ConvNets
文章平均质量分 96
bea_tree
只要不断按着梯度来,最差也可以进入局部最优解
展开
-
Improving Semantic Segmentation via Video Propagation and Label Relaxation
一篇使用视频信息提升semantic segmentation 精度的工作,可以看成合理的进行data augmentation方法,文章试验做的很全面,总体来说非常扎实。文章继承了英伟达该组之前的sdc net (见本文附录)的工作。Methodology使用SDC-net 预测某片段前后k 帧图像motion vectors从而得到相应的image和label,增加了网络的训练数据。...原创 2019-06-03 22:32:53 · 2724 阅读 · 4 评论 -
论文笔记 | Metric Learning with adaptive density discrimination
Authors Oren Rippel(有趣的是作者喜欢中文) Manohar Paluri Piotr Dollar Lubomir BourdevAbstract本文介绍了一种Distance Metric Learning (DML),效果比triplet还要好,而且需要的迭代次数更少。1 Introduction 与传统的DML相比,Magenet不仅考虑intra-class v原创 2016-10-09 14:34:59 · 3695 阅读 · 0 评论 -
论文笔记 |What makes for effective detection proposals?
在faster rcnn中提到的proposal的综述:J. Hosang, R. Benenson, and B. Schiele, “How good are detection proposals, really?” in British Machine Vision Conference(BMVC), 2014.J. Hosang, R. Benenson, P. Doll´ar, a原创 2016-07-08 12:48:33 · 2596 阅读 · 2 评论 -
论文笔记 | FaceNet: A Unified Embedding for Face Recognition and Clustering
AuthorsFlorian Schroff Dmitry Kalenichenko James Philbin Florian SchroffAbstract本文提出了FaceNet system,直接从face images 学习到 compact Euclidean 欧几里德 space 从而得到face的相似程度。这样一来 face recognition, v原创 2016-07-28 21:29:48 · 5107 阅读 · 0 评论 -
论文笔记 | Training Region-based Object Detectors with Online Hard Example Mining
AuthorsAbhinav Shrivastava Abhinav Gupta Ross Girshick Abhinav Shrivastava 这篇文章很多人说不值得在cvpr oral,不过有ross Girshick,还是要看一看的。Abstract本文提出了online hard example mining(OHEM),使用的是Bootstrapping技术,又一次说明了原创 2016-07-07 10:56:02 · 2809 阅读 · 0 评论 -
kaggle | Digit recognizer with caffe
Maybe this competition is easy to complete and it might be a good way to practice my Caffe skill. https://www.kaggle.com/c/digit-recognizer 1 Csv -> lmdbhttp://www.cnblogs.com/dcsds1/p/5205669.原创 2016-06-19 20:29:16 · 2216 阅读 · 2 评论 -
论文笔记|Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition
AuthorsKaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun 何恺明 #原创 2016-06-19 20:24:58 · 1622 阅读 · 0 评论 -
论文笔记|You Only Look Once: Unified, Real-Time Object Detection
AuthorJoseph Redmon, Santosh Divvalay, Ross Girshick{, Ali Farhadiy Joseph Redmon #原创 2016-06-19 20:27:30 · 2327 阅读 · 0 评论 -
论文笔记 | CNN-RNN:A Unified Framework for Multi-label Image Classification
AuthorsJiang Wang Yi Yang Junhua Mao Zhiheng Huang Chang Huang Wei Xu Wang Jiang Abstract利用了CNN和RNN,考虑了类别之间的dependency,取得了不错的分类效果1 Introductionmulti_label 的一些文献Y. Gong, Y. Jia, T. Leung, A. Toshe原创 2016-09-28 19:17:01 · 7286 阅读 · 6 评论 -
论文笔记| 几分钟看完ResNet的融合特性及冗余性分析的三篇文章
本文是博主在paper reading时的ppt,主要涉及三篇论文: 1. Residual Networks Behave Like Ensembles of Relatively Shallow Networks(nips) 2. On the Connection of Deep Fusion to Ensembling(微软亚研) 3. Wider or Deeper: Revisi原创 2017-02-15 15:39:29 · 6843 阅读 · 9 评论 -
神经网络梯度下降优化算法及初始化方法小结
An overview of gradient descent optimization algorithms and Weight initialization methods. 神经网络重要的一点就是调参炼丹,这里复习一下网络的初始化方法及优化方法。 然而知道这些并没有什么用, 平时多实验才是王道网络优化方法1 SGD2 Momentum3 Nesterov4 Adag原创 2017-10-09 18:53:53 · 5993 阅读 · 0 评论 -
论文水记|How to Train Triplet Networks with 100K Identities?
