DeRPN: Taking a further step toward more general object detection

DeRPN is a novel region proposal network which concentrates on improving the adaptivity of current detectors. The paper is available here.

Recent Update

· Mar. 13, 2019: The DeRPN pretrained models are added.

· Jan. 25, 2019: The code is released.

Contact Us

Welcome to improve DeRPN together. For any questions, please feel free to contact Lele Xie ([email protected]) or Prof. Jin ([email protected]).

Citation

If you find DeRPN useful to your research, please consider citing our paper as follow:

@article{xie2019DeRPN,
  title     = {DeRPN: Taking a further step toward more general object detection},
  author    = {Lele Xie, Yuliang Liu, Lianwen Jin*, Zecheng Xie}
  joural    = {AAAI}
  year      = {2019}
}

Main Results

Note: The reimplemented results are slightly different from those presented in the paper for different training settings, but the conclusions are still consistent. For example, this code doesn't use multi-scale training which should boost the results for both DeRPN and RPN.

COCO-Text

training data: COCO-Text train

test data: COCO-Text test

Pascal VOC

training data: VOC 07+12 trainval

test data: VOC 07 test

Inference time is evaluated on one TITAN XP GPU.

( "*": On Pascal VOC dataset, we found that it is more suitable to train the DeRPN+R-FCN model with 900 proposals. For other experiments, we use the default proposal number to train the models, i.e., 2000 proposals fro Faster R-CNN, 300 proposals for R-FCN. )

MS COCO

training data: COCO 2017 train

test data: COCO 2017 test/val

Getting Started

1. Requirements
2. Installation
3. Preparation for Training & Testing
4. Usage

Requirements

1. Cuda 8.0 and cudnn 5.1.
2. Some python packages: cython, opencv-python, easydict et. al. Simply install them if your system misses these packages.
3. Configure the caffe according to your environment (Caffe installation instructions). As the code requires pycaffe, caffe should be built with python layers. In Makefile.config, make sure to uncomment this line:

WITH_PYTHON_LAYER := 1

4. An NVIDIA GPU with more than 6GB is required for ResNet-101.

Installation

1. Clone the DeRPN repository

   git clone https://github.com/HCIILAB/DeRPN.git
   

2. Build the Cython modules

   cd $DeRPN_ROOT/lib
   make

3. Build caffe and pycaffe

   cd $DeRPN_ROOT/caffe
   make -j8 && make pycaffe

Preparation for Training & Testing

Dataset

1. Download the datasets of Pascal VOC 2007 & 2012, MS COCO 2017 and COCO-Text.

2. You need to put these datasets under the $DeRPN_ROOT/data folder (with symlinks).

3. For COCO-Text, the folder structure is as follow:

   $DeRPN_ROOT/data/coco_text/images/train2014
   $DeRPN_ROOT/data/coco_text/images/val2014
   $DeRPN_ROOT/data/coco_text/annotations  
   # train2014, val2014, and annotations are symlinks from /pth_to_coco2014/train2014, 
   # /pth_to_coco2014/val2014 and /pth_to_coco2014/annotations2014/, respectively.

4. For COCO, the folder structure is as follow:

   $DeRPN_ROOT/data/coco/images/train2017
   $DeRPN_ROOT/data/coco/images/val2017
   $DeRPN_ROOT/data/coco/images/test-dev2017
   $DeRPN_ROOT/data/coco/annotations  
   # the symlinks are similar to COCO-Text

5. For Pascal VOC, the folder structure is as follow:

   $DeRPN_ROOT/data/VOCdevkit2007
   $DeRPN_ROOT/data/VOCdevkit2012
   #VOCdevkit2007 and VOCdevkit2012 are symlinks from $VOCdevkit whcich contains VOC2007 and VOC2012.

Pretrained models

Please download the ImageNet pretrained models (VGG16 and ResNet-101, password: k4z1), and put them under

$DeRPN_ROOT/data/imagenet_models

We also provide the DeRPN pretrained models here (password: fsd8).

Usage

cd $DeRPN_ROOT
./experiments/scripts/faster_rcnn_derpn_end2end.sh [GPU_ID] [NET] [DATASET]

# e.g., ./experiments/scripts/faster_rcnn_derpn_end2end.sh 0 VGG16 coco_text

Copyright

This code is free to the academic community for research purpose only. For commercial purpose usage, please contact Dr. Lianwen Jin: [email protected].