search

ShaoqingRen / faster_rcnn

Faster R-CNN
Other
2.71k stars 1.22k forks source link

train with my own dataset, what files should be prepared？ #42

Open Pratyeka opened 8 years ago

Pratyeka commented 8 years ago

Hi all: I would like to train faster-rcnn on my own dataset, but i do not know what data files should be prapared？? i.e. the images, annotations? Any reference blogs or tutorials will be highly appreciated. Thanks

rayzhang0346 commented 8 years ago

I setup a .mat with imdb and roidb for both training and testing datas according to the structure of VOC sample in imdb/cache.

duygusar commented 8 years ago

@rayzhang0346 I have cleaned the cache and converted my data and annotations into VOC format along with image sets list etc. But then I get this error: "proposal_prepare_image_roidb>scale_rois (line 110) Non-singleton dimensions of the two input arrays must match each other." I don't see why this is not enough, why did you set a .mat for imdb and roidb? Is it because your data format is different than VOC?

rayzhang0346 commented 8 years ago

@duygusar Yes, i read through the code and find the key difference is VOC devkit. So i write code to deal with my data & notations and output the same format as imdb & roide. That's how I skip the data_prepare process. I am afraid i can't help you with your error.

duygusar commented 8 years ago

@rayzhang0346 That's alright, thanks for your answer. Well I was hoping I wouldn't run into problems since I converted my dataset to VOC format instead. Anyhow, I will try to see what's wrong

fengyuxi55 commented 8 years ago

@duygusar , I got same error as yours, and I don't know what's going wrong. If you find anything, plz tell me

duygusar commented 8 years ago

@fengyuxi55 I have just remembered I have the same problem while training a DPM too, I basically had to rewrite the whole xml parser part because it was not able to retrieve inner tags. So it makes sense to use own script to retrieve imdb and roidb

fengyuxi55 commented 8 years ago

@duygusar , thanks for your advice. I rewrite the xml parser to read my annotation files, the error disappear. Now I'm running the training code to train my own data.

arasharchor commented 8 years ago

@duygusar , I would like to retrain faster-rcnn using the the following dataset: http://www.svcl.ucsd.edu/projects/anomaly/dataset.htm could you tell me how to start to do so?