CV
文章平均质量分 51
John2King
machine learning & computer vision 小白
展开
-
[论文笔记3]Aggregated channel features
the aggregated pixels in all subsampled channels are then vectorized into a pixel look-up table.由于特征直接从下采样的通道里由像素值提取而来,而不是使用积分图计算不同位置和比例的矩形区域求和。特征算的快,并且特征池变小。跟提升分类器配合好。Feature design:1.channel原创 2015-11-04 13:27:31 · 4950 阅读 · 1 评论 -
论文笔记:CTC
1.label error rate:归一化的编辑距离。LER= 总ED/num2.pie关于x的条件概率:其中,y = Nw(x),是网络的输出序列,可以看做时刻t输出为k的概率。T是输出序列长度L+{blank} define as L’ 的长度.pie 是L’的元素,被称为path。3.指定label l 关于输入x的条件概率1) beta(-aa原创 2016-11-10 18:08:04 · 1594 阅读 · 0 评论 -
论文笔记:An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application
1.历史方法1)基于字符的DCNN,比如photoOCR.单个字符的检测与识别。要求单个字符的检测器性能很强,crop的足够好。2)直接对图片进行分类。9万个单词,组合成无数的单词,无法直接应用3)RNN,训练和测试均不需要每个字符的位置。但是需要预处理,从图片得到特征序列,独立做的,无法端到端。2.propose method1)端到端2)任意长度3)模型尺寸原创 2016-10-12 15:20:40 · 3934 阅读 · 0 评论 -
论文笔记:Precomputed Real-Time Texture Synthesis with Markovian Generative Adversarial Networks
1.什么是deconvolution反向映射,用来可视化深度特征,也可以用来重建图片。2.合成图片两种方式:1)生成全图的模型,效果不错但只对小图work,保真度fidelty低,一般用auto encoder2)马尔科夫模型,同时生成texture。可以捕获局部patch的统计信息。本文是第二种3.主要通过strided convolutional network取代po原创 2016-08-08 16:19:30 · 2580 阅读 · 1 评论 -
论文笔记:Perceptual Losses for Real-Time Style Transfer and Super-Resolution[doing]
1.transformation: image to image 2.perceptual losses: psnr是per-pixel的loss,值高未必代表图片质量好,广泛应用只是因为计算比较简单对图片做super resolutiondefine 4 loss:1.pixel L2 :feature reconstruction loss2.gram mat原创 2016-07-13 12:07:11 · 3084 阅读 · 0 评论 -
[论文笔记5]A Discriminatively Trained, Multiscale, Deformable Part Model
1.物体检测的模型组成:一个 global root filter 用来确定物体的位置(获取coarse features) 以及若干part model,包括spatial model 和 part filter(finer scale features)2.特征:hog一个block=2×2 cell = 4× 9(bins) = 36 D所以一个filter的weight是:4×原创 2015-11-16 15:34:18 · 1015 阅读 · 0 评论 -
[论文笔记1]Integral Channel Features
Introduced channel features and modified the VJ framework to compute integral images (and Haar wavelets) over the channels. Substantially outperformed HOG and at faster speeds.原创 2015-11-04 13:19:08 · 987 阅读 · 0 评论 -
[论文笔记4]Robust Object Detection via soft cascade
the trade-off between accuracy and speed1.降低计算量的一个常用方式是把分类器拆分成一个线性序列,cascade,其中每个次分类器,叫做stage,是二分类器,被训练来拒绝明显不是脸的部分,并且使的几乎所有的脸通过下一个stage.每个连续的stage基于之前stage通过的非脸部分来训练.复杂度降低通过早期比较容易拒绝掉大多数非脸来体现2.casc原创 2015-11-04 15:05:35 · 1924 阅读 · 0 评论 -
[论文笔记2]a fast feature pyramid
优化分类速度:for optimizing classification speed given precomputed image feature:1.cascades structure---soft cascade ,rather than attentional cascade structure in VJ.soft cascade的应用。Adaboost分类器包括2原创 2015-11-04 13:24:25 · 1546 阅读 · 0 评论 -
A Brief History of CNNs in Image Segmentation: From R-CNN to Mask R-CNN
At Athelas, we use Convolutional Neural Networks(CNNs) for a lot more than just classification! In this post, we’ll see how CNNs can be used, with great results, in image instance segmentation.转载 2017-04-25 13:29:09 · 3236 阅读 · 0 评论