Dealing with small images in dataset

Part 1 (2018)
1

I am working on a practice problem called Age-Prediction hosted by Analytics Vidhya. https://datahack.analyticsvidhya.com/contest/practice-problem-age-detection/

The size of the images in the dataset is quite varied:

image
image.png873×714 53.1 KB

I have trained with sz = 16,32,64,100 with resnet50, resnext101 and inceptionv4 but unable to make any significant improvements on my score. Its a 3-class classification problem and I am stuck at the ~0.7 mark on accuracy.

Anyone has ideas on how to get the fastai framework to work with small images but of substantial proportion in the dataset?

2

The kaggle seedlings competition has the same type of size issue. I just trained at sz=299 and it worked great for me…