这是来自猎户星空的关于人脸识别的文章 作者 Chong Wang ;Xue Zhang ;Xipeng Lan https://arxiv.org/abs/1709.02940好久没有写博客了,水一篇。。。一句话总结对应triplet的训练,多采用OHNM的方式挖掘困难负样本,然而随着训练数据的增加,easy-triplet更多,困难样本的搜索空间增大,于是本文将训练数据分为若干小部分原创 2017-09-26 20:28:51 · 3981 阅读 · 0 评论 -
caffe 实例笔记 2 LeNet详细解读及实现
1 温习1.1 关于caffe的名称:caffe = convolutional architecture for fast feature embedding 1.2 caffe.protoProtocol Buffers顾名思义这是一种协议接口,这是了解caffe功能之后,需要了解的第一件事情。有很多相关博客。简单看一下其结构:package xx;#xx将作为原创 2016-06-19 20:21:32 · 10899 阅读 · 7 评论 -
论文笔记 | R-FCN: Object Detection via Region-based Fully Convolutional Networks
AuthorsJifeng Dai,Yi Li,Kaiming He,Jian Sun 代季峰 代码里还有百度云盘的连接,为国人考虑的真周到,只是自己的显卡不足~Abstract本文提供了region-based,fully convolutional networks,用于快速精确的目标检测。Fast或者Faster在per-region的时候都需要subnetwor原创 2016-07-04 06:15:12 · 14592 阅读 · 15 评论 -
Caffe 实例笔记 1 CaffeNet从训练到分类及可视化参数特征 微调
本文主要分四部分 1. 在命令行进行训练 2. 使用pycaffe进行分类及特征可视化 3. 进行微调,将caffenet使用在图片风格的预测上1 使用caffeNet训练自己的数据集主要参考: 官方网址: http://caffe.berkeleyvision.org/gathered/examples/imagenet.html 数据集及第一部分参考网址:http://www.lx原创 2016-06-06 19:41:15 · 19682 阅读 · 2 评论 -
几分钟走进神奇的光流|FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
故事背景 那是15年的春天,本文的作者和其他几个人,看着美丽的春光,突发奇想使用CNN做光流估计,于是FlowNet成了第一个用CNN做光流的模型,当时的结果还不足以和传统结果相匹配。2016年冬天,作者和一群小伙伴又基于Flow Net的工作进行了改进,效果得到了提升,可以与传统方法相匹敌。 15年的思想主要是把两张用来估计光流信息的图片输入网络,经过训练使网络学到光流信息,后面会原创 2017-03-28 02:56:38 · 32625 阅读 · 27 评论 -
CNN入门必读经典:Visualizing and Understanding Convolutional Networks
本文主要是借助deconvnet来可视化卷积网络,这对于理解卷积网络还是非常重要的,同时本文又是13年ImageNet分类任务的冠军。 代码: https://github.com/guruucsd/CNN_visualization1 Deconvolution首先我们先对Deconvolution有个了解,这里推荐知乎里面的一个回答: http://zhihu.com/question/4原创 2017-04-03 12:55:38 · 12012 阅读 · 3 评论 -
论文笔记 | BoxSup: Exploiting Bounding Boxes to Supervise Convolutional Networks for Semantic Segmentati
AuthorsJifeng Dai Kaiming He Jian Sun 都是老熟人了就不贴照片了~Abstract本文使用bbox 来代替或者部分代替mask进行图像像素级分割,节省了标注时间,充分利用了bbox的数据集。基本思路是automativally generating region proposals and training convolutional networks 相互交替原创 2016-08-08 23:20:11 · 4227 阅读 · 5 评论 -
论文笔记| Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
AuthorShaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun Shaoqing RenAbstractIn this work , it introduces a Region Proposal Network(RPN) that shares full-image convolutional features wi原创 2016-06-19 20:26:41 · 2711 阅读 · 0 评论 -
论文笔记|Fast R-CNN
AuthorRoss GirshickAbstractCompared to SPP net, Fast R-CNN trains VGG16 x3 faster, tests 10x faster, and is more accurate.1 IntroductionLocalization challenges: 1. numerous candidate ob原创 2016-06-19 20:25:46 · 2867 阅读 · 0 评论 -
论文笔记|Rich feature hierarchies for accurate object detection and semantic segmentation
