目标检测 | Receptive Field Block Net for Accurate and Fast Object Detection

目标检测 | Receptive Field Block Net for Accurate and Fast Object Detection

  • Paper:https://eccv2018.org/openaccess/content_ECCV_2018/papers/Songtao_Liu_Receptive_Field_Block_ECCV_2018_paper.pdf
  • GitHub: https://github.com/ruinmessi/RFBNet

简介

  • RFB Net主要用于目标检测,模型参数少。文中提出了RF block模块, 接近人类视觉系统,检测框架使用ssd,具有精度高,速度快的特点,在pc上(gpu)可以达到实时。

目标检测 | Receptive Field Block Net for Accurate and Fast Object Detection

NET

  • RF BLOCK, 主要由两部分构成:1、多个分支的卷积层及不同卷积核,2、空洞卷积 ,这样可以形成不同的感受野和不同中心的特征。
    目标检测 | Receptive Field Block Net for Accurate and Fast Object Detection

  • 多个经典的 RF layer

目标检测 | Receptive Field Block Net for Accurate and Fast Object Detection

  • RFB-s, 借鉴Inception and Resnet, RFB-S通过两个3 ×\times 3卷积替换 5×55\times5

目标检测 | Receptive Field Block Net for Accurate and Fast Object Detection

  • RFB-Net300, 在不同的特征层采用的RFB模块不一定相同,conv4_3后面采用的是RFB-S, 最小两个特征由于尺寸太小,就没加 RFB。

目标检测 | Receptive Field Block Net for Accurate and Fast Object Detection

EXPERIMENT

  • Pascal voc 2007 test, epoch 250, lr strategy: first 5 epochs, use “warmup” 106to4×10310^{-6} to 4 \times 10^{-3}, divided by 10 at 150 and 200 epochs. Sgd, weight decay: 0.0005, momentum 0.9

目标检测 | Receptive Field Block Net for Accurate and Fast Object Detection

  • MS COCO

目标检测 | Receptive Field Block Net for Accurate and Fast Object Detection