/Rainbow-SSD

Enhancement of SSD by concatenating feature maps for object detection

By Jisoo Jeong, Hyojin Park, Nojun Kwak

Intoroduction

SSD Images vs R-SSD Images

The conventional SSD has a couple of points to be supplemented

  • Each layer in the feature pyramid is used independently (the same object can be detected in multiple scales)
  • Small objects are not detected well (this is not the problem only for SSD)

We tackle this problems as follows

  • The classifier network is implemented considering the relationship between layers in the feature pyramid
  • The number of channels in a layer is increased efficiently
  • The proposed network is suitable for sharing weights in the classifer network for different scales, resulting in a single classifier network

For more details, please refer to our arXiv paper

Citing R-SSD

Please cite R-SSD in your publications if it helps your research

@article{jeong2017enhancement,
  title={Enhancement of SSD by concatenating feature maps for object detection},
  author={Jeong, Jisoo and Park, Hyojin and Kwak, Nojun},
  journal={arXiv preprint arXiv:1705.09587},
  year={2017}
}

Installation & Preparation

We experimented with R-SSD using the SSD framework. To use our model, complete the installation & preparation on the SSD homepage

Models

Pascal VOC model

Models training batch size mAP
R-SSD300 32 78.7 (higher than paper)
R-SSD521 4 80.8
R-SSD300 with one classifier(6 boxes) 8 77.0

To test this model, check "sh" file

# check your path in shell script (.sh file)
# cd /home/soo/caffe_ssd -> cd /your/path
./R_SSD_300model_32.sh  (in file folder)