| Summary: | In this dissertation, we present a novel humanoid robot pose dataset and its construction process. Our purpose is to provide a high-quality dataset that enables more studies on humanoid robot pose related tasks. The goal is to design a dataset collection pipeline that can automatically collect, filter, annotate raw data collected from the internet. We use object detection, pose estimation and many other computer vision algorithms and techniques in the designing process. More than 750 raw images of humanoids were collected from the internet and 150 videos were crawled for all humanoid robot models. Detailed annotations were made on over 1000 images of humanoid robots. A survey was conducted on various humanoid robot models and their structures. Humanoid robot skeleton was constructed by determining the appropriate keypoints and their relation- ships, and a dataset collection pipeline was designed. Our pipeline can be used to collect data more efficiently and reduce time and labor costs.
|
|---|