Using Fast-RCNN with a ILSVRC2012 model?

[Anu Chandra]

Hi,

I’m new to caffe. I was using the bvlc_reference_caffenet.caffemodel to do some basic detection. I recently came across Fast-RCNN and have got the demo working. The demo gives the option of using three models which appear to use the PASCAL VOC classes ie 20 different object categories. Is there a way to use the Fast-RCNN with the something similar to the bvlc_reference_caffenet.caffemodel which uses ILSVRC2012 1000 categories? Ideally, I want to re-write the demo code to use a ILSVRC2012 model.

Cheers

[georgi angelov]

Bump! I'm also interested in that. The 20 classes are decent but more will definitely be better.

[Sean Bell]

Sure, you could train one by computing your own selective search proposals, and then use the ImageNet 200 category detection challenge data.

You could probably actually just use the existing 1000-category ImageNet model to detect all 1000 classes and use some heuristic that if the max probability is low, then the box counts as background.  This will be missing bbox regression and won't do as well as a properly trained model, but at least it has the right number of categories.

[Evgeny Smirnov]

If you want Fast-RCNN detector with 1000 ImageNet categories, you should train your own model using something like this method:

http://arxiv.org/abs/1407.5035

воскресенье, 28 июня 2015 г., 18:04:24 UTC+3 пользователь Anu Chandra написал:

[X.T. Li]

Hi! I am also interested in training fast-rcnn on imagenet, did you successfully training the model?

在 2015年6月28日星期日 UTC+8下午11:04:24，gobob写道：

[Rafael Ruiz Muñoz]

Did you figure out how to make it?

I've trained a lot of regular Caffemodels but I haven't got idea how to train this with boxes.