ProstaTD

ProstaTD: A Large-scale Multi-source Dataset for Structured Surgical Triplet Detection

Yiliang Chen, Zhixi Li, Cheng Xu, Alex Qinyang Liu, Xuemiao Xu, Jeremy Yuen-Chun Teoh, Shengfeng He, Jing Qin

Download Access:

Please find our dataset at https://github.com/SmartHealthX/ProstaTD

Abstract

ProstaTD is a large-scale surgical triplet detection dataset curated from 21 robot-assisted prostatectomy videos, collectively spanning full surgical procedures across multiple institutions, featuring 60,529 annotated frames with 165,567 structured surgical triplet instances (instrument-verb-target) that provide precise bounding box localization for all instruments alongside clinically validated temporal action boundaries. The dataset incorporates the ESAD and PSI-AVA datasets with our own added annotations (without using the original data annotations). We also include our own collected videos. It delivers instance-level annotations for 7 instrument types, 10 actions, 10 anatomical/non-anatomical targets, and 89 triplet combinations (excluding background). The dataset is partitioned into training (14 videos), validation (2 videos), and test sets (5 videos), with annotations provided at 1 frame per second.