Core Concepts
EYOC proposes an unsupervised method for distant point cloud registration, adapting to new data distributions without global pose labels.
Abstract
EYOC introduces a progressive self-labeling scheme to train a feature extractor for distant point cloud registration. The method includes spatial filtering and speculative registration to improve correspondence quality. Experiments show comparable performance with supervised methods but with lower training costs.
Key points:
- EYOC is an unsupervised method for distant point cloud registration.
- It uses a progressive self-labeling scheme and spatial filtering to improve correspondence quality.
- Experiments demonstrate comparable performance with supervised methods at a lower training cost.
Stats
Experiments show that EYOC can achieve comparable performance with state-of-the-art supervised methods at a lower training cost.
Quotes
"In this paper, we propose Extend Your Own Correspondences (EYOC), a fully unsupervised outdoor distant point cloud registration method requiring neither pose labels nor any input of other modality."
"We evaluate EYOC design with trace-driven experiments on three major self-driving datasets, i.e., KITTI [16], nuScenes [6], and WOD [43]."