How to train fast-rcnn on ImageNet Detection Dataset?

X.T. Li

unread,

Sep 6, 2015, 5:47:03 AM9/6/15

to Caffe Users

I want to train fast rcnn on imagenet 200 detection dataset. Before I started, I want to make sure that I am not inventing wheels again.

Is there any existed work on this goal?

Thanks!

林子達

unread,

Sep 6, 2015, 9:41:55 PM9/6/15

to Caffe Users

In my experience, if you are not planing to training bounding box regressor, you can train your model by following standard Caffe training process. However, the training image in your dataset should be cropped by its bounding box. After you are done with training, you should change the last pooling layer to ROIPoolingLayer in deploy.prototxt

X.T. Li

unread,

Sep 7, 2015, 4:55:01 AM9/7/15

to Caffe Users

Thanks.

Actually, I plan to train on imagenet by methods similar to this one. This one uses fast rcnn code directly with bounding box regression.

在 2015年9月7日星期一 UTC+8上午9:41:55，林子達写道：

Marx HA

unread,

Oct 27, 2015, 7:55:52 AM10/27/15

to Caffe Users

Can you paste conten of annotation.txt?

X.T. Li

unread,

Nov 2, 2015, 1:48:21 AM11/2/15

to Caffe Users

The annotation is the same format as pascal voc in xml. I managed to train on imagenet thanks to this link: https://github.com/zeyuanxy/fast-rcnn/tree/master/help/train
在 2015年10月27日星期二 UTC+8下午7:55:52，Marx HA写道：

Ali Lemus

unread,

Nov 5, 2015, 5:43:23 AM11/5/15

to Caffe Users

Hi X.T. I'm on the same path as you it seems, also interested in the "faster" (http://arxiv.org/abs/1506.01497) version, but I want to start here: https://github.com/zeyuanxy/fast-rcnn/tree/master/help/train
did you manage? any recommendations?

X.T. Li

unread,

Nov 8, 2015, 9:17:51 PM11/8/15

to Caffe Users

I wrote some notes on my blog about training fast rcnn on imagenet.

在 2015年11月5日星期四 UTC+8下午6:43:23，Ali Lemus写道：

Chiou Meng-Jiun

unread,

Nov 26, 2015, 3:51:53 AM11/26/15

to Caffe Users

Hi everyone,

Following X.T. Li's blog and zeyuanxy's instruction, after dealing with some other errors I also train Fast-RCNN successfully on my dataset. Here is my instruction of training a new Fast-RCNN on another dataset. I also provided solutions to some common errors may be encountered when trying to training.

M.-J. Chiou

2015年11月9日月曜日 10時17分51秒 UTC+8 X.T. Li:

Ali Lemus

unread,

Dec 3, 2015, 7:50:31 PM12/3/15

to Caffe Users

Im sorry it took me so long to say thanks!!!

I actually managed to train imagenet with the help of your blog!!! thank you so much!!!

I had actually tried to do it myself, but your code was much simpler than mine and noticed some things that I had gotten wrong like the -1 in the annotation files.

Thanks Again!!!

Reply all

Reply to author

Forward