The performance of overfit when finetuning and how to avoid it?

[Hao Chen]

Hi,

I finetune a model based on VGG for a pixel-wise prediction task. But my test results (images) show very similar, however, the train loss and validate loss donnot decrease a lot, they are still very big and unstable, does my model overfit? since my trainset for finetuning is 600 images initially, and 600*12 images after augmentation. 

My model only need fine one fully connecttion after the pool5 layer of VGG, on this condition, does it easy to overfit, if does, how to avoid it?

Many thanks!

[Carlos Treviño]

Hi!

I would suggest you to modify your network using 2 fully convolutional layers based on this paper:

arXiv:1411.4038
https://gist.github.com/longjon/ac410cad48a088710872#file-readme-md

Can you talk about the groundtruth dataset generation? I'm still struggling to make the image parsing with a FCNN.

Thanks in advance!