grinding

Posted on 2018-02-04 | In life

Words count in article: 12

Nothing is given, Everything is earned.

— Lebron James

Take it easy, babe …

orbslam

Posted on 2018-03-19 | In Robotics , SLAM

Words count in article: 661

ORB-SLAM: a Versatile and Accurate Monocular SLAM System

Introduction

Use ORB features
Real time operation, using covisibility graph
real time loop closing based on essential graph
real time relocalization
initialization & model selection
survival of the fittest approach to the map point & keyframe selection

System Overview

Feature Choice

Three Threads: Tracking, Local Mapping & Loop Closing

Tracking
localizing camera and deciding when to insert a keyframe
- initial feature matching & motion-only BA
- IF tracking is lost
  - place recognition > global relocalization
- initial camera pose & feature matchings
  - local visible map using covisibility graph of keyframes
- matches found by reprojection & camera pose opt again
- decide whether to insert keyframe
Local Mapping
- process new keyframes and local BA
- unmatched ORB searched in connected keyframes in the covisibility graph to triangulate new points
- exigent point culling
Loop Closing
- search for loops with every new keyframe
- IF loop is detected
  - compute similarity transformation
  - both sides of the loop is aligned and duplicated points fused
  - pose graph opt > global consistency
    - MAIN NOVELTY: opt on Essential Graph

MapPoints, KeyFrames, Selection

MapPoint p_i

3D position: X_w,i
The viewing direction: n_i
ORB descriptor: D_i
d_max,d_min: maximum distance & minimum distance the point can be observed

KeyFrame K_i

T_iw, tranformation world2camera
Camera intrinsics, including focal length & principal point
Undistorted coordinates of ORB features

Covisibility Graph & Essential Graph**

Covisibility Graph: Undirected Weighted Graph
- node: keyframe
- edge: shared observations of the same MapPoints
Essential Graph
- spanning tree
  - INSERTED: new keyframe linked tokeyframe which shares most point observations
  - ERASED: update links affected by that KeyFrame

BoW Place Recognition

DBoW2: loop detection & relocalization

Automatic Map Initialization

GOAL

compute relative pose between 2 frames & triangulate an initial set of MapPoints.

STEPS

Find intial correspondences
- ORB
Parallel computation of two models
- homography H_cr
- fundamental matrix F_cr
  
  each interation, compute score S_m
Model selection
$R_H=\frac{S_H}{S_H+S_F}$
- IF R_H > 0.45
  - choose homography
- ELSE
  - fundamental matrix
Motion and SfM
- check if there is one solution with most points seen with parallax, in front of both cameras and with low reprojection error
BA
- full BA to refine initial reconstruction

Tracking

ORB Extraction
- FAST corners at 8 scale levels
- compute ORB descriptors & orientation
Initial Pose Estimation from Previous Frame
- IF tracking is SUCCESSFUL
  - constant velocity model
  - wider search
Initial Pose Estimation via Global Relocalization
- IF tracking LOST
  - convert frame into BoW
  - query for keyframe candidates for global relocalization
- alternatively RANSAC iterations for each keyframe, PnP
- camera pose with enough liniers
- camera pose opt
Track Local Map

once get initial estimation of camera pose & feature matches, we can project map into frame and search more map point correspondence

The local map also has a reference keyframe K_ref, which shares most map points with current frame
- compute mapPoints projection & discard outliers
- compute the angle between current viewing ray v and mapPoint mean viewing direction n. Discard if v * n<cos(60°)
- compute distance, discard if out of [dmin, dmax]
- compute scale
- compare D
New KeyFrame Decision

To insert a frame:
- More than 20 frames passed global reloc
- Local mapping idle OR >20 frames pass last keyframe insertion
- current frame tracks >=50 points
- current frame tracks <=90% of K_ref

Local Mapping

KeyFrame insertion
Recent MapPoints culling

in order to be retained in the map, pass a restrictive test during te firsr 3 keyframes after creation
- 25% the frames visible
- observed from >= last 3 key frames
New MapPoint Creation

Created by triangulating ORB from connected keyframes K_c
Local BA
- current processed KeyFrame K_i
- all keyframes connected to it
- all MapPoints seen bt those
  NOTE all other KFs not connected but see,remain fixed in opt
Local KF culling

GOAL detect redundant KFs & delete

Loop Closing

Loop Candidates Detection
- compute similarity
- discard below lowest score in DB
Compute Similarity Transformation
- ORB
- RANSAC iterations, find ST using Horn
  - IF enough inliers found
    - opt & guided search for more correspondences4
    - opt again & check
Loop Fusion
- fuse duplicated MapPoints
- insert new edges
Essential Graph Optimization

effectively correct drifts

Compileorbslam2

Posted on 2018-03-18 | In Robotics , SLAM

Words count in article: 142

Running ORB-SLAM2

ORB-SLAM2

GITHUB PROJECT: orbslam2

The instructions are written in detail. Additionally, I want to address some points.

