Recently, object proposal generation has shown value for various vision tasks, such as object detection, semantic instance segmentation, multi-label image classification, and weakly supervised learning, by hypothesizing object locations. We are motivated by the fact that many traditional proposal methods generate dense proposals to cover as many objects as possible but that i) they usually fail to rank these proposals properly and ii) the number of proposals is very large. For example, the well-known object proposal generation methods, Edge Boxes and Selective Search, can achieve high detection recall with thousands of proposals per image. But the large number of generated proposals makes subsequent analyses difficult due to the large number of false alarms and heavy computation load. o significantly reduce the number of proposals, we design a computationally ightweight neural network to refine the initial object proposals. The refinement consists of two parallel processes, re-ranking and box regression. The proposed network can share convolutional features with other high-level tasks by joint training, so the proposal refinement can be very fast. We show a joint training example of object detection in this paper. Extensive experiments demonstrate that our method can achieve state-of-the-art performance with a few proposals compared with some well-known proposal generation methods.

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

@article{liu2020refinedbox,
  title={{RefinedBox}: Refining for Fewer and High-quality Object Proposals},
  author={Liu, Yun and Li, Shi-Jie and Cheng, Ming-Ming},
  journal={Neurocomputing},
  volume={406},
  pages={106--116},
  year={2020},
  publisher={Elsevier}
}
@article{cheng2019bing,
  title={{BING}: Binarized Normed Gradients for Objectness Estimation at 300fps},
  author={Cheng, Ming-Ming and Liu, Yun and Lin, Wen-Yan and Zhang, Ziming and Rosin, Paul L and Torr, Philip HS},
  journal={Computational Visual Media},
  volume={5},
  number={1},
  pages={3--20},
  year={2019},
  month={Mar},
  publisher={Springer},
  doi={10.1007/s41095-018-0120-1},
  url={https://doi.org/10.1007/s41095-018-0120-1}
 }

1. Clone the RefinedBox repository

   git clone https://github.com/yun-liu/RefinedBox.git

2. Follow the well-known py-faster-rcnn repository to install this software, because this code is based on Faster R-CNN.

3. Download the precomputed edge boxes proposals for the VOC2007 dataset here and put the extracted files in the data folder.

4. Link the proposal data to refined_box/mat_data by running

   cd RefinedBox_ROOT_PATH/data
   mkdir refined_box
   cd refined_box
   ln -s YOUR_PROPOSAL_DATA_PATH mat_data

   Otherwise, you can also change the proposal data path by modifying __C.TRAIN.RERANK_PROP_PATH and __C.TEST.RERANK_PROP_PATH in lib/fast_rcnn/config.py.

5. Run the following command to execute the script experiments/scripts/refined_box_alt_opt.sh for training and testing RefinedBox on the VOC2007 dataset:

   cd RefinedBox_ROOT_PATH
   ./experiments/scripts/refined_box_alt_opt.sh YOUR_GPU_ID VGG16 voc_2007

6. Run RefinedBox with other proposal generation methods following the above way. Note that you should prepare your proposal data first in the format required by the py-faster-rcnn repository.

This code is based on Faster R-CNN. Thanks to the contributors of Faster R-CNN.

@inproceedings{ren2015faster,
  title={Faster {R-CNN}: Towards real-time object detection with region proposal networks},
  author={Ren, Shaoqing and He, Kaiming and Girshick, Ross and Sun, Jian},
  booktitle={Advances in neural information processing systems},
  pages={91--99},
  year={2015}
}