Train the model on images without void space in corners

NotJoeMartinez / APDL_Quality_Control

Computer Vision Quality control for the Advanced Particle Detector Laboratory at Texas Tech University

0 stars 1 forks source link

Closed NotJoeMartinez closed 3 years ago

NotJoeMartinez commented 3 years ago

The activation maps seem to indicate that the model picks up on these black corners on the augmented images

Train the model on images that do not produce these dark corners
Modify the augmentations script to replicate jordans dataset
- 480x480
- No duplicate images
- The dataset balance is below

NotJoeMartinez commented 3 years ago

Fixed with mode=edge using skimage library but this introduced another problem with image stretching