Related Publication: ORB-SLAM2: an Open-Source SLAM System for Monocular, Stereo and RGB-D Cameras [pdf]

PreRequisites:

C++11 OR C++0x Compiler
Pangolin
OpenCV
Eigen3
DBoW2 & g2o
ROS (optional)

1. about OpenCV

suggest DON NOT try version > 3.0, I have failed multiple times.
Eventuall, I successfully run on OpenCV 2.4.11

compile OpenCV, check here

IF you want to uninstall OPENCV here

2. about Building

./build.sh

Remove -j of make -j to make sure your code can run smoothly.
IF your PC is powerful enough, IGNORE this!!!

3. Tips

Good Tutorial can be found here

OPENGL may be needed for Pangolin here

Stereo Datasets

The EuRoC MAV Dataset here
KITTI Dataset here
Others on CSDN here

orbslam2

Posted on 2018-03-18 | In Robotics , SLAM

Words count in article: 384

ORBSLAM2 (Stereo Camera)

Points can be reliably triangled if their depth is less than ~40 times the stereo baseline.
modern stereo SLAM: keyframe-based & Local BA
stereo LSD SLAM
direct method: performance could be severely degraded by unmodeled effects like rolling shutter or non-lambertian reflectance.

Structure

INPUT
Frame >
TRACKING
pre-process > pose prediction(motion model) OR relocalization > track local map > new keyframe decision >
LOCAL MAPPING
keyframe insertion > recent mappoints culling > new points creation > local BA > local keyframe culling >
LOOP CLOSING
querying DB > compute SE3 > loop fusion > opt essential graph >
FULL BA
full BA > update map

Tracking.preprocess: ORBExtractor (left & right rectified images) > stereo matching > stereo & mono keypoints

ORB-SLAM2

tracking: localize the camera with every frame by finding feature matches to the local map and minimizing the reprojection error applying motion-only BA
local mapping: local BA
loop closing: detect large loops & correct drifts

Notice: place recognition module based on DBoW2 for relocalization, reinitialization and loop detection.

use the same ORB features for tracking, mapping and place recognition tasks.

system handles mono & stereo keypoints (close OR far)

A. stereo point

close: associated depth < 40 times the stereo/RGBD baseline
far: > 40 times (accurate rotation, but weak scale & translation), triangulate when supported by multiple views

monocular point
- on the left image and stereo match could not be found OR invalid depth value

B. bootstrapping:

keyframe with the first frame, its pose to the origin, and create initial map with all stereo keypoints

C. BA with mono & stereo constraints

motion-only BA R,t, minimizing reprojection error
Local BA covisible frame and others (remain fixed in optimization)
full BA all key frames & points in the map

D. loop closing and full BA

detected and validated
corrected

stereo/depth info makes scale observable and geometric validation and the pose-graph opt based on rigid body transformation instead of similarity

IF a new loop is detected while opt, opt aborted and close the loop, which will launch a full BA again.
FROM updated to non-updated: spanning tree

E. keyframe insertion

inserted if num of the tracked close points drops below T1 and up T2

F. localization mode

lightweight in well mapped areas

loop closing deactivated, camera is continuously localized by tracking. VO matches ORB & 3D points

orbslam2-code

Posted on 2018-03-18 | In Robotics , SLAM

Words count in article: 640

ORB-SLAM2 code

structure

Tracking.cpp
LocalMapping.cpp
LoopClosing.cpp
Viewer.cpp

stereo_EuRoC

LoadImages()
create SLAM system: ORB_SLAM2::System SLAM
vector for tracking time statistics: vTimesTrack
rectify images: rectifyer
pass images to SLAM: SLAM.TrackStereo
wait to load next frame
SLAM.Shutdown()
time statistics
save trajectory

code

system

System::System()

