search

pidahbus / deep-image-orientation-angle-detection

MIT License
83 stars 20 forks source link

Images with varying lens distortion effects #9

Open dedkamaroz opened 1 year ago

dedkamaroz commented 1 year ago

Firstly, this is an excellent implementation of a very clever paper/topic - so thank you for your efforts!

Do you have any advice or tips on how best to deal with images that have some sort of a geometric distortion?

eg. Optical Distortion that can make straight lines curved, etc.. and Perspective Distortion where a square subject looks like a trapezoid, etc..

Even after completing training using a dataset of 150 perfectly aligned images, I find that the angles being detected have an error rate greater than 50% when performing inference on a large dataset filled with images, many of which have one or both types of distortion.

I can easily generate hundreds (and thousands) of known angle images for the test/validation phase but editing even 150 more images to be oriented to 0 degrees is very time consuming and inaccurate.

pidahbus commented 1 year ago

Is it possible to share a few examples of the distorted images you are talking about?