Automated Object Keypoints Dataset Generation Using Blender

Abstract

This research paper explores the use of Blender, a powerful 3D computer software, as a convenient alternative to the traditional hand labeling method for generating keypoints datasets for object pose estimation. By leveraging Blender's capabilities and automating the generation of labeled keypoints data; this offers significant time and resource savings when compared to manual labeling, thereby reducing the workload of keypoints annotation. This study aims to compare the speed and model accuracy achieved by using the Blender-generated dataset against the conventional hand labeling approach. Through experimentation, the synthetic datasets generated from Blender demonstrated an accuracy that rivals traditionally hand-labeled datasets yet demanded a fraction of the time. Shortening the process of taking pictures and labeling keypoints of objects is desirable for manipulation tasks such as service robots set up in unknown environments.