目标检测论文(尤其针对一些小目标的可能改进方法)

Posted ylhe

tags:

篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了目标检测论文(尤其针对一些小目标的可能改进方法)相关的知识,希望对你有一定的参考价值。

------------------------------------------------------------------------
About Face detection
------------------------------------------------------------------------
1、Finding Tiny Faces
    Code:https://github.com/peiyunh/tiny
    小目标检测难3大原因:目标本身尺度变化、图像分辨率以及环境因素。本文针对多尺度训练了不同的检测器,这些检测器所用特征来自同一网络的不同层级。此外,还充分利用了目标周边信息。

2、Seeing Small Faces from Robust Anchor’s Perspective
    基于anchor设计原理解决小脸检测不到的问题。

3、Face-MagNet: Magnifying Feature Maps to Detect Small Faces
    Code:https://github.com/po0ya/face-magnet
    基于Faster-RCNN框架提出Face-MagNet网络(在人脸建议和分类前放大特征图的判别能力)而无需任何跳过或残差连接。在RPN中和ROI前都加了一组反卷积层。另外,评估了其他3种针对尺度问题而有较好调整架构的方法:context pooling, skip connections, and scale partitioning.

4、Detecting and counting tiny faces
    Code:https://github.com/alexattia/ExtendedTinyFaces
    对Finding Tiny Faces这篇文章的深入理解,类似的方法。

5、SSH: Single Stage Headless Face Detector
    Code:https://github.com/mahyarnajibi/SSH
    单阶段检测器,速度快,占用内存少,在不同深度的网络层进行人脸检测,用于检测大、中、小人脸。

6、S3FD: Single Shot Scale-invariant Face Detector
    Code:https://github.com/sfzhang15/SFD
    (1) proposing a scale-equitable face detection framework to handle different scales of faces well.
    (2) improving the recall rate of small faces by a scale compensation anchor matching strategy.
    (3) reducing the false positive rate of small faces via a max-out background label.

7、Multi-Path Region-Based Convolutional Neural Network for Accurate Detection of Unconstrained “Hard Faces”
    a two-stage cascaded face detection framework:
    (1) a Multi-Path Region Proposal Network(MP-RPN),利用3个平行特征图的输出预测不同尺度的候选人脸区域,嵌有带有上采样过滤的卷积层和新提出的产生“难”例采样层。
    (2) a Boosted Forests classifier,利用候选人脸区域内的深层面部特征和周围更大区域的上下文特征,大大减少 hard negative samples.

8、Scale-Aware Face Detection
    先对图片上的人脸进行尺度估计,再在特定尺度上进行人脸检测(使用RPN,只使用一种anchor,且每次只检测一张脸)。不用在各个尺度下对人脸检测,因此速度较快。

9、Detecting Faces Using Inside Cascaded Contextual CNN
    不是使用多个CNN网络来级联的,而是使用一个CNN中不同网络层来做级联。CNN网络的前几层完成简单的人脸检测,后面的网络完成难度较大的人脸检测,采用data routing机制来使不同的卷积层由不同类型的样本来训练,关注于去掉不同类型的非人脸样本。 同时使用 body part localization 来辅助人脸检测。

10、Face Detection through Scale-Friendly Deep Convolutional Networks
    核心方法类似SSD。在网络不同阶段引出分支检测对应范围的人脸。训练时针对不同分组只用对应尺度的样本进行训练。

11、A Multi-Scale Cascade Fully Convolutional Network Face Detector
    基于FCNs的3层级联结构。It first proposes the approximate locations where the faces may be, then aims to find the accurate location by zooming on to the faces. Each level of the FCN cascade is a multi-scale fully-convolutional network, which generates scores at different locations and in different scales. A score map is generated after each FCN stage. Probable regions of face are selected and fed to the next stage. The number of proposals is decreased after each level, and the areas of regions are decreased to more precisely fit the face.

12、Face Detection using Deep Learning: An Improved Faster RCNN Approach
    对Faster RCNN的一些改进策略: feature concatenation, hard negative mining, multi-scale training, model pretraining, and proper calibration of key parameters. 

