Testing Caffe on ImageCLEF

[Qiaosong Wang]

Hi everyone,

I'm new to Caffe and I want to directly use Caffe for the ImageCLEF plant identification task:

http://www.imageclef.org/2011/Plants

The training and testing data are scan images of the leaves (white background) with a species category. There are about 5000 training images and 5000 testing images with about 70 categories. Sample image:

http://www.tela-botanica.org/appli:cel-img:000155003XS.jpg

My question is, should I directly use the imagenet model, resize all images to 256*256 and create a training txt file like this:

/path/1.jpg 0

/path/2.jpg 56

...

/path/5000.jpg 70

Then create the leveldb dataset, compute image mean and start training? Is there anything else I should be aware of or parameters I need to tune to achieve good accuracy?

Thank you.