I'm building a model to recognize fishes in the aquarium (150 different fishes).
I'm using a faster_rcnn_inception_v2_coco_2018_01_28 model for transfer learning from TF object detection API. I have labelled around (80 fishes) and (55) get detected only.
when I labelled more (30 fishes) and added it to the model and retrained it again, the number of detected fishes reach to (66 fishes), but some of the fishes which detected in the previous model are not detected after adding more classes and retrain the model.
I don't know why that happened. and I have some concerns about the model.
While labelling the images, each image contains multiple fishes types, and I labelled the clear ones only.
Should I label the images that taken when I applied zoom in to the camera to focus on some fishes, in this case, the fish will appear in a bigger size than the real size.
Some of the fishes are big and some of them are small and some of them are too small, is it applicable to detect all of them in one model?
Are there any suggestions to deal with similar fishes?
Is there another benefit for training using GPUs other than reducing time-consuming for training?
Thanks in advance for helping.
Update: I have added more class (20 classes more) and the count of detected fishes reach to ( 87) but not more, even I put it for more time in training.