Thermal AI improvements

We have rolled out some improvements to our thermal AI models, both on the server and on the camera.

Improved accuracy

A recent update to our thermal AI model has increased the accuracy of our thermal predictions by 51% to 74% across our test set of 54,000 recordings. We still have relatively poor performance for sheep and deer, but these will improve as we get more data. Many thanks to everyone who has been tagging thermal videos.  Your tags definitely help. Please keep it up.

This confusion matrix shows the model performance. For each animal you can see it's accuracy and what it will identify the animal as if it doesn't get it right. You can see for most animals it gets the classification right, or will mark it as unidentified.

On camera model

Note: This hasn't been released yet, but will be shortly.

Faster

The new model on the camera is faster. Under optimum conditions the time to process a video has dropped from 0.9s to 0.76s. This is mainly of use for those using the camera to control a trap. At the moment there are a handful of people testing this. It means the trap will be able to arm faster when it sees a fast moving predator such as a stoat. The accuracy improvements above also apply to the on-camera classification.

Post processing

Up until now we have done "real-time" classification on the camera. This means the camera broadcasts what it thinks it is seeing as quickly as it can. This is useful for controlling a trap, but is not so useful for monitoring. For monitoring it does not need to be real-time and the whole video can be analysed to estimate what animal or animals were most likely present. We have added an option under configuration setting in advanced to do this.

If you tick post process then the recording will be processed after it has been finished.