深度学习论文阅读
CH-Yuan
这个作者很懒,什么都没留下…
展开
-
generalized mean pooling
GMP原创 2022-06-17 15:39:56 · 575 阅读 · 0 评论 -
BV-Person: A Large-scale Dataset for Bird-view Person Re-identification
论文代码:目前没有提供ICCV2001作者考虑到目前在ATM、超市收银台等地方存在较高视角的摄像头,拍摄的图像特征中头、肩是主要的信息,下身被自身遮挡。提出了一个BirdView的数据集 https://git.io/BVPerson,同时提供一个基于cross-attention 的解决模型。三个挑战:(1)如何将bird-view和horizental-view的图像建立联系;——cross attention(2)头肩在bird-view 图像中占很大部分,而horizontal-view图原创 2021-11-05 10:38:38 · 387 阅读 · 0 评论 -
论文阅读:cvpr2021 :Joint Generative and Contrastive Learning for Unsupervised Person Re-identification
将GAN、contrastive module 联合学习GAN 利用身份特征和mesh-based映射视角生成新的图像(图2b)contrastive module从feature memory 中提取正负对,便于视角不变的特征学习(图2c)原创 2021-07-28 10:23:51 · 586 阅读 · 0 评论 -
论文阅读:Joint Discriminative and Generative Learning for Person Re-identification
pdf: Joint Discriminative and Generative Learning for Person Re-identificationgithub:https://github.com/layumi/Person_reID_baseline_pytorch https://github.com/NVlabs/DG-Net原创 2021-07-28 10:10:35 · 140 阅读 · 0 评论 -
论文阅读: Omni-Scale Feature Learning for Person Re-Identification
pdf:code:https://github.com/KaiyangZhou/deep-person-reidAbstract:As an instance-level recognition problem, person re-identification (re-ID) relies on discriminative features, which not only capture different spatial scales but also encapsulate an arbitr原创 2020-11-17 18:38:37 · 348 阅读 · 1 评论 -
谷歌系列 :Inception v1到v4
Paper列表:[v1] Going Deeper with Convolutions, 6.67% test error, http://arxiv.org/abs/1409.4842 [v2] Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift, 4.8%原创 2016-11-10 15:13:27 · 27544 阅读 · 1 评论 -
1:ImageNet Classification with Deep Convolutional Neural Networks
这里写链接内容ImageNet Classification with Deep Convolutional Neural Networks 在看caffe教程中example/ImageNet 时,看这篇文章,了解该网络结构,各个layer在网络中的设置和作用。一、ImageNet介绍ImageNet由15 million 有标签的高分辨率图像组成,有大约22,000个类别。 开始于2010年原创 2015-12-03 16:20:53 · 3155 阅读 · 0 评论 -
论文阅读:DENSELY CONNECTED CONVOLUTIONAL NETWORKS
概述作者观察到目前的深度网络有一个重要的特点是shorted connected,认为能够训练极深的网络很可能就是由于这个原因。 于是作者提出了Densely connected convolutional network, 该网络和ResNet类似,采用了短路链接。但是不同的是,在接收前层的feature maps时,ResNet采用了summation,而DenseNet采用的则是concate原创 2017-03-01 10:00:12 · 3003 阅读 · 5 评论 -
论文阅读:Multi-Scale Triplet CNN for Person Re-Identification
中国科技大学的Jiawei Liu等发表在Acm on Multimedia Conference上的论文。作者提出一个多尺度的triplet CNN,在一个数据集Market1501上表现出好的性能。网络结构如下: Triplet loss 主要改进:同时训练了三个不同尺度的网络,对于低尺度用了shallow network(2个卷积+池化),不变尺度的用了5个卷积(池化)。最后进行特征的组合原创 2017-03-02 10:55:46 · 2473 阅读 · 1 评论 -
论文阅读:Joint Learning of Single-image and Cross-image Representations for Person Re-identification
哈尔滨工业大学(/香港理工大学)的作者Faqiang Wang等采用pairwise comparison and triplet comparison联合学习特征,并在CUHK03,CUHK03,VIPeR上进行实验。网络结构: 文章分析了SIR(single image representation) 是CIR(cross image representation)的特殊形式。(欧式距离,马氏原创 2017-03-02 20:33:41 · 1144 阅读 · 0 评论 -
SPP pooling layer
参考:Spatial Pyramid Pooling in Deep Convolutional Networks for Visual RecognitionSPP_net的提出因为传统的CNNs网络只能接受固定size的images。实际上,卷积层可以输入任意尺度的, 但这样就会导致输出的feature map 大小不同;而InnerProduct layer 的输入必须是固定的。因此在CNN原创 2016-10-18 21:02:31 · 3315 阅读 · 0 评论 -
Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification
序:在ISVRC2015 on ImageNet2012 classification dataset 上,取得了4.94%的 top-5 test error,这是第一次超越了人的识别了率(5.1%). 文章提出了两个内容: (1) Parametric Rectified Linear Unit (PReLU) :使模型以接近0的额外计算代价和较小的过拟合风险训练模型。 (2) MSRA:原创 2016-10-17 17:33:07 · 2457 阅读 · 0 评论 -
