SavING the SCANDINAVIAN Arctic Fox

The Arctic fox lives far up in the north and is adapted to a life in the high mountains and an Arctic climate. Even though there are many Arctic foxes globally, they are rare in Scandinavia. Their numbers here began to decline dramatically in the mid-1800s; despite long efforts to protect the species, this decline has continued into the present day. Changes to rodent life history patterns and a growing population of red fox are the primary reasons for this, but a number of other factors are also involved. The Arctic fox is therefore now a critically endangered species in both Sweden and Norway. The present-day Arctic fox population is so small that it is entirely dependent on conservation measures for its long term survival. Understanding of the reasons why the numbers of Arctic fox have declined means that we are able to take the necessary action to begin rebuilding the population. Currently, these actions mainly take place as part of research projects, which makes it possible to evaluate their effects and subsequently improve the methods that are used. [1]

SAVING PHD Students from boring work

There are multiple scandinavian research projects that seek to map the arctic fox populations. The primary method of research is gathering population statistics by observation. This is done by placing automated camera traps in multiple locations with known arctic fox populations.

The camera takes automatically a photo every 10 or 15 minutes.

Each season can generate hundreds of thousands of images that are then manually processed by PhD students. 99% of the images contain nothing. It is tedious work, but it also requires concentration as an animal may be difficult to spot.

 

The Deep Fox system uses a combination of a 3D convolutional network, an LSTM and a direct attention sampler mechanism controlled by the LSTM

The Deep Fox system uses a combination of a 3D convolutional network, an LSTM and a direct attention sampler mechanism controlled by the LSTM

Our system, DeepFox, detects and tags all animals in each frame, eliminating the need for manual processing. It can handle high resolution images and images where the presence of an animal is implicitly inferred based on surrounding frames. 

Deep Fox was developed in collaboration with the Department of Zoology @ Stockholm University, who provided us with all the data painstakingly labeled by their PhD students.

The system is general in nature and can be trained on any labeled image data, so it is not limited to detecting animals in photos, but we felt that helping the arctic fox conservation effort was a good starting point.