search

42dot / VFDepth

VFDepth Self-supervised surround-view depth estimation with volumetric feature fusion
Apache License 2.0
93 stars 11 forks source link

World to cam coordinates #1

Closed sandhawalia closed 1 year ago

sandhawalia commented 1 year ago

Hi there,

Thanks for the amazing work. This seems like a bug. World (voxel) to cam should use camera_extrinsics and not camera_extrinsics_inv ?

JungHeeKim29 commented 1 year ago

Hi, thank you for your interest in our work. And our backproject function aims to obtain the inverse warping function, which maps 3D voxel points to camera viewpoints. Here, camera extrinsic indicate associated mapping with the reference of the common vehicle frame, and therefore we applied the inverse of extrinsic, and multiplied with intrinsic matrix to project 3D voxel points(x,y,z) into pixel coordinates.

sandhawalia commented 1 year ago

Hi there,

Indeed looking at DDAD this is how Extrinsics are defined (cam to world), though its opposite of the convention

DDAD

Each sensor has associated extrinsics mapping it to a common vehicle frame of reference (datum['extrinsics']).

Kornia

These terms are usually expressed by the joint rotation-translation matrix which is also known as the extrinsic matrix. It is used to describe the camera pose around a static scene and transforms the coordinates of a 3D point from the world coordinate system to the camera coordinate system.

Thanks for the clarification. closing the issue.