RON: Reverse Connection with Objectness Prior Networks for Object Detection

RON is a state-of-the-art visual object detection system for efficient object detection framework. The code is modified from py-faster-rcnn. You can use the code to train/evaluate a network for object detection task. For more details, please refer to our CVPR paper.

***There is also a tensorflow re-implementation of RON at RON_Tensorflow, thanks @HiKapok!

Citing RON

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

@inproceedings{KongtCVPR2017,
    Author = {Tao Kong, Fuchun Sun, Anbang Yao, Huaping Liu, Ming Lu, Yurong Chen},
    Title = {RON: Reverse Connection with Objectness Prior Networks for Object Detection},
    Booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
    Year = {2017}
}

PASCAL VOC detection results

Method	VOC 2007 mAP	VOC 2012 mAP	Input resolution
Fast R-CNN	70.0%	68.4%	1000*600
Faster R-CNN	73.2%	70.4%	1000*600
SSD300	72.1%	70.3%	300*300
SSD500	75.1%	73.1%	500*500
RON320	74.2%	71.7%	320*320
RON384	75.4%	73.0%	384*384

MS COCO detection results

Method	Training data	AP(0.50-0.95)	Input resolution
Faster R-CNN	trainval	21.9%	1000*600
SSD500	trainval35k	24.4%	500*500
RON320	trainval	23.6%	320*320
RON384	trainval	25.4%	384*384

Note: SSD300 and SSD500 are the original SSD model from SSD.

RON Installation

0. Clone the RON repository

   git clone https://github.com/taokong/RON.git
   
   

1. Build Caffe and pycaffe

   cd $RON_ROOT/
   git clone https://github.com/taokong/caffe-ron.git
   cd caffe-ron
   make -j8 && make pycaffe
   *this version use CUDNN for efficiency, so make sure that "USE_CUDNN := 1" in the Makefile.config file.
   

2. Build the Cython modules

   cd $RON_ROOT/lib
   make
   

3. installation for training and testing models on PASCAL VOC dataset

   3.0 The PASCAL VOC dataset has the basic structure:

    $VOCdevkit/                           # development kit
    $VOCdevkit/VOCcode/                   # VOC utility code
    $VOCdevkit/VOC2007                    # image sets, annotations, etc.
   

   3.1 Create symlinks for the PASCAL VOC dataset

    cd $RON_ROOT/data
    ln -s $VOCdevkit VOCdevkit2007
    ln -s $VOCdevkit VOCdevkit2012
   

4. Test with PASCAL VOC dataset

   Now we provide two models for testing the pascal voc 2007 test dataset. To use demo you need to download the pretrained RON model, please download the model manually from BaiduYun(Google Drive), and put it under $data/RON_models.

   4.0 The original model as introduced in the RON paper:

    ./test_voc07.sh
    # The final result of the model should be 74.2% mAP.
   

   4.1 A lite model we make some optimization after the original one:

    ./test_voc07_reduced.sh
    # The final result of the model should be 74.1% mAP.
   

5. Train with PASCAL VOC dataset

   Please download ImageNet-pre-trained VGG models manually from BaiduYun(Google Drive), and put them into $data/ImageNet_models. Then everything is done, you could train your own model.

5.0 The original model as introduced in the RON paper: 

    ./train_voc.sh
    
5.1 A lite model we make some optimization after the original one:

    ./train_voc_reduced.sh