BlueFlame-RGBT: A Paired RGB–Thermal Dataset for Gaseous Flame Recognition

BlueFlame-RGBT is a paired RGB–thermal dataset designed for close-range gaseous flame recognition under conditions in which flames may exhibit weak or unstable visible contrast. The dataset contains 6,660 nominal multimodal acquisitions, including 4,558 flame-present and 2,102 flame-absent observations. Each nominal acquisition links a visible-spectrum RGB image, an 8-bit thermal visualization (IR8), and the corresponding raw 16-bit thermal frame (IR16) acquired using a TOPDON TC001 thermal camera. The RGB and thermal sensors were mounted on a common rigid support with an approximate separation of 15 cm between their optical axes, and acquisitions were performed at scene distances of approximately 0.5–4 m. RGB and thermal observations correspond to the same nominal acquisition event but are not pixel-wise registered. The release contains 6,659 complete RGB/IR8/IR16 triplets. Eleven IR16 frames are known all-zero frames and are retained for transparency with modality-specific usability information. The flame-absent class includes thermally active objects and regions intended to provide challenging negative examples. To reduce scene-family leakage, sequential observations were first organized into 1,521 refined visual groups. Visually related groups were subsequently linked into leakage components. Each leakage component was assigned entirely to one of the training, validation, or test partitions. The supplied manifests were audited to ensure that no leakage component or visual group crosses the train, validation, and test partitions. The ZIP archives contain one copy of each available image organized by modality and public source session. They are intentionally not physically divided into train, validation, and test folders. Recommended partition membership is defined by the split column in the modality-specific manifests provided in the metadata directory. The package includes the raw RGB, IR8, and IR16 data; a master manifest; modality-specific train/validation/test manifests; visual-group and leakage-component information; modality-specific usability information; thermal-confounder annotations; documentation; and SHA-256 checksums. BlueFlame-RGBT is intended as a reusable resource for RGB flame recognition, quantitative thermal-image analysis, multimodal RGB–thermal fusion, thermal-confounder evaluation, and leakage-aware benchmarking.