Beta testing is underway! Please contact us if you would like to get involved.
Our dream is to one day be able to identify specific individuals of every species using artificial intelligence, thus allowing users to get incredibly accurate species counts, alongside detailed data on species movement patterns and other habits. However, this is a very challenging problem with current technology – the ability to identify individuals has only been achieved for specific species, under very specific conditions which do not convert well to camera trap images. It is even very challenging for humans, and who can only do this for specific species, in the right conditions.
As such, we are approaching this goal in baby steps, focusing on only one species at a time, beginning with species that are known to be individually identifiable from camera-trap data. Initially, we plan to build an interface that helps humans to perform this task more effectively, by providing suggestions based on simple heuristics – images of the same species, of the same sex, with the same camera orientation. Further, we intend to limit the suggestions to images of animals that are separate temporally and spatially in such a way that they could physically be at the location of the currently-viewed image, at that point in time. Thereafter, once we have generated sufficient data, we can begin to train an AI model that gives ranked match suggestions for each image in the above-mentioned interface that still meet the heuristic requirements. This will make the individual-identification process easier for a human workers, allowing them to generate more data, more quickly, in turn allowing us to train better models.
We would like to add the ability to add additional information to each individual animal detection such as its age and sex. This information could then be used to improve the results one can generate, as well as assist in determining the approximate group size of a species contained in an image cluster, and improve suggestions for the individual identification process.
Unknown-Animal Classification Interface
When a cluster only contains a blur, a whited-out close up, a tail, or some whiskers it can be almost impossible for the average person to identify the species of animal contained in the images, and as such, these are typically labelled as unknown. We would like to provide a dedicated interface for trying to identify these unknown species by providing the context of the chronologically previous and next clusters along with their times, and general species statistics of the camera involved. This interface could then be used by the appropriate experts in your team to attempt to identify the animals using all the possible information at hand.
An additional piece of information that can be useful in statistical analyses is the number of individual animals in the group captured within a cluster. Ideally, one would like to be able to identify the individuals in the group and generate a count that way, but in the absence of an accurate AI for that purpose, this is unreasonable. Instead, we can generate an approximate minimum count based on the maximum number of males and females in any one of the images. However, such a count would be very rough for any large group, so we would like to additionally provide a dedicated interface for a human worker to manually determine the group size based on the images in each cluster.