Presence-Only Geographical Priors for Fine-Grained Image Classification

Paper: Link

Objective:

• Explores how can we use additional meta-data available to make better classification (in this case animal species).
• Explores how to make best use of additional meta data which comes with most images today.

Summary:

• Knowing where a given image was taken can provide a strong prior for what objects it may contain.
• Paper provide a novel training loss to capture these relationship
• The data they assemble can have unrelated image and location dataset as long as both contain the same categories.
• At test time, given an image and where and when it was taken, they aim to estimate which category it contains.
• Location information is incorporated as bayesian spatio-temporal prior. Also, during the modelling, spatio temporal (longitude, latitude, time) are independent from the image classifier
• It is difficult and time consuming to have information on where and when a given category has been a.) observed to be present and b.) observed to be absent. Hence, the paper explores presence-only setting (novelty)