目标检测--RON: Reverse Connection with Objectness Prior Networks for Object Detection

RON: Reverse Connection with Objectness Prior Networks for Object Detection
CVPR2017
https://github.com/taokong/RON

本文可以看作是对 SSD 的改进， SSD 对不同尺度特征图进行独立的检测，这里我们 reverse connection block 将相邻的特征图联系起来。同时使用 objectness prior 来有效降低目标的搜索空间。

3 Network Architecture
我们使用 VGG16 作为基础模型， VGG16有13个卷积层，3个全连接层，这里我们将第 14,15 全连接层变为卷积层，使用一个2×2卷积核步长为2 来降采样 FC7 的尺寸。用于检测的特征图尺寸分布为输入图像尺寸的 1/8 (conv 4 3), 1/16 (conv 5 3), 1/32 (conv 6) and 1/64 (conv 7)

RON object detection overview

3.1. Reverse Connection
怎么把相邻的特征图联系起来了？我们提出了一个 reverse connection block 来解决这个问题。相邻的特征图尺寸差异通过 deconv 反卷积来消除

3.2. Reference Boxes
这里我们对每个尺度的特征图使用 2 scales and 5 aspect ratios 的 default boxes，
aspect ratios { 1/3，1/2, 1, 2,3}

3.3. Objectness Prior
针对正负样本比例严重失调， the ratio between object and non-object samples is seriously imbalanced，这里我们使用 Objectness Prior 来过滤大部分负样本

3.4. Detection and Bounding Box Regression

A=10个 Reference Boxes

inception block 我们使用了一个比较简单的结构

3.5. Combining Objectness Prior with Detection

经过 objectness prior 过滤，只有很少一部分样本进行 back-propagation

4 Training and Testing
4.1. Loss Function
损失函数包括三个子损失函数： objectness prior，矩形框坐标，物体类别

其中α =β =1/3

4.2. Joint Training and Testing
Data augmentation ：flipped，Randomly sampling a patch

Detection results on PASCAL VOC 2007 test set

Results on PASCAL VOC 2012 test set

MS COCO test-dev2015 detection results

6.1. Do Multiple Layers Help?

low resolution 384×384 input size，With 1.5G GPU memory at test phase, the speed of the network is 15 FPS, 3× faster than the Faster R-CNN counterpart

目标检测--RON: Reverse Connection with Objectness Prior Networks for Object Detection相关推荐

目标检测之RPN网络（Feature Pyramid Networks for Object Detection）
最近读了几篇WSDN的文章,有一篇的paper用到了RPN网络的一些思想,因此去拜读一下RPN这篇paper.这篇paper的attribution就是取长补短的思想,在卷积神经网络中,网络层数越浅, ...
【目标检测】36、OTA: Optimal Transport Assignment for Object Detection
文章目录一.背景二.方法 2.1 Optimal Transport 2.2 OT for label assignment 2.3 Center prior 2.4 Dynamic k Esti ...
目标检测--Enhancement of SSD by concatenating feature maps for object detection
Enhancement of SSD by concatenating feature maps for object detection BMVC 2017 本文是对SSD 的改进,通过牺牲一点速度 ...
两阶段3D目标检测网络 SIENet: Spatial Information Enhancement Network for 3D Object Detection from Point Cloud
本文介绍一篇两阶段的3D目标检测网络:SIENet. 这里重点是理解本文提出的 Hybrid-Paradigm Region Proposal Network 和 Spatial Informatio ...
目标检测论文阅读：Multi-scale Location-aware Kernel Representation for Object Detection（CVPR2018）
Multi-scale Location-aware Kernel Representation for Object Detection 论文链接:https://arxiv.org/abs/180 ...
CVPR2020-SEPC-单阶段目标检测提升4个点 | Scale-Equalizing Pyramid Convolution for Object Detection
启发应该是来源于SIFT中高斯金字塔带来的尺度不变性.论文出来很久了,感觉的确创新点很强,效果也是爆炸! 论文地址:http://openaccess.thecvf.com/content_CVPR_ ...
显著性目标检测之Global Context-Aware Progressive Aggregation Network for Salient Object Detection
Global Context-Aware Progressive Aggregation Network for Salient Object Detection 文章目录 Global Contex ...
【论文阅读】【三维目标检测】VoteNet：Deep Hough Voting for 3D Object Detection in Point Clouds
文章目录 Hough Voting VoteNet 网络结构 Voting in Point Clouds Object Proposal and Classification from Votes ...
【CVPR 2022】高分辨率小目标检测：Cascaded Sparse Query for Accelerating High-Resolution Smal Object Detection
QueryDet: Cascaded Sparse Query for Accelerating High-Resolution Small Object Detection 摘要: 动机: 实现: ...

目标检测--RON: Reverse Connection with Objectness Prior Networks for Object Detection

目标检测--RON: Reverse Connection with Objectness Prior Networks for Object Detection相关推荐

最新文章

热门文章