AuthorsRoss Girshick /Jeff Donahue/Trevor Darrell /Jitendra Malik Ross Girshickhttp://blog.csdn.net/chenriwei2/article/details/38110387 http://blog.csdn.net/u013488563/article/details/420278原创 2016-06-19 20:23:50 · 4446 阅读 · 2 评论 -
Object detection
Review and summrize原创 2016-06-22 10:28:42 · 775 阅读 · 0 评论 -
论文笔记 | SSD: Single Shot MultiBox Detector
AuthorsWei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, Alexander C. Berg Wei LiuAbstractOur approach, named SSD, discretizes the output space of bounding box原创 2016-06-21 19:35:47 · 5155 阅读 · 0 评论 -
论文笔记 | Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning
1 IntrouductionIn this work we study the combination of the two most recent ideas: Residual connections and Inception v3. We replace the filter concatenation stage of the Inception architecture with re原创 2016-06-29 17:19:52 · 10708 阅读 · 1 评论 -
论文笔记 | Going deeper with convolutions
AuthorsChristian Szegedy Wei Liu Yangqing Jia Pierre Sermanet Scott Reed Dragomir Anguelov Dumitru Erhan Vincent Vanhoucke Andrew Rabinovich Christian Szegedy3 Motivation and high level co原创 2016-06-27 16:12:10 · 2273 阅读 · 0 评论 -
论文笔记 | VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE -SCALE IMAGE RECOGNITION
AuthorsKaren Simonyan & Andrew Zisserman Karen SimonyanAbstractIn this work we investigate the effect of the convlutional network depth on its accuracy in the large-scale image recognation.1 Intro原创 2016-07-03 01:18:54 · 1236 阅读 · 0 评论 -
论文笔记 | Exploit All the Layers: Fast and Accurate CNN Object Detector with SDP and CRC
AuthorsFan Yang Wongun Choi Yuanqing Lin fan Yang Abstract本文提出了两种目标检测的措施,兼具精度与效率:1.scale-dependent pooling (精度)2. layer wise casaded rejection classifiers(效率)1 Introduction首先作者简要介绍了RCNN等方法,FRCNN的原创 2016-07-12 19:38:30 · 3760 阅读 · 0 评论 -
论文笔记 | Improving neural networks by preventing co-adaptation of feature detectors
AuthorsG. E. Hinton , N. Srivastava, A. Krizhevsky, I. Sutskever and R. R. Salakhutdinov HintonAbstract训练时随机忽略一半的feature detectors 能够防止因训练集太小带来的过拟合问题。这能够防止一些detectors联合在一起才起作用的情况,每个神经元预测一个特征有利于提高准原创 2016-07-10 15:54:04 · 4619 阅读 · 0 评论 -
论文笔记 | Deep Residual Learning for Image Recognition
AuthorsKaiming He Xiangyu Zhang Shaoqing Ren Jian SunAbstractResidual Networks are easier to optimize and gain accuracy from considerably increased depth, but it have lower complexity than VGGnets.1 In原创 2016-06-23 01:22:52 · 3539 阅读 · 0 评论 -
论文笔记 | OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks
AuthorPierre Sermanet David Eigen Xiang Zhang Michael Mathieu Rob Fergus Yann LeCun Pierre SermanetAbstractWe show how a multi-scale and sliding window approach can be implemented in a ConvNet. We a原创 2016-06-24 19:49:58 · 1322 阅读 · 0 评论 -
