Speeding up image annotation and AI-model training for wildlife images

Wildlife research is going through a technological revolution by increasingly relying on remote sensing, putting more effort in the conversion of complex signals to data. Camera-traps being the spearhead of that revolution, numerous approaches, workflows, pipelines, online platforms and packages have been proposed to help users convert images to data. Users without specialized training may quickly feel overwhelmed by the variety of these tools, and the most accessible, general-purpose solutions rarely offer satisfying performance either in manual processing efficiency or automated image classification. Hence, as can be seen from the rising number of publications presenting new wildlife classification models, custom models often need to be trained to answer research objectives, a daunting task for wildlife practitioners without computer-vision training. Here, we take the reader through a Timelapse and R-based pipeline, starting with the annotation of raw images and ending with the training of a custom AI-model. This pipeline focuses on efficient image-level manual annotation and transfers the tedious object-detection and instance segmentation tasks to pre-trained general models. We demonstrate the advantages of using this pipeline through a proof-of-concept using a dataset on North American small mammals, dividing manual annotation time by five and generating models with performance metrics (F1 score, mAP) reaching over 90% for all classes despite heavily unbalanced class frequency. To help readers have a smooth introduction to computer-vision work, we provide all the code underlying this pipeline and detailed instructions online.