Authors
A Ferone, M Lazzaro, VM Scarrica, A Ciaramella and A Staiano
Abstract
Computing a reliable estimate of optical flow is a fundamental step in many computer vision applications, mainly based on the analysis of the motion of objects in the scene. Therefore, in recent years, many research applications have led to numerous solutions and methodologies for its estimation. In recent literature, several approaches based on deep neural networks have been proposed for the estimation of optical flow, performing with high accuracy when trained on large labeled datasets. Since the difficulty of obtaining optical flow ground truths from real-world image sequences, many efforts have been devoted to producing large synthetic datasets for effectively training such models. Nevertheless, being able to obtain an accurate estimate of the flow in real-life scenarios is still a great challenge due to the complexity of the environments. The problem is even more complex if an underwater environment is considered, due to sudden changes in lighting, water turbidity, movements of the background, particles, and other objects. In this perspective, our work presents a synthetic dataset of underwater scenes, endowed with optical flow labels, to demonstrate the benefits of training a specific deep neural model for estimating optical flow in the considered environment. To this aim, an experimental comparison between a general-purpose deep neural model and the same model trained specifically with the new proposed dataset has confirmed an increase in accuracy of the final estimation.