13、Face R-CNN
    对Faster RCNN改进:new multi-task loss function design, online hard example mining, and multi-scale training strategy

14、Face Detection Using Improved Faster RCNN
    multi-scale training, multi-scale testing, light-designed RCNN, keep the small proposals at training and testing stage, directly select top-ranked proposals (e.g., 6000) without NMS in the RPN stage for R-CNN, a vote-based NMS ensemble strategy.

15、Anchor Cascade for Efficient Face Detection
    propose a context pyramid maxout mechanism for anchor cascade。大大减少计算量和提高检测精度。同时对于训练小规模模型也有很高的检测精度。

16、SFace: An Efficient Network for Face Detection in Large Scale Variations
    解决大尺度变化问题。提出新算法SFace:整合了anchor-based methods(类似RetinaNet)和anchor-free based methods(类似UnitBox)。

-----------------------------------------------------------------------

-----------------------------------------------------------------------
1、Single-Shot Refinement Neural Network for Object Detection
    Code:https://github.com/sfzhang15/RefineDet
    可看做将Faster RCNN的two stages检测方法和SSD结合。
    propose a novel one-stage framework(RefineDet) consists of two inter-connected modules. the former aims to (1) filter out negative anchors to reduce search space for the classifier, and (2) coarsely adjust the locations and sizes of anchors to provide better initialization for the subsequent regressor. The latter module takes the refined anchors as the input from the former to further improve the regression and predict multi-class label. Meanwhile, we design a transfer connection block to transfer the features in the anchor refinement module to predict locations, sizes and class labels of objects in the object detection module. The multitask loss function enables us to train the whole network in an end-to-end way.

2、An Analysis of Scale Invariance in Object Detection-SNIP
    可看成改版版本的Image Pyramid。
    分析了小尺度与预训练模型尺度之间的关系, 提出了Scale Normalization for Image Pyramids (SNIP):在训练中,每次只回传那些大小在一个预先指定范围内的proposal的gradient,而忽略掉过大或者过小的proposal;在测试中,建立大小不同的Image Pyramid,在每张图上都运行这样一个detector,同样只保留那些大小在指定范围之内的输出结果,最终在一起NMS。这样就可以保证网络总是在同样scale的物体上训练,也就是标题中Scale Normalized的意思。

3、Cascade R-CNN: Delving into High Quality Object Detection 
    Code:https://github.com/zhaoweicai/cascade-rcnn
    基于two-stage detector。Cascade R-CNN是R-CNN的multi-stage延伸,由一系列随着IOU临界值增加而训练的检测器构成,从而对close false positives更具有选择性。R-CNN阶段的cascade是按顺序训练的,使用一个阶段的输出来训练下一阶段。类似于boostrapping methods,不同点是Cascade R-CNN的重采样过程并不旨在mine hard negatives,而是通过调整bounding boxes,每个阶段的目的都是为了找到一组好的false positive来训练下一阶段。

4、Single-Shot Object Detection with Enriched Semantics
    在SSD网络基础上,增加了语义分割分支和全局激活模块。前者增加低层检测特征,后者通过学习特征通道和目标类别的语义关系来进行高层目标检测特征。

5、Multi-scale Location-aware Kernel Representation for Object Detection
    Code:https://github.com/Hwang64/MLKP
    提出了一种新颖的多尺度位置感知核表示(MLKP),将判别性高阶统计量结合到object proposals的表示中以进行有效的对象检测。MLKP基于多项式核近似,可以有效生成低维高阶表示,其固有的位置保持性和敏感性也保证了可以灵活地用于目标检测。

6、A-Fast-RCNN: Hard Positive Generation via Adversary for Object Detection
    Code:https://github.com/xiaolonw/adversarial-frcnn
    提出学习一个可以生成遮挡和变形样本的对抗网络,对抗器的目标是生成让目标检测器难以进行分类的样本。在我们的框架中,原始检测器和对抗器都是以联合的方式学习的。

7、Detecting Small Signs from Large Images
    large images are broken into small patches as input to a Small Object-Sensitive-CNN (SOS-CNN) modified from a Single Shot Multibox Detector (SSD) framework with a VGG-16 network as the base network to produce patch-level object detection results. Scale invariance is achieved by applying the SOS-CNN on an image pyramid. Then, image-level object detection is obtained by projecting all the patch-level detection results to the image at the original scale.

