Showing 1–1 of 1 results for author: Minoi, J L
-
Unsupervised Segmentation of Action Segments in Egocentric Videos using Gaze
Authors:
I. Hipiny,
H. Ujir,
J. L. Minoi,
S. F. Samson Juan,
M. A. Khairuddin,
M. S. Sunar
Abstract:
Unsupervised segmentation of action segments in egocentric videos is a desirable feature in tasks such as activity recognition and content-based video retrieval. Reducing the search space into a finite set of action segments facilitates a faster and less noisy matching. However, there exist a substantial gap in machine understanding of natural temporal cuts during a continuous human activity. This…
▽ More
Unsupervised segmentation of action segments in egocentric videos is a desirable feature in tasks such as activity recognition and content-based video retrieval. Reducing the search space into a finite set of action segments facilitates a faster and less noisy matching. However, there exist a substantial gap in machine understanding of natural temporal cuts during a continuous human activity. This work reports on a novel gaze-based approach for segmenting action segments in videos captured using an egocentric camera. Gaze is used to locate the region-of-interest inside a frame. By tracking two simple motion-based parameters inside successive regions-of-interest, we discover a finite set of temporal cuts. We present several results using combinations (of the two parameters) on a dataset, i.e., BRISGAZE-ACTIONS. The dataset contains egocentric videos depicting several daily-living activities. The quality of the temporal cuts is further improved by implementing two entropy measures.
△ Less
Submitted 23 June, 2021; v1 submitted 30 September, 2017;
originally announced October 2017.
