Algorithms can help automatically analyze images and videos, but they usually need to be taught what to do first! There are already a lot of efficient algorithm to categorize pictures and videos with big mammals on them. But we are working with species not commonly studied with cameras : 🐸🐍🦎 (learn why here). We therefore need to gather a lot of data with our new camera technology, in order to train the algorithm to recognize the species of interest.
Now that we have received the data of the first month, we are manually going through them and labelling them to train the algorithm. At this stage we are screening the data video by video… to truly perfect the algorithm later on, this will have to be done frame by frame! But that is a problem for future us.
In the image above, you can see a screenshot of the first results : the algorithm is drawing a “box” around the animal it thinks it has detected. It also shows the label “snake” and the how confident the algorithm is about its identification (0.81 probability).
What happens when you don’t properly train an algorithm? During the lockdown in 2020, the data collection was obviously postponed. We had a few videos from a pilot study, but not enough to properly train an algorithm. So we tried a random algorithm for fun, only trained to detect common objects and animals (NOT including reptiles and amphibians).
Here are the results when you don’t properly trained an algorithm! 🤨