liz6688 / Hide-and-Seek

Hide-and-Seek (HaS)

Code for the [Hide-and-Seek: Forcing a Network to be Meticulous for Weakly-supervised Object and Action Localization, ICCV 2017] Krishna Kumar Singh, Yong Jae Lee (http://krsingh.cs.ucdavis.edu/krishna_files/papers/hide_and_seek/hide_seek.html) and [Hide-and-Seek: A Data Augmentation Technique for Weakly-Supervised Localization and Beyond] Krishna Kumar Singh, Hao Yu, Aron Sarmasi, Gautam Pradeep, Yong Jae Lee. If you use our work, please cite:

@inproceedings{singh-iccv2017,
  title = {Hide-and-Seek: Forcing a Network to be Meticulous for Weakly-supervised Object and Action Localization},
  author = {Krishna Kumar Singh and Yong Jae Lee},
  booktitle = {International Conference on Computer Vision (ICCV)},
  year = {2017}
}

@inproceedings{singh-arxiv2018,
  title = {Hide-and-Seek: A Data Augmentation Technique for Weakly-Supervised Localization and Beyond},
  author = {Krishna Kumar Singh, Hao Yu, Aron Sarmasi, Gautam Pradeep, and Yong Jae Lee},
  booktitle = {Arxiv},
  year = {2018}
}

Pre-requisites

1. Torch (http://torch.ch/docs/getting-started.html)
2. For training use the code https://github.com/soumith/imagenet-multiGPU.torch
3. For the visualization and generating Class Activation Maps(CAM) use the code https://github.com/metalbubble/CAM

Training

1. Please download train.lua and opts.lua and replace it in https://github.com/soumith/imagenet-multiGPU.torch
2. The new code has two additional arguments patchSize and hideProb. patchSize decides the size of the patch to be hidden. For example to hide the patches of size 32 give argument -patchSize 32. Multiple patch sizes can be provided seperated by comma, for example -patchSize 0,16,32,44,56. Here, 0 indicates no patch will be hidden. hideProb indicates by what probability patches will be hidden. For example to hide patches with 50% probability give the argument -hideProb 0.5.

Pre-trained Models

1. AlexNet-HaS-Mixed: https://drive.google.com/open?id=1QIrXJV5Sw0eYyXjauW6SlxkQXe3uDnmL
2. GoogLeNet-HaS-32: https://drive.google.com/open?id=1N3zgRmD0trCMfYOw1vo_DbesW4Ug7qx5
3. Please subtract mean and divide by standard deviation (meanstdCache.t7). For class ordering refer classes.t7.
4. ResNet-50-HaS(trained for classification task): https://drive.google.com/open?id=1CkrXmpqDtXGOTiL4v381WJ5Rn_tg32G1

Data Augmentation

If you need to hide the image patches for data augmentation, please refer the code hide_patch.py. This can be used for both PyTorch and MXNet.

Results

Method	No HaS	HaS	Boost
Weakly-supervised object localization	54.90	58.75	+3.85
Weakly-supervised semantic seg	60.80	61.45	+0.65
Weakly-supervised action localization	34.23	36.44	+2.21
Image classification	76.15	77.20	+1.05
Semantic segmentation	48.00	49.31	+1.31
Emotion recognition	93.65	94.88	+1.23
Person re-identification	71.60	72.80	+1.20

Person re-identification Comparison

	Market-1501		Duke
Methods	Rank-1	mAP	Rank-1	mAP
IDE+CamStyle	87.6	67.3	74.8	52.4
IDE+CamStyle+Random Erasing	89.4	71.5	78.3	57.6
IDE+CamStyle+HaS	90.2	72.8	79.9	57.2

Visualization of AlextNet-HaS for object localization:

Click Here

Contact

Please contact [e-mail] for any questions.