[CVPR2016]Learning Deep Feature Representations with Domain Guided Dropout for Person Re-id
论文题目:Learning Deep Feature Representations with Domain Guided Dropout for Person Re-identification 还是香港中文大学的牛牛们写的paper.有源码:https://github.com/Cysu/person_reidMotivation这篇paper主要是为了解决一些issues(例如person r原创 2016-11-05 11:22:23 · 914 阅读 · 0 评论 -
Deep Metric Learning for Person Re-Identification
Deep Metric Learning for Person Re-Identification(ICPR2014,中国科学院自动化研究所模式识别国家重点实验室:Dong Yi, Zhen Lei, Shengcai Liao and Stan Z. Li)网络结构 实验结果 总结创新点: 1.把siamese最后的对照损失函数替换成了cosin 层 2.两个网络对共享相同权重——跨数原创 2016-03-30 11:18:23 · 4693 阅读 · 10 评论 -
DeepReID (2014 CVPR)
DeepReID: Deep Filter Pairing Neural Network for Person Re-Identification(香港中文大学Wei Li Rui Zhao Tong Xiao XiaogangWang∗)三个贡献1.提出了一个filter paring neural network(FPNN) 2.采用了dropout,data augmentation,dat原创 2016-03-30 09:27:32 · 4011 阅读 · 6 评论 -
泛读:CVPR2014:Discriminative Deep Metric Learning for Face Verification in theWild
Discriminative Deep Metric Learning for Face Verification in theWild(Junlin Hu, Jiwen Lu, Yap-Peng Tan School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore)文章用了原创 2016-03-31 11:08:49 · 1621 阅读 · 0 评论 -
Deep Residual Learning for Image Recognition 笔记
下面转自博客:http://blog.csdn.net/u014114990/article/details/50505331今天跑了把googlenet1应用到自己的分类任务中,识别率有大约有0.8%的提高。 看来大网络还是有用的, 只要你能很好的解决随着网络层数增多,容易引起过梯度消失或爆炸等问题。https://github.com/ShaoqingRen/faster_rcnn/issues转载 2016-03-08 10:00:08 · 940 阅读 · 0 评论 -
VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE-SCALE IMAGE RECOGNITION
这篇文章发表在ICLR2015 (International Conference on Learning Representations),这个会议从2013年开始举办,上升快大牛多,关于Deep Learning的。这篇文章解决ConvNet 结构设计中“深度”这个重要问题。提出用3*3的卷积增加网络的深度,当weighted layers增加到16-19层时,accuracy得到明显提升。网络原创 2016-03-22 13:03:49 · 2526 阅读 · 2 评论 -
OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks
可以参考博文:[http://blog.sciencenet.cn/blog-1583812-844178.html](http://blog.sciencenet.cn/blog-1583812-844178.html) (大概了解一下,没仔细看) 文章发表在ICLR2014,yan Lecun实验室。提出了一个用卷积网络进行分类、定位和检测的联合框架。介绍了如何在一个ConvNe原创 2016-03-23 16:24:43 · 919 阅读 · 1 评论 -
caffe:Data 层
在看caffe教程中Fine-tuning for style recognition时比较image_net和style微调网络结构时,发现两个不同的Data layer的类型Data 和 ImageData,如下:**name: "CaffeNet"**layer { name: "data" type: "Data" top: "data" top: "label" in原创 2016-03-01 10:04:12 · 2882 阅读 · 1 评论 -
GoogLeNet
GoogLeNet v1参考:http://blog.csdn.net/sunbaigui/article/details/50807362 论文网络结构定义中,accuracy-1,accuracy-5,辅助loss层指定所占比例,定义:layer { name: "loss1/ave_pool" type: "Pooling" bottom: "inception_4a/outp原创 2016-05-16 17:17:32 · 1505 阅读 · 0 评论 -
contrastive loss function (papers)
1.Learning a Similarity Metric Discriminatively, with Application to Face Verification 2,Dimensionality Reduction by Learning an Invariant Mapping 3.Signature Verification using a “Siamese”Time Delay原创 2016-09-10 21:15:50 · 2564 阅读 · 0 评论 -
cvpr 2015 2016论文地址
cvpr 2016http://www.cv-foundation.org/openaccess/CVPR2016.py 关于person re-identification的文章: 1、Learning a Discriminative Null Space for Person Re-Identification 2、Learning Deep Feature Representation原创 2016-09-02 09:17:30 · 4913 阅读 · 0 评论 -
What is 'lr_policy' in caffe?
reference: https://stackoverflow.com/questions/30033096/what-is-lr-policy-in-caffe/30045244I just try to find out how I can use Caffe. To do so, I just took a look at the different .prototxt原创 2015-12-08 19:55:04 · 3199 阅读 · 0 评论