VGGNets for Scene Recognition

Here we release our trained VGGNet models on the large-scale Places205 dataset, called Places205-VGGNet models, from the following report:

http://arxiv.org/abs/1508.01667

Places205-VGGNet Models for Scene Recognition
Limin Wang, Sheng Guo, Weilin Huang, and Yu Qiao, in arXive 1508.01667, 2015

Performance on the Places205 dataset

Model	top-1 val/test	top-5 val/test
Places205-VGGNet-11	58.6/59.0	87.6/87.6
Places205-VGGNet-13	60.2/60.1	88.1/88.5
Places205-VGGNet-16	60.6/60.3	88.5/88.8
Places205-VGGNet-19	61.3/61.2	88.8/89.3

We use 5 crops and their horizontal flippings of each image for testing.

Performance on the MIT67 and SUN397 dataset

Model	MIT67	SUN397
Places205-VGGNet-11	82.0	65.3
Places205-VGGNet-13	81.9	66.7
Places205-VGGNet-16	81.2	66.9

We extract the fc6-layer features of our trained Places205-VGGNet models, which are further normalized by L2-norm.

Download

• Places205-VGGNet-11:
  http://mmlab.siat.ac.cn/Places205-VGGNet/siat_scene_vgg_11.caffemodel
• Places205-VGGNet-13:
  http://mmlab.siat.ac.cn/Places205-VGGNet/siat_scene_vgg_13.caffemodel
• Places205-VGGNet-16:
  http://mmlab.siat.ac.cn/Places205-VGGNet/siat_scene_vgg_16.caffemodel
• Places205-VGGNet-19:
  http://mmlab.siat.ac.cn/Places205-VGGNet/siat_scene_vgg_19.caffemodel
• Mean file:
  http://mmlab.siat.ac.cn/Places205-VGGNet/places205_mean.mat

These models are relased for non-conmercial use. If you use these models in your research, thanks to cite our above report.

Multi-GPU Implementation

In order to speed up the training procedure of VGGNets, we use a Multi-GPU extension of Caffe toolbox:

https://github.com/yjxiong/caffe/tree/action_recog

Meanwhile, we add the strategies of multi-scale cropping and corner cropping provided by this extension, which has been proved to be effective for action recognition in videos.

Questions

Contact

• Limin Wang
• [Sheng Guo] (mailto:sheng.guo@siat.ac.cn)
• Weilin Huang