Key Points
We generate pixel-level pseudo-annotations for training a segmentation network from object image registration results. This combines object recognition, pose estimation, and segmentation without requiring additional manual region annotations.
Abstract
To pick up an object in a warehouse, a robot needs to know what it is, which way it faces, and which part of the image belongs to it rather than the background. Image registration aligns images of the same object to help identify it and estimate its orientation. However, conventional registration was not designed to separate the object from its background.
We add a learning system that finds the object's region in the image. Training such a system often requires people to label individual pixels, the small points that make up an image, as object or background. Our method automatically creates provisional labels from similarities between two aligned images. It uses these labels to learn to refine the object region and remove background pixels, without additional manual region labeling.
Tests on several object-image datasets showed more accurate region extraction than existing methods that learn from limited labeling. Object identification and orientation estimation also improved. The added system worked with several existing image registration methods.
Publications
- Shunsuke Yoneda, Go Irie, Masashi Nishiyama,
Canonical Plane Segmentation without Annotating Pixel-level Object Regions for Image Registration, IEEE Access, vol. 12, pp. 39484 - 39494, March 2024. - Shunsuke Yoneda, Go Irie, Takashi Shibata, Masashi Nishiyama, Yoshio Iwai,
Deep Segmentation Network without Mask Image Supervision for 2D Image Registration, Proceedings of International Workshop on Frontiers of Computer Vision (FCV), OS5-5, pp. 408 - 422, February 2022. [Short movie]