- load ORB vocabulary (ORBVocabulary class, ORBVoc.txt)   
- create keyframe database (KeyFrameDatabase class, initialized with *mpVocabulary*) 
- create map  
- create drawers (used by map)   
- initialize tracking thread  
- initialize local mapping thread & launch 
- initialize loop closing thread & launch
- initialize Viewer thread & launch
- set pointers between threads

some important names:

//---ORB
mpVocabulary;
//---keyfrane
mpKeyFrameDatabase;
//---map
mpMap;
//---drawer
mpFrameDrawer;
mpMapDrawer;
//---tracking
mpTracker;
//---local mapping
mpLocalMapper;
mptLocalMapping;
//---loop closing
mpLoopCloser;
mptLoopClosing;
//---viewer
mpViewer

System::TrackStereo

- check GUI options   
- mpTracker->GrabImageStereo

System::SaveTrajectoryTUM

- mpMap->GetAllKeyFrames()

- transforamtion (1st keyframe is the origin) *GetPoseInverse()*

- framepose stored relative to its reference keyframe (lRit), the timestamp (lT), tracking state (lbL)

- if reference keyframe was culled, traverse the spanning tree to get a suitable keyframe

Tracking

Tracking::GrabImageStereo

- RGB to Gray

- mCurrentFrame 
Frame(mImGray,imGrayRight,timestamp,mpORBExtractorLeft,mpORBExtractorRight,mpORBVocabulary,mK,mDistCoef,mbf,mThDepth);

- Track()

- return mTcw (camera pose W2C)

Frame

Frame::Frame

// stereo initialization    
- Frame ID 
- get scale level info (ORBextractor class)   
- ORB Extraction   
- threadLeft (Frame::ExtractORB > .join())   
- threadRight (Frame::ExtractORB > .join())   
- UndistortKeyPoints()   
- ComputeStereoMatches(): compute depths if matches   
- depth info: mvuRight & mvDepth  
- mvpMapPoints, mvOutlier

Frame::UndistortKeyPoints

- *N* feature points  
- cv::undistortPoints()   
- mvKeysUn: corrected // mvKeys, mvKeysRight   
- **redundant** in stereo case

Frame::ComputeStereoMatches

- assign keypoints to row table // vRowIndices
- compute range of rows
- set limits for search //minD, maxD, minZ
- for each left keypoint search a match in the right
- SAD (subpixel match by correlation, IF |deltaR| >1, continue)
- matched points culling // mvuRight, mvDepth

Frame::UnprojectStereo

// backproject a keypoint (if stereo/depth info available) into 3D world coordinates.

// Rotation, translation & camera center
mRcw;   //Rotation from world2camera
mtcw;   //Translation from world2camera
mRwc;   //Rotation from camera2world
m0w;    //Translation from camera2world;

Tracking

//      |fx 0   cx|
//  K = |0  fy  cy| 
//      |0  0   1 |
//  corrected coef: [k1 k2 p1 p2 k3]
//  mThDepth: Threshold4 close/far points

Tracking::Track()

IF NOT_INITIALIZED
- StereoInitialization()
ELSE
- CheckReplacedInLastFrame()
    - IF mVelocity.empty()
    TrackReferenceKeyFrame()
    - ELSE
    TrackWithMotionModel()

Tracking::StereoInitialization

IF N > 500
- set pose to the origin
- create keyframe
- insert keyframe in the map
- create mappoints and associate to keyframe 
// Frame::UnprojectStereo()
// MapPoint::ComputeDistinctiveDescriptors() > find best descriptors for MapPoint; using median of dists
// MapPoint::UpdateNormalAndDepth() > update observations: mNormalVector & mfMaxDistance, mfMinDistance

Tracking::TrackReferenceFrame()

- mCurrentFrame.ComputeBoW();
- ORBmatcher.SearchByBoW()
- initialize pose by *mLastFrame*
- Optimizer::PoseOptimization(&mCurrentFrame)
- discard outliers

Tracking::TrackWithMotionModel

- Tracking::UpdateLastFrame
    - Const Velocity Model, estimate current pose
    - project points seen in previous frame
- Based on CVM, tracking MapPoints in the last frame
    - IF nmathces < 20, uses a wider window search (th > 2*th) 
    //---ORBmatcher.SearchByProjection
- optimize frame pose with all matches 
// Optimizer::PoseOptimization(&mCurrentFrame)
- discard outliers of mvpMapPoints(feature >> MapPoint)

