Publication year and journal Link

• 2018(CVPR 2018)
• https://arxiv.org/abs/1712.07629

Introduction

• Keypoint/Interest point detection are building blocks of many slam algorithms. Many previous algorithms such as (ORB, FAST, SIFT) are based on hand-crafted features.
• SuperPoint is self-supervised deep learning algorithms for keypoint detection i.e interest points.
• There are two types of keypoints:
  • Semantic keypoints: It has semantic meaning for objects in images. For e.g. left eye corner of face.
  • Interest points are low level points. For e.g .corner point of line.

Contribution

• MASK RCNN (ICCV 2017) and pifpaf (CVPR 2019) are representative learning methods for detecting semantic keypoints. These are supervised methods.

• SuperPoint don’t require explicit human annotations.

• It produces more correct and dense matches compared to other previous approaches.

  

Details

• It includes three steps: interest point pre-training, interest point self-labeling and joint training.