Accurate perception of surgical instruments underpins automated skill assessment and computer-assisted training in minimally invasive surgery, but publicly available video-kinematic datasets are scarce, particularly for non-in-vivo tasks, with existing resources being image-only, entirely simulated, or captured on costly da Vinci platforms that do not represent non-robotic training realities. This work presents LASK (Laparoscopic Skill and Kinematics), a peg-transfer dataset featuring synchronised HD video and 7 degree-of-freedom ground-truth kinematics for two surgical graspers, collected on a standard laparoscopic box trainer using an NDI Aurora electromagnetic tracking system and jaw angle sensors. The dataset comprises 114 trials totalling approximately 3 hours from 38 low-skill, 41 medium-skill and 35 high-skill expert surgeons, providing 324,101 frames with time-aligned kinematics for both tools and tooltips, of which 3,725 frames carry annotated bounding boxes, together with a complete 2,680-frame validation sequence. Data were collected at the 7th and 8th Urology Boot Camps in Leeds in 2023 and 2024 and at the 2024 British Association of Paediatric Endoscopic Surgeons congress, and include surgeon-specific metadata on handedness (103 right-handed, 5 left-handed, 6 mixed-handed) and experience (84 adult and 30 paediatric surgeons). LASK captures two instruments throughout, with wider fields of view than typical in-vivo data and imaging conditions representative of box trainers, and is intended to support benchmarking of multi-class detection, tracking, pose estimation and skill classification, with preliminary multi-tool detection benchmarks using state-of-the-art YOLO models showing greater than 95% accuracy.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)