search

DLR-RM / AugmentedAutoencoder

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

[question] About the occlusion during generating training images. #67

Closed zeal-up closed 4 years ago

zeal-up commented 4 years ago

In the paper, it said: applying occlusions using random object masks or black squares. However, in the code I found that these two flags are set to be false:

 [Augmentation]
REALISTIC_OCCLUSION: False
SQUARE_OCCLUSION: False

Do you apply occlusions in your final experiments?

MartinSmeyer commented 4 years ago

Hi, Yes we apply occlusion in the form of spatial dropout, i.e. black squares, as this is cheaper to compute as realistic occlusions and has similar effects. See some lines below: Sometimes(0.5, CoarseDropout( p=0.2, size_percent=0.05) ),