PWISeg (PWISeg Surgical Instruments Dataset)

Overview

The Surgical Instruments Recognition Dataset is a groundbreaking collection of high-resolution images (1280x960 pixels) specifically designed for the recognition and categorization of surgical instruments. This dataset captures the intricate details and complexity of surgical tools, particularly when arranged in scenarios reminiscent of an operating room.

Data Acquisition

Our data collection approach was multifaceted to encompass a wide range of scenarios in which surgical instruments are typically viewed. We utilized: 1. Ceiling-mounted, program-controlled camera: This camera systematically rotated to capture the instruments from various angles. 2. Handheld camera: It was employed to take pictures of instruments in trays on the floor, creating images with natural shadows and lighting. 3. Real-world operating room imagery: We included close-up shots of instruments in use, offering a view akin to that of the surgical team, and capturing the tools in motion.

Annotations

The dataset adheres to the COCO results format, supporting object and keypoint detection tasks. For object detection, bounding boxes and class labels are provided for each instrument. Key point annotations identify precise instrument points, with over 10,000 instruments annotated throughout the dataset.

Dataset Structure

The dataset is divided into three parts: - Training set: 1788 images with annotations - Validation set: 200 images with annotations - Testing set: 185 images with annotations