Paper Note of Faster RCNN

Title: Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks

Contribution:

  • (motivation: break the bottleneck in 2 stage detection which is the generating procedure of proposals)
  • propose Region Prososal Networks (RPN) for generating proposals who shares conv layers with afterward detection networks
    • RPN: several conv layers - regress region bounds and objectness scores at each location on a grid
    • used nms on proposal regions based on cls scores
  • introduce anchor boxes as references at multi-scales/-aspect ratios
    • could be thought as pyramid of filters
    • key component for effectively sharing features
  • propose an alternating training scheme for the 2 stage detection
  • evaluated with different param/structure setting comprehensively (PASCAL VOC/ COCO)

Pros: end-to-end, provided code & detailed hyper-parameters for reimplement

Show comments from Gitment