From the figures below, all models converge, among which Dense Convolutional Network (DenseNet) has the lowest training loss and evaluation loss. From the third figure, it can be observed that DenseNet has the highest accuracy over epochs compared to other classification models. DenseNet also has the highest accuracy with a relatively lower parameter number. From the training results, it seems that DenseNet has the best performance.
Pictures were taken in a grocery store to be used as testing data. From the figures above, it can be observed that the lighting reflection will lead to over exposure. Additionally, for the steaks packaged in boxes, the labels will cover parts of the meat.
Recall the initial purpose, we want to train a general model with the ability to adapt to the real world conditions. Although models perform well in the training/validation dataset, their accuracy rates drop while facing different unkonwn input. The winner "Densenet" above may be over-trained, and "InceptionNet" and "EfficientNet" may be robust to the light changes at the cost of little training accuracy.