Tracking::UpdateLastFrame()

- update pose according to reference keyframe
// mlRelativeFramePoses: store the reference keyframe for each frame and its relative transformation
- IF stereo OR RGBD
    - sort points according to measured depth
    - rank depths in ascending order
    - IF nPoints > 100, break

Tracking::Relocalization()

Relocalization is performed when tracking is lost
- compte BoW Vector
- mpKeyFrameDB->DetectRelocalizationCandidates(&mCurrentFrame)
- ORB matching with each candidate 
    - IF enough matches, set up PnP solver
- perform iterations of P4P RANSAC, until a camera pose supported by enough inliers
- Optimizer::PoseOptimization(&mCurrentFrame)
- IF few inliers,search by projection & optimization

ORBmatcher

ORBmatcher::SearchByProjection

SearchByProjection(currentFrame, lastFrame, th, bMono)

1. project MapPoints in the last frame
2. match & culling

KeyFrameDatabase

KeyFrameDatabase::DetectRelocalizationCandidates

find similar keyframes in relocalization
- search all keyframes that share a word with current frame
- find keyframes share enough words & decide 
    Th: minCommonWords = maxCommonWords*0.8f
- compute similarity score
- accumulate score by covisiblity
    One Group: Keyframe + GetBestCovisibilityKeyFrames(10)
    >> bestAccScore & minScoreToRetain = 0.75f*bestAccScore
    return group(>minScoreToRetain) memeber with highest score

slam-learning

Posted on 2018-02-26 | In slam

Words count in article: 452

SLAM

slam, (Simultaneous Localization and Mapping)

传感器数据 → 前端视觉里程计 → 后端非线性优化 → 建图
→ 回环检测 →

核心

运动方程

$x_k=f(x_{k-1},u_k,w_k)$

其中，u_k是运动传感器的参数，w_k是噪声

观测方程

$z_{k,j}=h(y_j,x_k,v_{k,j})$

y_j是路标点，z_k,j是观测数据
$a\times{b}=\begin{Vmatrix}i&j&k\\a_1&a_2&a_3\\b_1&b_2&b_3\end{Vmatrix}=\begin{bmatrix}{a_2}{b_3}-{a_3}{b_2}\\{a_3}{b_1}-{a_1}{b_3}\\{a_1}{b_2}-{a_2}{b_1}\\\end{bmatrix}=\begin{bmatrix}0&-a_3&a_2\\a_3&0&-a_1\\-a_2&a_1&0\\\end{bmatrix}b\triangleq{a\wedge{b}}$

特殊正交群
$SO(n)=\{R\in{R^{n\times{n}}|RR^T=I\}$

特殊欧式群
$SE(3)={T=\begin{bmatrix}R&t\\0^{T}&1\\\end{bmatrix}}$

罗德里格斯公式
$R=\cos\theta{I}+(1-\cos{\theta})nn^{T}+\sin\theta{n}^{\wedge}$

tr(R)=1+2cosθ
Rn=n

相似变换
$T_s=\begin{bmatrix}sR&t\\0^T&1\\\end{bmatrix}$

仿射变换
$T_A=\begin{bmatrix}A&t\\0^T&1\end{bmatrix}$

射影变换
$T_p=\begin{bmatrix}A&t\\a^T&v\end{bmatrix}$

李代数
$so(3)李代数$

$se(3)李代数$

SO(3)指数映射
$Z\begin{bmatrix}u\\v\\1\end{bmatrix}=\begin{bmatrix}f_x&0&c_x\\0&f_y&c_y\\0&0&1\end{bmatrix}\begin{bmatrix}X\\Y\\Z\end{bmatrix}\triangleq{KP}$

SE(3)指数映射
$ZP_{uv}=Z\begin{bmatrix}u\\v\\1\end{bmatrix}=K(RP_w+t)=KTP_w$

其中， $\frac{z-f}{z}=\frac{b-u_L+u_R}{b}$

李代数求导&扰动模型
李代数求导：
$\frac{\partial(RP)}{\partial\phi}=(-Rp)^\wedge{J_l}$

扰动模型(左)：
$\frac{\partial(Rp)}{\partial\varphi}=-(Rp)^\wedge$

相机与图像

相机模型:
$\frac{Z}{f}=-\frac{X}{X'}=-\frac{Y}{Y'}$

$Z\begin{bmatrix}u\\v\\1\end{bmatrix}=\begin{bmatrix}f_x&0&c_x\\0&f_y&c_y\\0&0&1\end{bmatrix}\begin{bmatrix}X\\Y\\Z\end{bmatrix}\triangleq{KP}$

