Trying to train ImageNet model with CNN (R-CNN)

Question

After months of working with caffe , I was able to successfully train my models. For example, in addition to my own models, I was able to train ImageNet with 1000 classes.

In my project, I am now trying to extract an area of ​​my interest class. After that, I compiled and ran the Fast R-CNN demo and it works fine, but the sample models only contain 20 classes, and I would like to have more classes, for example, all of them.

I have already loaded the bounding fields from ImageNet with real images.

Now I am gone, I can not understand the next steps, and there is no documentation on how to do this. The only thing I found is how to train the INRIA human model, and they provide a dataset + annotations + python script.

My questions:

• Is there any tutorial or guide that I missed?
• Is there an already prepared model with 1000 classes capable of classifying images and extracting bounding fields?

Thank you in advance.

Sincerely.

Raphael.

Answer 1

Dr. Ross Hirschik has done a great job finding objects. You can learn a lot from your detailed git on fast RCNN : you should find a caffe branch there, with a demo. I have not used it myself, but it seems very understandable.

Another area you might find interesting is LSDA : using weak control to teach object detection for many classes.

By the way, did you view fast-rcnn ?