Is 78% Model Accuracy Good

I just finished my Lesson 1 Part 1 2019, and was playing with this dataset from kaggle :

It contains 2357 images of 8 different categories, and I think is a very similar problem to the cats breed vs dogs breed problem in Lesson 1.

I split 20% of the train dataset for a validation dataset, and implemented resnet34.

Using model fine tuning, I selected the suitable learning rate slice, and after 40 epochs, I reached an accuracy of 78%.

Is this the best that can be achieved, or can this be improved (and how)?

For what it’s worth I created a dataset of pictures of my two kids and couldn’t get the loss better that 25%. I tried resnet50 up to 20 runs and that was its best. I assume we’ll get more info on why further into the course.

My guess on my issue is that the faces are not dead centre and maybe resizing is cutting the face out a lot of the time.