, which contains both an RGB sensor and a thermal sensor. The RGB sensor supports Full HD $1920\times1080$ at 30 Hz with up to 30× optical zoom. The thermal sensor operates in the LWIR range of $7.5\text{–}13.5 \,\mu m$ at $800\times600$ resolution, with a thermal sensitivity range reported as $-20^\circ C$ to $150^\circ C$. The camera is mounted on a gimbal, and MONET includes gimbal attitude metadata in addition to drone attitude metadata.

The software pipeline used the camera’s proprietary SDK over Ethernet, integrated into the drone ground station using ROS. RGB and thermal streams were transmitted as RTSP video streams, received by a Data Processing Unit, and published by ROS nodes. Image timestamps and metadata timestamps were both logged during acquisition. Metadata was captured at about 40 Hz on average, and image-to-metadata alignment was performed by nearest-neighbour assignment between image and metadata timestamps. Per image, the aligned metadata fields include:

• date in ISO 8601 format
• drone pitch, drone roll, drone yaw
• gimbal pitch, gimbal roll, gimbal yaw
• latitude, longitude
• altitude
• x-axis speed, y-axis speed, z-axis speed

Although the sensor package is intrinsically RGB-plus-thermal, the paper states that RGB images are not included in the current release because of unresolved privacy concerns. In practice, the released resource is therefore centered on thermal imagery plus aligned drone metadata (Riz et al., 2023).

3. Dataset composition, classes, and annotation protocol

MONET contains approximately 53K images / frames and 162K manually annotated bounding boxes. The paper reports the following site-wise totals:

Site	Frames	Bounding boxes
dirt-road	23.3K	83.4K
runway	29.4K	79.3K

The annotation categories are person, vehicle, and ignore. The vehicle category is defined as car-like objects. The ignore category marks a region near the hangar area in the runway scenario. The rationale is to prevent models from learning a contextual bias that people and vehicles often occur next to the hangar; pixels inside ignore regions were therefore zeroed out during training, while the original images remain available for alternative use (Riz et al., 2023).

Target density statistics are reported for frames that contain targets. For person, the average number of instances is 2.96, with max 6 and min 1. For vehicle, the average is 1.33, with max 4 and min 1. For ignore, the count is always 1 when present. Bounding boxes are linked with target identities, so the dataset supports object trajectories and multi-object tracking.

Annotations were created using CVAT by six people, with three people subsequently double-checking the labels for accuracy and consistency. The annotation rules are unusually explicit. Annotators were instructed to draw boxes as tight as possible, provided the object was clearly distinguishable from the background. Interpolation across frames was allowed, but every frame still had to be checked to ensure the box was centered correctly. Brightness, contrast, and saturation could be adjusted in CVAT to improve visibility. If a target was partially occluded or indistinguishable from the background, annotators were told to use the best guess and flag the box as occluded. Annotation of a target was to start when more than approximately 30% of its pixels were visible in the scene, and terminate when more than approximately 70% of its pixels were outside the scene. If a target exited and later re-entered, it received a new ID.

A notable property of MONET is its thermal-specific notion of visibility. In this dataset, “occlusion” includes not only geometric occlusion by another object but also cases where a target becomes thermally difficult to distinguish from the background or from nearby warm objects. The paper retains such boxes because they are informative for tracking and hard-case analysis (Riz et al., 2023).

4. Benchmark tasks and evaluation protocol

The paper benchmarks MONET primarily on object detection / localisation, using the COCO evaluation procedure and reporting AP, AP$_{50}$, and AP$_{75}$, together with class-wise AP. It evaluates nine object detection algorithms:

1. Faster R-CNN
2. SSD
3. CornerNet
4. FCOS
5. DETR
6. Deformable DETR
7. VarifocalNet
8. ObjectBox
9. YOLOv8

Most detectors were implemented through MMDetection, while ObjectBox and YOLOv8 used the authors’ original implementations. The paper also reports ObjectBox$^\dagger$ and YOLOv8$^\dagger$, which use the methods’ original augmentation policies (Riz et al., 2023).

To keep training conditions comparable, the paper states that all models were trained, as far as possible, with the same data augmentations, backbones, and optimization parameters. The common augmentation sequence was:

1. RandomCrop with crop factors in $[0.8,1.0]$
2. Resize between $7.5\text{–}13.5 \,\mu m$0 and $7.5\text{–}13.5 \,\mu m$1, preserving aspect ratio
3. RandomHorizontalFlip with probability 0.7
4. Padding to $7.5\text{–}13.5 \,\mu m$2
5. Normalization with mean 126.225 and std 73.338

The supplementary material states that non-maximum suppression used IoU threshold 0.5, and that all detector outputs with confidence above $7.5\text{–}13.5 \,\mu m$3 were evaluated.

