Real-Time System for Detection Hidden and Visible Keypoints of Rainbow Trout

In this paper, we describe the experience of applying deep learning methods to the detection of eight keypoints of rainbow trout. Using the two loss functions combination and hyperparameters tuning, our solution achieves performance comparable with state-of-the-art results of human keypoints detection. The obtained result enables the digital transformation of aquaculture industries and the automation of production processes, through the implementation of a wide range of video analysis algorithms providing such capabilities as non-contact fish counting, weighing, behavior analysis, and diagnosis of diseases.