• Files

Arxiv   Supp   Poster

• Codes

  Torch Version   PyTorch Version   Caffe Version

 Yanzhao helped upgrade the torch version. Caffe version is coming soon.

• Models

  SP-GoogLeNet(torch) for ImageNet


In this work, we propose a network component called Soft Proposal, which can be inserted after any CNN layer.
The Soft Proposal map iteratively evolves and jointly optimizes with convolutional filters, thus provides objectness information during the end-to-end learning procedure of CNN.
CNN equipped with Soft Proposal can be led to learn better object-centric filters, discover more discriminative visual evidence, and suppress background interference.


  title={{Soft Proposal Networks for Weakly Supervised Object Localization}},
  author={Zhu, Y. and Zhou, Y. and Ye, Q. and Qiu, Q. and Jiao, J.},