Training a Caffe model using a small dataset

[Isaac Mercieca]

Hi guys,

I just started using caffe to train my own model. I have a very small dataset containing only 100 training images and 50 test images for each category. All images are of products one finds in a local grocery store, and also the test images are not in the same condition as the ones in the training images. Finally each category represents a common product category such as cereal, chocolate etc.  

I have a few images in regards to my dataset and network. 

• Can model be trained using my dataset or do I need to use a fine tuning approach being that my dataset is really small? 
  
• And also can this be achieved using a multi-label approach? 
  
• One last question, will my model be over-fitted due to the low number of training images?

Thanks a lot in advance for your help.

Regards,

Isaac

[ChanChip]

Hi Issac,

It may help u....

http://sites.duke.edu/rachelmemo/2015/04/03/train-and-test-lenet-on-your-own-dataset/

Thanks

[Du Guowei]

Hey Isaac,

I am doing a project using caffe too.Have you had any success with using fine tuning? I am also not sure how it works. What is your accuracy using 100 training images and 50 test images?

I am working on classifying 2 labels => 1) Traversible pathways 2)  Non-traversible pathways for path detection of autonomous vehicles.
How many training and val images is it recommended for me to get?

Thanks!