See how accurate and precise the inatVisionAPI implementation is compared to the existing implementation. Ideally we could configure the algorithm to get exactly the same results. Then, we could tweak the configuration to see if we could improve accuracy and precision. Ultimately it would be great to have an evaluation that shows improvements, and settle on an optimal configuration for an initial release of the new common ancestor approach for end users
Comparison has been done. In a test of 2000 observations today with our working version of the common ancestor algorithm, looking at completely random observations, we find on average:
92.8% of all images get a common ancestor
99.6% of the time that common ancestor is consistent with the target taxon
0.4% of the time that common ancestor is inconsistent with the target taxon (i.e. incorrect)
92.4% of all images get a common ancestor that is correct
See how accurate and precise the inatVisionAPI implementation is compared to the existing implementation. Ideally we could configure the algorithm to get exactly the same results. Then, we could tweak the configuration to see if we could improve accuracy and precision. Ultimately it would be great to have an evaluation that shows improvements, and settle on an optimal configuration for an initial release of the new common ancestor approach for end users