Creating a Dataset for the Detection of Flying Birds

We present a pipeline for collecting a dataset for the detection of flying birds in videos. We treat the creation of a dataset as an iterative task, which allows to train an inference system periodically, and to improve gradually its performance. We follow a three stage pipeline, that includes repetitive video recording, video annotation and frame sampling for training a Deep Learning detector. Furthermore, we would like to point out that the end goal of our work is to protect birds flying through wind farms, in order to take protective measures. Therefore it is crucial to find the moment that a bird enters and exits the field of view of the camera. Thus we use our dataset to setup a Temporal Action Detection (TAD) task, and we evaluate our inference pipeline using an appropriate measure.