Skip to content

Single-Shot Refinement Neural Network for Object Detection

License

Notifications You must be signed in to change notification settings

Baby47/RefineDet

 
 

Repository files navigation

Single-Shot Refinement Neural Network for Object Detection

License

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.

RefineDet Structure

System VOC2007 test mAP FPS (Titan X) Number of Boxes Input resolution
Faster R-CNN (VGG16) 73.2 7 ~6000 ~1000 x 600
YOLO (GoogleNe) 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

RefineDet results on multiple datasets

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 directory that you cloned Caffe into $RefineDet_ROOT.
git clone https://github.com/sfzhang15/RefineDet.git
  1. 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
  1. 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
  1. 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
  1. Change the ‘self._devkit_path’ in test/lib/datasets/pascal_voc.py to yours.

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

  3. 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
  1. 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):

  2. COCO models:

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

About

Single-Shot Refinement Neural Network for Object Detection

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C++ 78.0%
  • Python 12.3%
  • Cuda 5.7%
  • CMake 2.3%
  • MATLAB 0.7%
  • Makefile 0.6%
  • Other 0.4%