| Paper ID | D-3-3.5 |
| Paper Title |
Data reduction using cluster sampling |
| Authors |
Ye seung Park, Mingyu Jang, Jungwoo Huh, Kyoungoh Lee, Sanghoon Lee, Yonsei University, Republic of Korea, Korea (South) |
| Session |
D-3-3: Image and video processing based on deep learning |
| Time | Thursday, 10 December, 17:30 - 19:30 |
| Presentation Time: | Thursday, 10 December, 18:30 - 18:45 Check your Time Zone |
|
All times are in New Zealand Time (UTC +13) |
| Topic |
Image, Video, and Multimedia (IVM): Special Session: Image and video processing based on deep learning |
| Abstract |
It is natural to use larger and more diverse datasets to get better performance in pose study. Learning with a large scale is essential to improve the model performance to a level similar to human recognition, but there is a problem with gradually increasing learning time and data redundancy. This can also lead to a lack of data storage. Our study proposes a new way to solve these problems: Data shaping Using Cluster Sampling (DUCS). In this paper, we propose a sampling framework that clusters a pose dataset and extracts only a small number of random frames from each cluster. To ensure the consistency of pose data, the data is normalized, and a preprocessing process of aligning the entire joint based on the pelvic joint is performed, and an optimal parameter search in DBSCAN is proposed to improve the performance of clustering. This process can greatly reduce the redundancy due to the specific posture bias. To demonstrate the effectiveness of our method, we trained a 3D pose estimation model with sampled datasets of Human3.6M and shown competitive results despite the drastic compression rate of over 95%. |