SSM (the Unofficial Version of Pytorch Implementation)

Towards Human-Machine Cooperation: Self-supervised Sample Mining for Object Detection

Keze Wang, Xiaopeng Yan, Dongyu Zhang, Lei Zhang, Liang Lin

Sun Yat-Sen University, Presented at CVPR2018

License

For Academic Research Use Only!

Strict Requirements

Python 3.6

OpenCV

PyTorch 0.3

Note: PyTorch 0.4 or Python 2.7 is not supported !

Citing SSM

If you find SSM useful in your research, please consider citing:

@inproceedings{wang18ssm,
    Author = {Keze Wang, Xiaopeng Yan, Dongyu Zhang, Lei Zhang, Liang Lin},
    Title = {{SSM}: Towards Human-Machine Cooperation: Self-supervised Sample Mining for Object Detection},
    Journal = {Proc. of IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    Year = {2018}
}

Dependencies

The code is built on top of https://github.com/ruotianluo/pytorch-faster-rcnn. Please carefully read through the pytorch-faster-rcnn instructions and make sure pytorch-faster-rcnn can run within your enviornment.

Datasets/Pre-trained model

1. In our paper, we used Pascal VOC2007/VOC2012 and COCO as our datasets, and res101.pth model as our pre-trained model.

2. Please download ImageNet-pre-trained res101.pth model manually, and put them into $SSM_ROOT/data/imagenet_models

Usage

1. training

   Before training, please prepare your dataset and pre-trained model and store them in the right path as R-FCN. You can go to ./tools/ and modify train_net.py to reset some parameters.Then, simply run sh ./train.sh.

2. testing

   Before testing, you can modify test.sh to choose the trained model path, then simply run sh ./test.sh to get the evaluation result.

Misc

Tested on Ubuntu 14.04 with a Titan X GPU (12G) and Intel(R) Xeon(R) CPU E5-2623 v3 @ 3.00GHz.