DLR-RM / AugmentedAutoencoder

Official Code: Implicit 3D Orientation Learning for 6D Object Detection from RGB Images
MIT License
341 stars 97 forks source link

About Distance Estimation #83

Closed ghost closed 4 years ago

ghost commented 4 years ago

Hello,

I'm studying your paper and trying to understand the estimation of the object distance from the camera.

I have some problems due to the diag_bb, since for different object orientation this value causes some issues.

For example taking into accont the following image

in

where the left object is the closest and the center object is the farthest. The z estimation wrongly estimate the distances due to the diag_bb which depends on object orientation.

Any explanation? Thanks in advance

MartinSmeyer commented 4 years ago

Hi @Zaxorn,

you are correct the diag_bb depends on the object orientation. That is why we save a different diag_bb for every possible orientation when we create the codebook. At inference time you first predict the orientation and then retrieve the corresponding diag_bb of the synthetically rendered view. Afterwards you compute its size ratio with the observed diag_bb in the real image. As long as the orientation estimate is not too far off, this should yield a decent (+-3cm) distance estimate.