A network is a training model? A model is only corresponding to a set of data? Said simple point is, for example, I can better the program have gone on to the cat did the semantic segmentation figure, but I also want to make it to the dog, people figure, also do semantic segmentation glass figure, can I use other data sets to train it? One thousand training finished it the cat map segmentation effect becomes poor again, I understand is that the network to determine their own a set of parameters, weight and migration of what, this group of parameters will be according to the training of more and more have been adjusted, destroyed one thousand for training, finally become "can split but can split very well",,

And, the same network performance difference under different data sets is normal, a data set assessment mIoU 0.8, 0.3, another only, asked other people, a lot of to add class, is just with a new data set to train? Just contact the depth of the neural network learning, very white, blind area there are a lot of knowledge, thanks to answer