name | caffemodel | caffemodel_url | sha1 | gist_id |
---|---|---|---|---|
ParseNet on PASCAL |
VGG_VOC2012ext.caffemodel |
99cc76c373dc522fd70f80208b30a43ab2fba2f6 |
This is a model presented in the paper
ParseNet: Looking Wider to See Better
Wei Liu, Andrew Rabinovich, Alexander C. Berg
arXiv:1506.04579
This is the ParseNet model trained on PASCAL (using extra data from Hariharan et al. and finetuned from the fully convolutional reduced VGGNet).
You should be able to train/eval this model with http://github.com/weiliu89/caffe/tree/fcn. This branch introduces filter_stride (used for 'atrous' algorithm as described in Deeplab), L2-norm layer, evaluation code on the fly, etc.
The model should obtain 69.55 mean IoU on PASCAL 2012 segmementation val dataset. Please feel free to send me email (wliu@cs.unc.edu) if you have any questions.
Hello, I want train your ParseNet from scratch on pascal voc 2012 segmentation task. Follow your the experiment in your paper, I use the augmented dataset and the train_val.prototxt, solver.protoxt you provided above.
But I always get the all zero output which stand for background label for all pixels in the image. I also tried other learning rate and trained several times, but the results are all zero.
I think it's because the network is stuck in local minima, since the training set is in balance, the background label occupy nearly 75% of all pixels.
So, Have you ever meet this problem during your experiments?
If not, do you know what causes this problem ?
I have been bothered by this problem for two weeks, I also trained FCNs, DeconvNet and my segmentation networks, but the results are all the same, all zero.
Wish your reply, thankyou