Training issues

[JYC59]

Hi, me and my team have three questins about training.
Q1. We used a large image that includes various species of corals to train our model. However, we would like to confirm if the generated results are accurate. If not, how to optimize the model to improve accuracy?
Q1 result photo link:

Q2. We also attempted to generate datasets using ten identical images of the same coral species. We would like to know if the results obtained from this training process are accurate. If not, how to optimize the model to improve accuracy?
Q2 result photo link:

Q3. If we want our model to be applicable to various coral images, what should we do?

[maxcorsini]

Hi,

Regarding you questions:

In the first case the performance are good (mIoU greater than 0.9 is a good result). The training graph shows that there is an overfitting problem with the data, so pay attention that we expect that this network will work well for similar orthoimages but not for non-similar ones.

Q2: For ten images you mean ten orthoimages or ten tiles ? In this last case, the number of images is very low to obtain reliable results.

Q3: difficult to say. It depends on the number of classes and their distributions on the maps. For example, if you have RARE classes we expect that the network will have difficulties to find them. Before to launch the training you can select the classes that the custom classifiers will recognize. I suggest you to do some tests and choose the best configuration. I know that this takes time but without seeing your data and the relative distributions is difficult to give you a proper answer.

Best,