[论文阅读]Faster R-CNN：Towards Real-Time Object Detection with Region Proposal Networks

Research questions

 

  

Motivation

Previous work

Approach

Training of Faster R-CNN(4-step training)

1. train RPN ( initialized with ImageNet-pre-trained model, and fine-tuned end-to-end for the region proposal task.  )    
2.  train a detector network ( initialized with ImageNet-pre-trained  model ) by Fast R-CNN using the proposals generated  by the                   Step-1 RPN  
3. use the detector to initialize RPN training, but fix the  shared convolutional layers and only fine-tune the layers unique to RPN  
4. keeping the shared convolutional layers fixed, we finetune  the unique layers of Fast R-CNN

Results

Summary

Contributions

1. Region Proposal Networks(RPNs)
2. By sharing convolutional features with the down-stream  detection network, the region proposal step is nearly cost-free.
3. Our method enables a unified, deep-learning-based object detection system to run at near real-time frame rates.(run at 5-17 fps).
4. The learned RPN also improves region proposal quality and thus the overall object detection accuracy.

Limitations

• doesn't achieve absolute real time yet
• the accuracy should be improved further

Code

Matlab version: https://github.com/ShaoqingRen/faster_rcnn

Python version: https://github.com/rbgirshick/py-faster-rcnn