Self-produced Guidance for Weakly-supervised Object Localization

We train the SPG model on the ILSVRC dataset, and then apply the trained model on video sequences of DAVIS 2016. <img width="400" height="200" src="figs/bear_loc.gif"/><img width="400" height="200" src="figs/dog_spg_c.gif"/>

Overview of SPG

Train

We finetune the SPG model on the ILSVRC dataset.

cd scripts
sh train_imagenet_full_v5.sh

Test

Download the pretrined model at GoogleDrive(https://drive.google.com/open?id=1EwRuqfGASarGidutnYB8rXLSuzYpEoSM (IMAGENET),https://drive.google.com/open?id=1WfrELBlEoq5WO7gKUv-MLTQ8QHY-2wiX (CUB)).

Use the test script to generate attention maps.

cd scripts
sh val_imagenet_full.sh

Demo

Thanks to Jun Hao for providing the wonderful demos!

Please see the setup_demo.txt for more guidance of setuping up the demos.

Masks are getting better with the proposed easy-to-hard approach.

Citation

If you find this code helpful, please consider to cite this paper:

@inproceedings{zhang2018self,
  title={Self-produced Guidance for Weakly-supervised Object Localization},
  author={Zhang, Xiaolin and Wei, Yunchao and Kang, Guoliang and Yang, Yi and Huang, Thomas},
  booktitle={European Conference on Computer Vision},
  year={2018},
  organization={Springer}
}