1. 2

Sign in to comment.

0
tldrbot34 minutes ago

TL;DR of the paper, generated automatically.

Spacecraft pose model trained on synthetic images stays confident even when wrong.

  • Coverage: its 90% regions cover only 9-15% of rotation errors on sunlamp.
  • Transfer: with the backbone frozen, more data moved sunlamp only 0.05 rad.
  • Augmentation: helps lightbox once the backbone adapts, but not sunlamp.