caffe 实例笔记 4 Multilabel classification on PASCAL using python data-layers
This example is based on PASCAL VOC 2012, I think the matters of the blog are as follows: 1. Python data layer 2. SigmoidCrossEntropyLoss1 Preliminariesmake sure you compile caffe using WITH_原创 2016-06-19 20:22:50 · 3201 阅读 · 0 评论 -
Caffe 实例笔记 3 Brewing Logistic Regression then Going Deeper
From now on, I will try to write some blogs in English to improve my English writing skills. If there is anything wrong in the blogs ,please let me know. Thanks.1 Feature extraction with Caffe C++原创 2016-06-19 20:22:06 · 1189 阅读 · 0 评论 -
论文笔记 | Learning Deep Features for Discriminative Localization
作者Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, Antonio Torralba Bolei Zhou Abstract受到NIN 的启发,将global average pooling 用于 定位1. IntroductionGlobal average pooling layer 不仅是一个regularizer, 经过原创 2016-07-04 23:40:48 · 6050 阅读 · 8 评论 -
论文笔记 | HyperNet: Towards Accurate Region Proposal Generation and Joint Object Detection
作者Tao Kong Anbang Yao Yurong Chen Fuchun Sun 摘要本文提出了一种HyperNet,同时用于处理region proposal和object detection。Hyper特征是将各层featuremaps的特征聚集然后统一到一个空间。在cpu上处理速度是5fps。1 Introduction首先说了Rcnn 到Faster原创 2016-07-04 06:14:11 · 4825 阅读 · 3 评论 -
论文笔记 | Identity Mappings in Deep Residual Networks
这一篇文章为上一篇ResNets提供了更加可信服的论据AuthorsKaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun 都是“老熟人”了就不介绍了Abstract上篇文章中(http://blog.csdn.net/bea_tree/article/details/51735788)因residual builing b原创 2016-07-04 06:13:30 · 4445 阅读 · 0 评论 -
论文笔记 | Network In Network
AuthorsMin Lin, Qiang Chen, Shuicheng Yan AbstractInstead of linear filters, we bulid micro neural networks( micro neural network) with more complex strutures to abstract the data within the原创 2016-07-04 06:12:57 · 1391 阅读 · 0 评论 -
论文笔记 | Rethinking the Inception Architecture for Computer Vision
醉醉的,写完之后竟然被新的文章代替了,只能重新写一遍了~AuthorsChristian Szegedy Vincent Vanhoucke Sergey Ioffe Jonathon ShlensAbstractHere we are exploring ways to scale up networks in ways that aim at utilizing the a原创 2016-06-28 20:26:16 · 2124 阅读 · 0 评论 -
caffe|Fine-tuning for driver
1 img–>lmdbfirstly we need the train.txt,and test.txtimport ospath='/path/path/'#we can also use os.walk(dir)def imgname_to_txt(dir,output_txt): output=open(output_txt,mode='w')#mode:write read原创 2016-06-27 16:09:37 · 771 阅读 · 0 评论 -
论文笔记 | Wide Residual Networks
AuthorsSergey Zagoruyko Nikos Komodakis Sergey ZagoruykoAbstract网络不断向更深发展,但是有时候为了得到少量的accuracy的增加,却需要将网络层数翻倍,也会减少feature的reuse,降低训练速度。作者提出了wide residual network,16层的表现就比之前的ResNet效果要好。1 Introductio原创 2016-07-10 02:08:49 · 14600 阅读 · 2 评论