Skip to content

v0.4.1

Choose a tag to compare

@yz3440 yz3440 released this 12 Feb 22:26
· 3 commits to main since this release

Bug Fixes

  • Fix perspective-to-sphere coordinate conversion: Replaced naive angle addition (local_yaw + yaw_offset, local_pitch + pitch_offset) with proper 3D rotation via perspective_to_sphere() when converting OCR bounding box coordinates to world spherical coordinates. The old approach ignored the coupling between yaw and pitch in 3D rotation, causing ~10° coordinate errors when the camera had a non-zero pitch offset or when text was near the edges of perspective images.

Impact

  • The same text detected in overlapping perspective views now maps to identical world coordinates (previously off by up to 10°+), enabling correct deduplication across views
  • Bounding box width/height are now computed from world-space corner positions with proper yaw wrap-around handling