Training Data Preparation for Caffe Classification vs. Caffe Faster RCNN

[Mato98]

I just started working with caffe and have a general question i could't answer with all your tutorials.

I worked sucessfully with Caffe Image Classification and also created a Caffemodel with my own data using the configuration from bvlc_reference_caffenet. My data is stored in a folder with all the images (only one class per image) and in train.text, i specified each image with its class.

Now i'm trying to use Faster RCNN (Python) Detection Demo, which works succesfully with the demo models VGG16 and ZF. My question is now, how to train a model (like VGG16 or ZF) with my own data?

• Can i use my previous Caffemodel from Classification for Faster RCNN?
• Can i use my existing training data with cropped images (only one class) or do i need other training images with multiple classes and annotation files for each with specified objectregion (like it is shown in most of the tutorials )?

Thanks

[foolyc]

Q1: using your previous caffemodel file as a pretrained model is ok, make sure that the convolution part of your faster-rcnn net has the same layer name with your previous classification net. Q2: your trainning data is enough, but you should better give the box information (some tools can be used such as labelImg ) as the target of bbox regression.