This is the Food 101 dataset, also available from https://www.vision.ee.ethz.ch/datasets_extra/food-101/
“Food-101 – Mining Discriminative Components with Random Forests” by Lukas Bossard, Matthieu Guillaumin and Luc Van Gool cir 2014 had an average accuracy of 50.76%.
Was able to achieve 78% accuracy on test set (train/test split) using an GTX 1080ti over a period of 2 days. Training images were heavily augmented. Started off with a non-trainable VGG16 image-net pretrained model. Investigated and analyzed the features within the bottleneck using TensorBoard and other tools; determined sufficient feature transfer to continue; expanded out to 3 dense layers.
VGG16 is quite amazing, obviously VGG19 does better when dealing with more classes and I didn’t expect this model to do so good with this little work on such a high number of classes.