A Data Augmentation Method for Temporal Action Segmentation in Laparoscopic Surgery Videos

Abstract

This report proposes a data augmentation method to improve the accuracy of fine-grained temporal action segmentation in laparoscopic surgery videos. Unlike previous methods that applied faster scaling (e.g., 2 × or higher), the proposed method enables slower temporal changes, such as scaling to 1.5 ×. This augmentation is selectively applied only to segments whose lengths exceed the average segment length of their respective action classes. We demonstrate the effectiveness of the proposed approach by applying it to MS-TCN, MS-TCN++, and their ensemble model.