philkr
Custom Data Set
You can generate proposals using propose_hf5.py for any images. If you want to train the proposals you'll need to load your dataset in C++.
How do I load my dataset? I'm new to this so I don't know exactly what to do.
Do you want to train or just run the proposal method?
On Friday, July 17, 2015 09:39:41 AM Christopher Alicea-Nieves wrote:
How do I load my dataset? I'm new to this so I don't know exactly what to do.
Reply to this email directly or view it on GitHub: https://github.com/philkr/lpo/issues/19#issuecomment-122335933
I want to generate the proposals and then pass it to Caffe in order to train a network using Fast R-CNN (R. Girshick).
Just use and/or modify the propose_hf5.py script and run it on your images. There is no need to load your dataset into lpo.
@philkr this part is confusing. the lpo seems to be trained on VOC2012, but if I want to use it to propose boxes for some 50 categories (say even flowers), will this work? my objective is similar to @cuatristapr
Despite being trained on VOC2007, it generalized quite well for example to MS COCO. See the paper for details. Some of the intuition for this behavior might be that from a low level perspective objects can be quite similar, and the proposals don't focus much on class, but much rather on boundary cues ... The same effect has been observed by many other proposal systems.
Philipp
On Monday, November 23, 2015 07:08:22 PM pradeepj247 wrote:
@philkr this part is confusing. the lpo seems to be trained on VOC2012, but if I want to use it to propose boxes for some 50 categories (say even flowers), will this work? my objective is similar to @cuatristapr
Reply to this email directly or view it on GitHub: https://github.com/philkr/lpo/issues/19#issuecomment-159138247