Joint False Alarm Dataset (JFAD)

This repository presents the Joint False Alarm Dataset (JFAD). The dataset is designed to support the training, evaluation, and validation of deep learning models for thermal anomaly detection and false alarm reduction in terrestrial remote sensing scenarios. JFAD provides a comprehensive representation of thermal interference sources that commonly trigger false detections, making it a strong benchmark for developing robust cascade-based detection architectures.

JFAD was constructed by consolidating two complementary repositories: the Addressing False Alarm Situations (AFAS) dataset and the Thermal Anomaly (TA) dataset. Together, these collections characterize a broad spectrum of challenging thermal conditions across both the near-infrared (NIR) and long-wave infrared (LWIR) domains. The dataset includes diverse environments such as urban infrastructure monitoring, rural surveillance, and wilderness fire-like anomaly scenarios.

A key contribution of this dataset is the re-annotation of the TA subset, extending traditional bounding-box labels into pixel-level instance masks. This enables models to better distinguish filament-like structures such as powerlines from irregular thermal signatures caused by incipient fires, solar glint, or exhaust emissions, an essential requirement for minimizing false alarms in real-world deployments.

Dataset Composition
The JFAD dataset comprises a total of 6098 thermal images, aggregated from multiple measurement campaigns and established open benchmarks, including:
1) FLIR Advanced Driver Assistance Systems (ADAS)
2) TarDAL M3FD
3) Powerline Image Dataset (PID)
4) In-house thermal acquisitions across urban, rural, and wilderness settings

The dataset reflects substantial cross-sensor variability, incorporating different platforms, resolutions, and monitoring conditions.

For experimental evaluation, JFAD is divided into:
4308 training samples and 1790 validation samples

Technical Specifications
File Formats: JPEG, TXT, NPY
Spectral Bands: NIR and LWIR
Resolution Range: 160×120 up to 640×512

Cameras used: FLIR Tau 2, FLIR A615, FLIR A35, Seek Mosaic, and FLIR Lepton

This repository contains the dataset used in our work. If you wish to replicate our experiments or access the full implementation, please refer to the link provided in the Related Links section, which directs you to the official GitHub repository associated with the paper.