Training FCN using pretrain ResNet model

[Terry Chen]

I use ResNet50 model to finetune on Pascal Context dataset, and got only 31.41 mean I/U after 80k iterations, while I use VGG 16 model and get 33.7 mean I/U.  Anyone has same experience for FCN training with ResNet?

[Xin Li]

Hi, I wonder what's the batch size when you fine-tuning ResNet50, and how is the training speed?

What about ResNet101 and ResNet152?

在 2016年2月28日星期日 UTC+8上午3:28:39，Terry Chen写道：

[Forrest Iandola]

This isn't code, but the following paper does a broad set of experiments with ResNet + FCN. I found it interesting:

Zifeng Wu, Chunhua Shen, Anton van den Hengel. High-performance Semantic Segmentation Using Very Deep Fully Convolutional Networks. arXiv:1604.04339, 2016.

    https://arxiv.org/pdf/1604.04339v1.pdf

[Howard Mahé]

Hi Terry,

I'm also trying to train FCN with ResNet50. Did you succeed in integrating ResNet50 inside the skip architecture of FCN-8s ?
I still get some issues with crop layer due to 'noninteger crop offset'. I believe the source of this noninteger offset issue remains in the conv3, slide2 layer in ResNet.

So, if anybody ever succeed in generating a prototxt of ResNet50 (or more) inside FCN with skip architecture, please let me know.

[ido ariav]

Hi,

just to bump this topic.

has anyone succeeded in using ResNet as the base network for FCN and care to share a prototxt??

thanks

[Laurent Denoue]

Have you seen this? https://github.com/azavea/raster-vision/tree/develop/src/rastervision/semseg/models