Using simulated scenes for improving water surface obstacle detection

Avtonomna plovila se zanašajo na robustne metode zaznavanja ovir. Najsodobnejše metode temeljijo na segmentacijskih mrežah, ki so naučene na velikih naborih podatkov. Ker je nabor realnih slik zelo omejen, ročna segementacija pa časovno zahtevna in podvržena človeškim napakam, predlagamo alternativo –- izdelavo simulacijskih slik s samodejno segmentacijo. V nalogi predlagamo simulacijsko okolje za generiranje simuliranih vodnih scen in njihovih segmentacijskih mask. Analizirane so možnosti izboljšave segmentacijskih mrež za detekcijo ovir na vodi preko uporabe generiranih simuliranih vodnih scen. Predstavljamo primerjavo rezultatov učenja segmentacijske mreže na zbirki realnih slik z rezultati učenja mrež na zbirki simuliranih vodnih scen ter z rezultati učenja mrež na kombinaciji obeh zbirk. Rezultati analize kažejo, da je F-mera znotraj nevarnega območja z mrežami, ki so naučene na kombinaciji obeh zbirk, za 5 % višja kot z mrežami, ki so trenirane na zbirki brez sintetičnih slik, F-mera za celotno območje pa je za 0,3 % višja.Autonomous vessels rely on robust obstacle detection methods. State-of-the-art methods are based on segmentation networks that are trained on large datasets. Since the dataset of real images is very limited, and manual segmentation is time-consuming and subject to human error, we suggest an alternative -- the creation of simulation images with automatic segmentation. In this paper, we propose a simulation environment for generating simulated water scenes and their segmentation masks. Possibilities of improving segmentation networks for detection of obstacles on water through the use of generated simulated water scenes are analyzed. We present a comparison of the results of training the segmentation network on a dataset of real images with the results of training the networks on the dataset of simulated water scenes and with the results of training the networks on a combination of both datasets. The results of the analysis show that the F-measure within the danger zone with networks learned on the combination of both datasets is 5 % higher than with networks trained on the dataset without synthetic images, and the F-measure for the whole area is 0.3 % higher