In this final thesis we analyse a panoptic segmentation method on a water surface detection problem. Panoptic segmentation is capable of object detection segmentation and segmentation of areas, that can't be classified as objects. So that we can better analyse the results of panoptic segmentation, we decided to use combination of the semantic segmentation method WaSR and instance segmentation Mask R-CNN as a reference. The training of these methods is done on the MaSTr1325 dataset, meant for semantic segmentation, which is why we process it in to appropriate from for instance and panoptic segmentation. Result analysis of the trained methods is done on the MODS dataset with an already prepared evaluator. The results of semantic segmentation are very similar where WaSR still lightly takes the lead with 3,1% F1 score advantage in the danger zone and 0,3% advantage in the entire zone. In the analysis of instance segmentation we notice that there isn't much difference in the F1 score between Mask R-CNN and Panoptic-Deeplab methods, however the main reason for this is due to how Panoptic-Deeplab draws out bounding boxes. Because Panoptic-Deeplab shows as an effective method with a parallel object detection branch, we propose, as a subject of further work, a similar implementation to the WaSR method, adding it the capability of outputting panoptic segmentation.
|