PyTorch Implementation of Multiple Instance Detection Network with Online Instance Classifier Refinement (OICR)

paper

How to Start

git clone http://www.github.com/jd730/OICR-pytorch

Dependencies

• Python 3.5 or higher
• Pytorch 0.4.0 (not 0.4.1)
• CUDA 8.0 or higher

Data preparation

• PASCAL_VOC 07+12: Please follow the instructions in py-faster-rcnn to prepare VOC datasets. Actually, you can refer to any others. After downloading the data, creat softlinks in the folder data/.

Selective Search

wget https://dl.dropboxusercontent.com/s/orrt7o6bp6ae0tc/selective_search_data.tgz
tar -xvf selective_search_data.tgz
rm -rf selective_search_data.tgz

move selective_search_data folder into data folder.

Pretrained Model

• VGG16: Dropbox, VT Server

Download them and put them into the data/pretrained_model/.

NOTE. We compare the pretrained models from Pytorch and Caffe, and surprisingly find Caffe pretrained models have slightly better performance than Pytorch pretrained. We would suggest to use Caffe pretrained models from the above link to reproduce our results.

If you want to use pytorch pre-trained models, please remember to transpose images from BGR to RGB, and also use the same data transformer (minus mean and normalize) as used in pretrained model.

Compilation

As pointed out by ruotianluo/pytorch-faster-rcnn, choose the right -arch in make.sh file, to compile the cuda code:

More details about setting the architecture can be found here or here

1. Install pip dependency pip install -r requirement.txt

2. Compile th ecuda dependencies cd lib & sh make.sh

Performance

Green rectangulars are the results of OICR and red rectangulars are ground truth.

Library description

trainval_net.py : main training code.

test_oicr.py : test code modified from oicr test code.

datasets/ : loading pascal_voc

roi_data_layer/ : loading batch, making roidb and minibatch

model/ includes network and roi_align,crop,pooling.

In model/oicr*/ there are two files. One is vgg network which assign each layers and the other is oicr class which decides how to make a forward and how network is composed of.

How to monitor

1. Use tensorboard --use_tb flag, but sometime tensorflow session is dead abruptly.
2. Use logger. In this code, the program automatically generate log.txt and progress.csv in your directory. You can check this using note.ipynb. You can easily understanding through reading the example code in the notebook.

file_path = 'jdhwang/1006_seq_tr3/log/progress.csv'
plot_reward_curve_seborn(file_path, mavg=True, mavg_v=1, n=N, target_field='midn_loss', print_header=False,txt_offset=1.0, newfig=False, conv=20)

Run Example

Training (multi GPU)

CUDA_VISIBLE_DEVICES=1,2 python3 trainval_net.py --dataset pascal_voc --net vgg16 \ --bs 4 --nw 4 --save_dir='output' --model='oicr' \ --lr 0.001 --cuda --disp_interval 50 --mGPUs --vis \ --checkpoint_interval=500

Testing

CUDA_VISIBLE_DEVICES=2 python3 test_oicr.py --dataset pascal_voc --net vgg16 --checkpoint 70000 --load_dir='output' --cuda --model='oicr'--vis

Notice

bs (batch_size) should be divisable by 2. On Caffe, batch is defined as how many forward operations before bacward, and the it does not divide the accumulated loss. To follow this definition and improve the performance, my code automatically forward twice and backward once without divison. See here

Reference

https://github.com/ppengtang/oicr

https://github.com/jwyang/faster-rcnn.pytorch