MONET is also used as a benchmark for transfer learning and domain adaptation. The paper studies training on one MONET site and testing on the other, as well as transfer between MONET and HIT-UAV. This protocol is central to the paper’s argument that thermal modality alone does not guarantee generalization across rural sites or across aerial thermal datasets.

5. Difficulty profile and empirical findings

The empirical results show that MONET is challenging even for modern detectors. A central finding is that runway is easier than dirt-road. Same-site AP values are consistently higher on runway, while dirt-road is harder because daytime ground heat makes people blend into the background. The paper repeatedly notes that person detection is generally harder than vehicle detection, especially in dirt-road scenes (Riz et al., 2023).

The focused Faster R-CNN transfer study illustrates this. When trained on dirt-road + runway, the paper reports:

• dirt-road V: AP 24.9
• dirt-road T: AP 36.3
• runway V: AP 44.5
• runway T: AP 42.1

When training with COCO init + pretrain on HIT-UAV + fine-tune on MONET, the reported AP becomes 39.1 on dirt-road T and 46.2 on runway T, indicating that thermal pretraining on HIT-UAV can help before MONET fine-tuning. However, direct cross-dataset transfer is poor: training only on HIT-UAV yields AP 8.9 on dirt-road T and 12.7 on runway T, while training on MONET and testing on HIT-UAV yields AP 6.7 on HIT-UAV V and 7.1 on HIT-UAV T. Cross-site transfer within MONET is also weak: training on dirt-road and testing on runway gives 18.0 AP on runway T, and training on runway and testing on dirt-road gives 21.3 AP on dirt-road T.

The multi-detector comparison reinforces these conclusions. In the same-scenario dirt-road $7.5\text{–}13.5 \,\mu m$4 dirt-road setting, the best reported result is YOLOv8$7.5\text{–}13.5 \,\mu m$5 with AP 33.3, followed by ObjectBox$7.5\text{–}13.5 \,\mu m$6 with AP 31.4. In same-scenario runway $7.5\text{–}13.5 \,\mu m$7 runway, the best reported result is YOLOv8 with AP 53.5. In cross-site settings, performance drops substantially. The paper identifies SSD and Deformable DETR as particularly strong in transfer settings, while YOLOv8 often has the strongest AP$7.5\text{–}13.5 \,\mu m$8, which the authors interpret as suggesting better box-size localization.

Qualitative analysis adds a thermal-specific explanation for these failures. In dirt-road scenes, a person can become nearly indistinguishable from the ground because of similar heat signatures. The paper also notes that some false positives during transfer from HIT-UAV to MONET runway may arise because HIT-UAV contains many parked vehicles with relatively low temperature, causing models to associate dark thermal patterns with vehicles. This suggests that cross-dataset transfer failure is not merely a capacity problem; it is also a problem of thermal semantics and scene structure.

6. Distinguishing characteristics, uses, and naming ambiguity

MONET differs from previous thermal drone datasets because it combines rural scenes captured with thermal cameras containing both person and vehicle targets, trajectory information, and timestamp-aligned metadata. Its defining contribution is not only dataset scale but the combination of moving multirotor viewpoints, two structurally different rural sites, identity-linked bounding boxes, and multimodal telemetry. This makes it useful for research on metadata-aware detection, motion-aware tracking, scale estimation from altitude, and multimodal fusion between image content and UAV state (Riz et al., 2023).

A common misconception arises from the overloading of the name “MONET” in arXiv literature. The same label has been used for unrelated systems and models, including an Android malware variants detection system (Sun et al., 2016), an unsupervised scene decomposition model (Burgess et al., 2019), and an open text-to-image corpus (Aubin et al., 20 May 2026). In the present sense, however, MONET Dataset refers specifically to the multimodal drone thermal dataset recorded in rural scenarios introduced in 2023 (Riz et al., 2023).

The dataset is publicly associated with the project page:

https://github.com/fabiopoiesi/monet_dataset

Its broader significance lies in making a particular failure mode of aerial perception measurable: even with the same thermal sensor and broadly similar rural geography, detectors trained on one site generalize poorly to another. This suggests that MONET functions not only as a benchmark for thermal object detection, but also as a controlled testbed for cross-site robustness, multimodal aerial perception, and thermal-domain transfer.