Single-Shot Refinement Neural Network for Object Detection

By Shifeng Zhang, Longyin Wen, Xiao Bian, Zhen Lei, Stan Z. Li.

Introduction

We propose a novel single-shot based detector, called RefineDet, that achieves better accuracy than two-stage methods and maintains comparable efficiency of one-stage methods. You can use the code to train/evaluate the RefineDet method for object detection. For more details, please refer to our arXiv paper.

System	VOC2007 test mAP	FPS (Titan X)	Number of Boxes	Input resolution
Faster R-CNN (VGG16)	73.2	7	~6000	~1000 x 600
YOLO (GoogLeNet)	63.4	45	98	448 x 448
YOLOv2 (Darknet-19)	78.6	40	845	544 x 544
SSD300* (VGG16)	77.2	46	8732	300 x 300
SSD512* (VGG16)	79.8	19	24564	512 x 512
RefineDet320 (VGG16)	80.0	40	6375	320 x 320
RefineDet512 (VGG16)	81.8	24	16320	512 x 512

Note: RefineDet300+ and RefineDet512+ are evaluated with the multi-scale testing strategy. The code of the multi-scale testing has also been released in this repository.

Citing RefineDet

Please cite our paper in your publications if it helps your research:

@article{zhang2017single,
  title = {Single-Shot Refinement Neural Network for Object Detection},
  author = {Zhang, Shifeng and Wen, Longyin and Bian, Xiao and Lei, Zhen and Li, Stan Z.},
  booktitle = {arxiv preprint arXiv:1711.06897},
  year = {2017}
}

Contents

1. Installation
2. Preparation
3. Training
4. Evaluation
5. Models

Installation

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

git clone https://github.com/sfzhang15/RefineDet.git

2. Build the code. Please follow Caffe instruction to install all necessary packages and build it.

cd $RefineDet_ROOT
# Modify Makefile.config according to your Caffe installation.
# Make sure to include $RefineDet_ROOT/python to your PYTHONPATH.
cp Makefile.config.example Makefile.config
make all -j && make py

Preparation

1. Download fully convolutional reduced (atrous) VGGNet. By default, we assume the model is stored in $RefineDet_ROOT/models/VGGNet/.

2. Download ResNet-101. By default, we assume the model is stored in $RefineDet_ROOT/models/ResNet/.

3. Follow the data/VOC0712/README.md to download VOC2007 and VOC2012 dataset and create the LMDB file for the VOC2007 training and testing.

4. Follow the data/VOC0712Plus/README.md to download VOC2007 and VOC2012 dataset and create the LMDB file for the VOC2012 training and testing.

5. Follow the data/coco/README.md to download MS COCO dataset and create the LMDB file for the COCO training and testing.

Training

1. Train your model on PASCAL VOC.

# It will create model definition files and save snapshot models in:
#   - $RefineDet_ROOT/models/VGGNet/VOC0712{Plus}/refinedet_vgg16_{size}x{size}/
# and job file, log file, and the python script in:
#   - $RefineDet_ROOT/jobs/VGGNet/VOC0712{Plus}/refinedet_vgg16_{size}x{size}/
python examples/refinedet/VGG16_VOC2007_320.py
python examples/refinedet/VGG16_VOC2007_512.py
python examples/refinedet/VGG16_VOC2012_320.py
python examples/refinedet/VGG16_VOC2012_512.py

2. Train your model on COCO.

# It will create model definition files and save snapshot models in:
#   - $RefineDet_ROOT/models/{Network}/coco/refinedet_{network}_{size}x{size}/
# and job file, log file, and the python script in:
#   - $RefineDet_ROOT/jobs/{Network}/coco/refinedet_{network}_{size}x{size}/
python examples/refinedet/VGG16_COCO_320.py
python examples/refinedet/VGG16_COCO_512.py
python examples/refinedet/ResNet101_COCO_320.py
python examples/refinedet/ResNet101_COCO_512.py

3. Train your model form COOC to VOC (Based on VGG16).

# It will extract a VOC model from a pretrained COCO model.
ipython notebook convert_model_320.ipynb
ipython notebook convert_model_512.ipynb
# It will create model definition files and save snapshot models in:
#   - $RefineDet_ROOT/models/VGGNet/VOC0712{Plus}/refinedet_vgg16_{size}x{size}_ft/
# and job file, log file, and the python script in:
#   - $RefineDet_ROOT/jobs/VGGNet/VOC0712{Plus}/refinedet_vgg16_{size}x{size}_ft/
python examples/refinedet/finetune_VGG16_VOC2007_320.py
python examples/refinedet/finetune_VGG16_VOC2007_512.py
python examples/refinedet/finetune_VGG16_VOC2012_320.py
python examples/refinedet/finetune_VGG16_VOC2012_512.py

Evaluation

1. Build the Cython modules.

cd $RefineDet_ROOT/test/lib
make -j

2. Change the ‘self._devkit_path’ in test/lib/datasets/pascal_voc.py to yours.

3. Change the ‘self._data_path’ in test/lib/datasets/coco.py to yours.

4. Check out test/refinedet_demo.py on how to detect objects using the RefineDet model and how to plot detection results.

python test/refinedet_demo.py

5. Evaluate the trained models via test/refinedet_test.py.

# You can modify the parameters in refinedet_test.py for different types of evaluation:
#  - single_scale: True is single scale testing, False is multi_scale_testing.
#  - test_set: 'voc_2007_test', 'voc_2012_test', 'coco_2014_minival', 'coco_2015_test-dev'.
#  - voc_path: where the trained voc caffemodel.
#  - coco_path: where the trained voc caffemodel.
# For 'voc_2007_test' and 'coco_2014_minival', it will directly output the mAP results.
# For 'voc_2012_test' and 'coco_2015_test-dev', it will save the detections and you should submitted it to the evaluation server to get the mAP results.
python test/refinedet_test.py

Models

We have provided the models that are trained from different datasets. To help reproduce the results in Table 1, Table 2, Table 4, most models contain a pretrained .caffemodel file, many .prototxt files, and python scripts.

1. PASCAL VOC models (VGG-16):

   • 07+12: RefineDet320, RefineDet512
   • 07++12: RefineDet320, RefineDet512
   • COCO: RefineDet320, RefineDet512
   • 07+12+COCO: RefineDet320, RefineDet512
   • 07++12+COCO: RefineDet320, RefineDet512

2. COCO models:

   • trainval35k (VGG-16): RefineDet320, RefineDet512
   • trainval35k (ResNet101): RefineDet320, RefineDet512

Note: If you can not download our pre-trained models through the above links, you can download them through BaiduYun.