TPD: Thermal Pose Dataset

TPD (Thermal Pose Dataset) is a paired RGB and long-wave infrared (LWIR) human pose dataset for surveillance-oriented 2D human pose estimation. It contains 1,895 registered frame pairs across 12 activity classes, with 2,364 people annotated on a 17-point extended-MPII skeleton. Existing thermal pose benchmarks are largely confined to controlled indoor scenes, fixed subject-to-camera distances, and limited activity diversity. TPD was captured across institutional interiors (public libraries, cafes, gyms), open outdoor spaces subject to solar thermal loading (parks), and high-occlusion natural cover (dense foliage), at subject-to-camera distances of 3 to 8 metres with more than 15 subjects of both sexes and varying body types. Every thermal frame has a spatially registered visible-light counterpart sharing the same filename stem, so the dataset supports thermal-only training, RGB-only training, and cross-modal work. Contents — thermal/ and rgb/ (1,895 PNG pairs each, 680x576); labels/ (17 lines of x y per person, plus skeleton edge list); annotations_coco/ (COCO keypoint format, full set plus fold-0 train/val); splits/ (scenario-aware 5-fold manifests); inter_annotator_agreement/; annotation_previews/. Annotation quality — a stratified 144-image subset was independently re-annotated by a second annotator without access to the primary labels. Over the 177 paired people, agreement is 2.12 px mean keypoint deviation, 99.7% PCK@0.2, and 0.923 mean OKS. Those labels are included so it can be reproduced. Acquisition — synchronised bi-spectral Hikvision DS-2TD2628T-3/QA, 256x192 vanadium-oxide uncooled FPA. The released 680x576 resolution is a common working grid, not the sensor's native resolution. TPD releases individually selected still frame pairs, not continuous video. Reference benchmark — fifteen models evaluated under a single protocol; best zero-shot transfer 12.1 px MPJPE, in-domain fine-tuning 7.9 px.