.png){kind=icon}
.png){kind=icon}
Return
Shape-based 3D human action retrieval using triplet network
.png){kind=icon}
PRE
AI
DOI：10.1007/s11042-023-16211-1.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Human action retrieval is a challenging task in computer vision and computer graphics. Existing action retrieval works are mainly for the human representation of skeleton sequences or videos. However, there are fewer researches for 3D human shapes. In this paper, we propose a novel action retrieval method for 3D human point cloud sequences. Specifically, a triplet network with the margin loss is adopted to learn embedding vectors, where their Euclidean distances are close for pairs of point cloud sequences with the same actions and are far away for pairs of sequences with different actions. Given a query point cloud sequences, the retrieval results are in ascending order via the Euclidean distances of the embedding vectors. Furthermore, we also construct a 3D human action dataset, which consists of 220 classes for evaluation of action retrieval. Extensive experiments show that the proposed method is better than the existing skeleton-based methods with a 0.05 higher precision.
Keywords:
Human action retrieval
Point cloud sequences
Triplet network
Journal
IF:
3
Papers:
1.9W
Citations:
3.2W