Photo color correction

[Daniel A. A. Pelsmaeker]

Hello,

We are preparing to upload hundreds of photoquadrat photos in a source and use CoralNet to annotate these. We are currently preprocessing the photos (color correction, lens correction, cropping), and were wondering about the best approach to color correct.

We have two approaches to automated color correction that produce different results. What is in your experience more important for the training of the classifier: to have higher contrast or to have more vibrant (accurate?) colors?

Are there other aspects of preprocessing that affect the accuracy of the classifier?

Thanks for your time and for maintaining CoralNet!

Kind regards,

Daniel

[David Kriegman]

We don't have any very specific recommendations on correction methods as we have not performed any systematic studies of accuracy as a function of different preprocessing methods.

Other users may have experience that they can share.

Except in extreme cases of say highly saturated images, the most useful thing to do is to correct in a fashion that makes classes of interest to have similar colors. 

CoralNet trains a classifier specifically on images from your dataset, and so if there is wide variability in the appearance of a given class, then recognition accuracy will be more impacted than if they were consistent.

David