[Feature] Transformation about kornia HOT 5 CLOSED

hi @versatran01 ! thanks for considering to contribute to the library :D

I totally agree with you about the naming convention. We are totally open to refactor it if that makes sense and the use of the package more clear. Could you list what do you have in mind to implement ? We could start from there to break down and define a proper API for the transforms submodule.

Regarding the epsilon value, this is to prevent NaN values during the backprop. It happened to me in the past that by missing this value, the gradients where quite unstable and prone to exploit. So that's a trick that pytorch core developers recommended to me to put in practice.

from kornia.

@versatran01 I put this link as reference: https://github.com/ros/geometry/blob/melodic-devel/tf/src/tf/transformations.py

some of the conversions were initially based on this.

from kornia.

The default epsilon value (1e-6) is too large. Datasets like ScanNet provide camera poses with very high precision (1e-5 ~ 1e-8). The epsilon value should be much smaller than this value.

Also, it would be good if the user can set epsilon by themselves. Currently "DepthWarper" called "relative_transformation" with default eps value (1e-6). The user who is using DepthWarper cannot set the eps value if they want high precision.

from kornia.

eps should not be part of the function signature, see #79 for more discussion.

from kornia.

closed in favor of #79

from kornia.