PFPNet_Pytorch

The unofficial implementation of ECCV 2018 paper "Parallel Feature Pyramid Network for Object Detection" in Pytorch. Some tricks learned from RefineDet.PyTorch

And the paper could be downloaded from "http://openaccess.thecvf.com/content_ECCV_2018/papers/Seung-Wook_Kim_Parallel_Feature_Pyramid_ECCV_2018_paper.pdf"

Contents

• 1.Environment

• 2.Preparation

• 3.Training

• 4.Demo

Environment

    Python3.5
    
    Pytorch1.1.0
    
    GPU:1080Ti

Preparation

1. Get the code. We will call the cloned directory as $PFPNet_Pytorch.

   https://github.com/junjieAI/PFPNet_Pytorch.git

2. Build the Cython modules, We can refer to Faster-Rcnn approach.

    cd $PFPNet_Pytorch/lib  
    make   
   

3. Download the basenet model VGGNET. By default, we assume the model is stored in $PFPNet_Pytorch/pretrained.

4. Download the trained model of myself, the Extract code: fh9uby. By default, we assume the model is stored in $PFPNet_Pytorch/pretrained.

5. Prepare the data basic structure.

    $VOCdevkit/                           # RootPath  
   
    $VOCdevkit/VOC2012                    # image sets, annotations, etc.  
   
    $VOCdevkit/VOC2012/Annotations                       # include .xml files. 
   
    $VOCdevkit/VOC2012/ImageSets/Main                    # include trainval.txt file.  
   
    $VOCdevkit/VOC2012/JPEGImages                        # include images.  
   

Training

1. Train your model on PASCAL VOC Format.

    cd $PFPNet_Pytorch
    python3 train_PFPNet.py
   

2. Train results, it will create two types file, '.pth' model and loss log file.

    # It will create model definition files and save snapshot models in:
    #   - $PFPNet_Pytorch/weights/PFPNet{input_size}_VOC_{iteration}.pth/
    # and the loss log in:
    #   - $PFPNet_Pytorch/'{}_{}_{}_{}:{}:{}loss.txt'.format(log_time.year, log_time.month, log_time.day,log_time.hour,'%02d'%log_time.minute, '%02d'%log_time.second)
   

Demo

1. Visualization.

    cd $PFPNet_Pytorch
    python3 demo.py
   

2. The script vis_loss.py is just to visualize the loss log.