privoxy用于socks5代理转换http代理。默认端口是8118。但是http代理没有密码，这个端口很容易被扫到，然后vps的流量就跑完了。所以注意要该端口，不要使用默认端口。
caffe 使用
数据准备
数据转成lmdb
build/tools/convert_imageset rootdir listfile output
listfile 里label从０开始
./build/tools/convert_imageset ./ ./data/collar/train.txt ./data/collar/collar_train_lmdb
./build/tools/convert_imageset ./ ./data/collar/test.txt ./data/collar/collar_test_lmdb
计算mean图
./build/tools/compute_image_mean ./data/collar/collar_train_lmdb ./data/collar/collar_mean.binaryproto
编写训练solver和模型文件
Rich feature hierarchies for accurate object detection and semantic segmentation
R-CNNs : combine region proposals with CNNs
we focused on two problems: localizing objects with a deep network and training a high-capacity model with only a small quantity of annotated detection data.
Deep Learning - Nature
transform the representation at one level (starting with the raw input) into a representation at a higher, slightly more abstract level.
- The first layer of representation typically represent the presence or absence of edges at particular orientations and locations in the image.
- The second layer typically detects motifs by spotting particular arrangements of edges, regardless of small variations in the edge positions.
- The third layer may assemble motifs into larger combinations that correspond to parts of familiar objects, and subsequent layers would detect objects as combinations of these parts.
matConvNet
Block : real 4D array, the first two dimensions index spatial coordinates (image rows and columns respectively), the third dimension feature channels (there can be any number), and the last dimension image instances.
caffe
softmax 是logistic回归的多值的扩展。
forward and backward
The Net::Forward() and Net::Backward() methods carry out the respective passes while Layer::Forward() and Layer::Backward() compute each step.
The Solver optimizes a model by first calling forward to yield the output and loss, then calling backward to generate the gradient of the model, and then incorporating the gradient into a weight update that attempts to minimize the loss.
Object Detection Using Strongly-Supervised Deformable Part Models
The success of training DPMs, hence, depends on the good initialization of model parts in general.
Part-based R-CNNs for Fine-grained Category Detection
Methods for pose-normalized representations have been proposed, but generally presume bounding box annotations at test time due to the difficulty of object detection.
The recent success of convolutional networks, like [27], on the ImageNet Challenge[23] has inspired further work on applying deep convolutional features to related image classification [14] and detection tasks [21].
Deformable part descriptors for fine-grained recognition and attribute prediction
Discriminative markings are often highly localized, leading traditional object recognition approaches to struggle with the large pose variation often
present in these domains.
As described in [23], what often differentiates basic-level categories is the presence or absence of parts (e.g. an elephant has 4 legs and a trunk), whereas subordinate
categories are more often discriminated by subtle variations in the shape, size and/or appearance properties of these parts (e.g. elephant species can be distinguished by
localized cues such as ear shape and size).
