Wrong viewpoint frequent issue for male killer whales

I’m not sure if this is more of a bug report or a feature request, but I am using flukebook for killer whales and did some checks to see how prevalent miss-assigned viewpoints was in my dataset (e.g. right sides automatically assigned as left sides). Looking at 133 photos from two sightings of several animals, 7.5% were assigned to the wrong side. It seems like mostly mature males (9 the 10 issues I found), likely because their fin shape is less typical/more symmetrical (e.g. L85). It would be great if some additional effort could be done to train the viewpoint classifier to have higher success with mature males. I’m happy to supply some photos for training if that would be helpful.

Here and here and here are some examples of the issue.

As is, 7.5% rate of mis-assigned viewpoints hinders efficient matching, given that each incorrect annotation needs to manually removed and re-added, which can be time consuming when a large volume of photos is processed.

1 Like

I spoke with my teammates and learned that this issue tends to arise from the images being taken from an angle that’s underrepresented in the dataset (like this one). Front-side and back-side angles are typically considered undesirable for fin/fluke photos because it obscures the relevant data.

Tagging @jason for visibility in case @CMKonrad’s offer of additional images is relevant to existing efforts to retrain FinFindr.

1 Like