Skip to content

Latest commit

 

History

History
55 lines (33 loc) · 3.05 KB

faster-rcnn.md

File metadata and controls

55 lines (33 loc) · 3.05 KB

Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (2016), Shaoqing Ren et al.

contributors: @GitYCC

[paper] [code]

Prerequisite

Fast R-CNN (2015), Ross Girshick.

Introduction

  • When ignoring the time spent on region proposals (e.g. Selective Search), Fast R-CNN achieves near real-time rates using very deep networks.
  • Region Proposal Networks (RPNs) are designed to efficiently predict region proposals with a wide range of scales and aspect ratios.
  • To unify RPNs with Fast R-CNN object detection networks, we propose a training scheme that alternates between fine-tuning for the region proposal task and then fine-tuning for object detection, while keeping the proposals fixed.
  • Briefly, Faster R-CNN is replacing "Selective Search" with "RPNs" on Fast R-CNN.

Faster R-CNN

Conv Layers

  • Zeiler and Fergus model (ZF)
    • last conv layer of ZF is Conv5, whose size is (6 x 6 x 256)
  • VGG-16
    • last conv layer of VGG16 is Conv13, whose size is (14 x 14 x 512)

Region Proposal Networks (RPNs)

  • To generate region proposals, we slide a small network over the convolutional feature map output by the last shared convolutional layer. This small network takes as input an $n × n$ ($n=3$) spatial window of the input convolutional feature map. Each sliding window is mapped to a lower-dimensional feature (256-d for ZF and 512-d for VGG, with ReLU following).

    • ZF: conv3-256 (padding size=1), generate 6 x 6 x 256 conv feature map
    • VGG: conv3-512 (padding size=1), generate 14 x 14 x 512 conv feature map

  • This feature is fed into two sibling fully-connected layers—a box-regression layer (reg) and a box-classification layer (cls). This architecture is naturally implemented with an $n×n$ convolutional layer followed by two sibling $1 × 1$ convolutional layers (for reg and cls, respectively).

    • Anchors: At each sliding-window location, we simultaneously predict multiple region proposals, where the number of maximum possible proposals for each location is denoted as $k$. So the reg layer has $4k$ outputs encoding the coordinates of $k$ boxes, and the cls layer outputs $2k$ scores that estimate probability of object or not object for each proposal.
      • For anchors, we use 3 scales with box areas of 128, 256, and 512 pixels, and 3 aspect ratios of 1:1, 1:2, and 2:1 ($k=9$)
    • cls layer: conv1-18 (no padding) -> softmax on pixel
    • reg layer: conv1-36 (no padding) -> regression on pixel

  • Some RPN proposals highly overlap with each other. To reduce redundancy, we adopt non-maximum suppression (NMS) on the proposal regions based on their cls scores. We fix the IoU threshold for NMS at 0.7, which leaves us about 2000 proposal regions per image.