A 35 GB open dataset of seven high-resolution retinal-surgery videos (≈ 87 k frames) with frame-level ground-truth annotations for:

• Tool-tip detection – bounding boxes + instrument type (light pipe, forceps, loop scrapers, tono scrapers)
• Depth classification – four categorical distances from the tool tip to the retinal surface (contact, near, intermediate, far)

Data were curated through a multi-stage consensus workflow involving three retinal surgeons and three trained graders; detailed methodology and baseline models are provided in the accompanying README. The archive is intended for computer-assisted surgery research, surgical-skill assessment, and general computer-vision benchmarking.
License: CC BY 4.0