Est F-scores in each video test cases. The highest PHA-543613 Protocol F-score of 0.79 was reached with all the algorithm using Mask R-CNN trained with all augmentationsSustainability 2021, 13,ten ofSustainability 2021, 13, x FOR PEER REVIEW10 ofAmong all the studied detectors, testing of your algorithm using the baseline model expectedly showed the lowest F-scores in both video test circumstances. The highest F-score of 0.79 C6 Ceramide Technical Information applied to thewith the algorithm utilizing Maskof the “Haul-back” video augmentations was reached “Towing” case video. Within the case R-CNN trained with all case, the algorithm using the “Towing” case video. Within the case of your “Haul-back” video case, the algorithm applied to Mask R-CNN trained with CP, geometric transformations and cloud augmentationMask R-CNN educated with CP, geometric transformations and cloud augmentation with showed a slightly higher F-score than that on the algorithm together with the detection primarily based showed a slightly greater F-score than that of on the model trained with all augmentations.the algorithm using the detection primarily based around the model educated with all(Table A1) containing the values of the calculated Precision, Recall The explicit table augmentations. The explicit table categories in the two case videos the presented in Appendix A. and F-score for all four(Table A1) containing the values of are calculated Precision, Recall and F-score for all four obtained within the two the videos are presented with all augmentaThe detection examples categorieswith applying caseMask R-CNN educated in Appendix A. The detection examples the “Towing” and “Haul-back” video frames with all augmentations tions as a detector onobtained with using the Mask R-CNN educated are presented in Figure as five. a detector around the “Towing” and “Haul-back” video frames are presented in Figure 5.Figure 5. Multi object detection examples obtained from the model trained with all tested augmentations and applied to: Figure five. Multi object detection examples obtained in the model educated with all tested augmentations and applied to: (A) “Towing” test video and (B) “Haul-back” test video with all the higher rate of occlusions and circumstances variation. (A) “Towing” test video and (B) “Haul-back” test video with the higher price of occlusions and circumstances variation.3.3. Comparison of Automated and Manual Catch Descriptions three.three. Comparison of Automated and Manual Catch Descriptions Automated count estimated per frame ofof the test videos was closer toground truth Automated count estimated per frame the test videos was closer towards the the ground truth count in theof the on the “Towing” test(Figure (Figure six), supporting the algorithms’ count within the case case “Towing” test video video six), supporting the algorithms’ higher Fhigher F-scores (Figure four). During the “Haul-back”, the automated count of Nephrops had scores (Figure four). During the “Haul-back”, the automated count of Nephrops had a tendency atowards underestimation by both algorithms,algorithms, whereas of round fish and flat tendency towards underestimation by both whereas within the case in the case of round fish and flatan opposite trend of overestimation was observed.wasthe case with the the case of fish classes fish classes an opposite trend of overestimation In observed. In other class, the other class, the algorithm based “Cloud” augmentations approximated the actual count the algorithm primarily based on instruction with on education with “Cloud” augmentations approximated the real count far better in comparison with the algorithm output with all test augment.