8、Perceptual Generative Adversarial Networks for Small Object Detection
    P-GAN将小目标的特征映射到相似的大目标特征上来缩小差别,便能将小目标足够近似到大目标来欺骗判别器,达到小目标检测的目的。

9、Feature Pyramid Networks for Object Detection
    特征金字塔网络。

10、SOD-MTGAN: Small Object Detection via Multi-Task Generative Adversarial Network
    提出一个对于小目标检测的标准的端到端的多任务生成对抗网络(MTGAN),适用于任何已有的检测器。In the MTGAN, the generator network produces super-resolved images and the multi-task discriminator network is introduced to distinguish the real high-resolution images from fake ones, predict object categories, and refine bounding boxes, simultaneously. More importantly, the classification and regression losses are back-propagated to further guide the generator network to produce super-resolved images for easier classification and better localization.

11、Deep Feature Pyramid Reconfiguration for Object Detection
    当前特征金字塔的设计在如何整合不同尺度的语义信息方面仍然不够高效。本文把特征金字塔转换为特征的再组合过程,创造性地提出了一种高度非线性但是计算快速的结构将底层表示和高层语义特征进行整合。该网络由两个模块组成:全局注意力和局部再组合。这两个模块分布能全局和局部地去在不同的空间和尺度上提取任务相关的特征。重要的是,这两个模块具有轻量级、可嵌入和可端到端训练的优点。

12、Parallel Feature Pyramid Network for Object Detection
    使用SPP模块通过扩大网络宽度而不是增加深度来生成金字塔形特征图。提出MSCA模块有效地组合了不同规模的上下文信息。

13、SAN: Learning Relationship between Convolutional Features for Multi-Scale Object Detection
    提出了Scale Aware Network (SAN),将来自不同尺度的卷积特征映射到尺度不变的子空间,并设计了一种独特的学习方法,纯粹考虑了没有空间信息的渠道之间的关系。所提出的SAN减少了标度空间中的特征差异并提高了检测精度。

14、A CLOSER LOOK: SMALL OBJECT DETECTION IN FASTER R-CNN
    介绍了一种生成anchor proposals的改进建议,并对Faster R-CNN进行修改,利用较高分辨率的小目标的feature maps。

15、Improving Small Object Proposals for Company Logo Detection
    we introduce an improved scheme for generating anchor proposals and propose a modification to Faster R-CNN which leverages higher-resolution feature maps for small objects.

16、Scale-aware Pixel-wise Object Proposal Networks
    提出Scale-aware Pixel-wise Object Proposal(SPOP)网络,可以生成具有高召回率和平均最佳重叠(ABO)的proposals,即使对于小目标也是如此。另外,引入了一个类似分段的像素定位网络来密集预测每个像素的对象坐标,并开发了一种尺度感知对象定位策略,该策略将来自大尺寸和小尺寸网络的预测与加权机制相结合,以提高各种对象尺寸的坐标预测精度。

———————————————————————————————
原文链接:https://blog.csdn.net/u014236392/article/details/83993730

以上是关于目标检测论文(尤其针对一些小目标的可能改进方法)的主要内容,如果未能解决你的问题,请参考以下文章

论文解读SMOKE 单目相机 3D目标检测(CVPR2020)

目标识别YOLOv5针对小目标检测的改进模型/添加帧率检测

CVPR2020 SMOKE 单目相机 3D目标检测环境搭建篇

CVPR2020 SMOKE 单目相机 3D目标检测环境搭建篇

CVPR2020 SMOKE 单目相机 3D目标检测环境搭建篇

基于改进SSD的车辆小目标检测方法