Learning to Identify Local Flora with Human Feedback (Author's Manuscript)
Indiana University Bloomington United States
Pagination or Media Count:
In this ongoing work, we are developing a method that involves a user in the loop to aid in the fine-grained recognition of a diverse set of tree species. Instead of asking users to provide attributes of trees, we instead ask them to judge the similarity between pairs of tree images, and then use this to learn the parameters of a discriminative distance metric for use with k-nearest neighbors. Over time, the discriminative distance function becomes a better approximation to the humans judgment of visual similarity. We present baselines and results of our human-guided approach on a collection of 20 tree species from five geographic locations.