fine-tune prelearned faster-rcnn to recognize new classes

rbgirshick / py-faster-rcnn

Faster R-CNN (Python implementation) -- see https://github.com/ShaoqingRen/faster_rcnn for the official MATLAB version

Other

8.11k stars 4.11k forks source link

fine-tune prelearned faster-rcnn to recognize new classes #100

Open Inret opened 8 years ago

Inret commented 8 years ago

good day

i'm trying to solve problem – recognize new classes, via using prelearned (voc2007) caffe model on faster-rcnn net. Because of lack of image collection to learn (about 100 images, juice packages on group photos on store shelf, 6 kinds (classes)), the idea is to use existent prelearned model. froze values of weights in bottom and middle layers, erase weights in classification layers and learn them using backprop (finetune)

could you write some helpful comments - could this idea lead to success?, and some links or comments how to do it technically

thank you

davidkh1 commented 8 years ago

Hi, I'd also like to fine-tune "Faster R-CNN" with my small dataset of new classes. Please, could share your experience with this task? Thanks!

andrewliao11 commented 8 years ago

Hi all, I don't know if it's suitable for faster rcnn, but here is the paper that discussing how to finetune the CNN Towards Good Practices for Very Deep Two-Stream ConvNets

shaih82 commented 8 years ago

I am trying this approach now, freezing the RPN net, and finetunning the RCNN part so far i haven't been too successful... I think that i am lacking a more rich negative examples

xylcbd commented 7 years ago

watch it.

zhouhao94 commented 6 years ago

@andrewliao11 How would you do that? Can you share your code? Thank you.