yolov5x6 detection model 1280x1280 trained on set of 901902-uavs images from a small number of classes (22). mAP50 of 0.672. best performance on Kelp and Bird classes
Input images: 1163; transformed images: 24564
30713 Labeled localizations
Input labels: {'Whale': 1, 'Kelp': 19122, 'Jelly': 287, 'Foam': 376, 'Bird': 4016, 'Sea_Lion': 32, 'Reflectance': 226, 'Wave': 11, 'Buoy': 9, 'Otter': 16, 'Boat': 50, 'RIB': 10, 'Person': 4, 'Egregia': 12, 'Seal': 6, 'Shark': 15, 'Wood': 4, 'Mooring': 1, 'Mola': 5, 'Secci_Disc': 3, 'Batray': 1, 'Surfboard': 1};
Transformed labels: {'Whale': 4, 'Kelp': 53377, 'Jelly': 901, 'Foam': 1055, 'Bird': 12528, 'Sea_Lion': 103, 'Reflectance': 586, 'Buoy': 17, 'Otter': 49, 'Boat': 93, 'RIB': 32, 'Wave': 10, 'Egregia': 34, 'Seal': 18, 'Shark': 53, 'Wood': 8, 'Mooring': 4, 'Mola': 14, 'Secci_Disc': 8, 'Batray': 4, 'Person': 6, 'Surfboard': 4}
After training, 60 epochs, VAL results
30k Class Images Labels P R mAP50 mAP50:95
all 2483 7185 0.83 0.591 0.672 0.464
Bird 2483 1452 0.879 0.886 0.929 0.634
Boat 2483 12 0.367 0.75 0.477 0.216
Buoy 2483 3 0.82 1 0.995 0.813
Egregia 2483 5 0.876 1 0.995 0.792
Foam 2483 116 0.727 0.879 0.888 0.613
Jelly 2483 89 0.686 0.888 0.846 0.485
Kelp 2483 5415 0.803 0.855 0.887 0.664
Mola 2483 3 1 0 0 0
Otter 2483 5 0.815 1 0.995 0.672
Person 2483 2 0.997 0.5 0.543 0.46
RIB 2483 6 1 0 0.542 0.184
Reflectance 2483 54 0.793 0.759 0.772 0.487
Sea_Lion 2483 13 0.875 0.923 0.956 0.791
Shark 2483 5 0.468 0.6 0.531 0.317
Surfboard 2483 1 1 0 0.0765 0.0689
Wave 2483 2 1 0 0.995 0.697
Wood 2483 2 1 0 0 0