$ZP_{uv}=Z\begin{bmatrix}u\\v\\1\end{bmatrix}=K(RP_w+t)=KTP_w$

切向畸变&径向畸变
双目相机模型
$\frac{z-f}{z}=\frac{b-u_L+u_R}{b}$

得到， $z=\frac{fb}{d},d=u_L-u_R$

非线性优化

非线性最小二乘

$\begin{Vmatrix}f(x+\Delta{x})\end{Vmatrix}_2^2\approx\begin{Vmatrix}f(x)\end{Vmatrix}_2^2+J(x)\Delta{x}+\frac{1}{2}\Delta{x}^{T}H\Delta{x}$

保留一阶梯度
$\Delta{x}*=-J^T(x)$

保留二阶梯度
$H\Delta{x}=-J^T$

高斯牛顿法
$f(x+\Delta{x})\approxf(x)+J(x)\Delta{x}$

最小二乘:
$J(x)^{T}J(x)\Delta{x}=-J(x)^Tf(x)$

令左式系数为H，右边为g： $H\Delta{x}=g$

列文伯格-马尔夸尔特法
定义ρ来确定信赖区域范围
$\rho=\frac{f(x+\Delta{x}-f(x))}{J(x)\Delta{x}}$

第k次迭代求解：
$\min_{\Delta{x_k}}\frac{1}{2}\begin{Vmatrix}f(x_k)+J(x_k)\Delta{x_k}\end{Vmatrix}^{2},s.t.\begin{Vmatrix}D\Delta{x_k}\end{Vmatrix}\leqslant\mu{,}$

其中，D=I
解得： $(H+\lambda{I})\Delta{x}=g$

视觉里程计1

前端 → 后端,提供好的初始值;特征点（Feature）作为路标,由关键点&描述子组成。

特征提取

ORB

FAST角点提取
- 极大值抑制(NMS)
- 灰度质心法(Intensity Centroid)
BRIEF描述子
- 汉明距离(Hamming Distance)
- 快速近似最近邻(FLANN)

根据匹配点估计相机运动

对极几何
P=[X,Y,Z]^T,
s₁p₁=KP, s₂p₂=K(RP+t)
得到：Ｅ＝t^R, F=K^-TEK^-1, x_2^TEx₁=p₂^TFp₁=0
- 本质矩阵
  E=t^R
- 单应矩阵
  $n^Tp+d=0,p2=Hp1,H=K(R-\frac{tn^T}{d})K^{-1}p_1$
- 本质矩阵
  $s_1x_1=s_2Rx_2+t$

ICP(双目相机 OR RGB-D)
PnP(3D-2D)
重投影误差
$\xi^*=arg\min_{\xi}\frac{1}{2}\sum_{i=1}^n\begin{Vmatrix}u_i-\frac{1}{s_i}Kexp(\xi^\wedge)P_i\end{Vmatrix}_2^2$
关于相机位姿的一阶变化关系：
$\frac{\partial{e}}{\partial\delta\xi}=\lim_{\delta\xi\rightarrow{0}}\frac{e(\partial\xi\oplus\xi)}{\partial\xi}=\frac{\partial{e}}{\partial{P'}}\frac{\partial{P'}}{\partial\delta\xi}$
同理，得到关于位置的一阶变化关系
$\frac{\partial{e}}{\partial{P}}=\frac{\partial{e}}{\partial{P'}}\frac{\partial{P'}}{\partial{P}}$

论文

《Real-Time Visual Odometry from Dense RGB-D Images》

energy minimizaiton
maximizing photoconsistency
coarse-to-fine

《Semi-direct tracking and mapping with RGB-D camera for MAV》

direct method
feature
measurement error function (depth & geometric)

ssd

Posted on 2018-02-22 | In deep learning

Words count in article: 942

SSD: Single Shot MultiBox Detector
从 Github 上面下载源工程代码：caffe-SSD

配置Caffe-SSD

进入caffe-ssd 主目录：cp /home/xxx/…/caffe-ssd/
1
cp Makefile.config.example Makefile.config
编译项目：（进入 caffe-ssd 主目录）
1
2
3
4
make -j8
make py
make test -j8
make runtest -j8
CUDA 版本比较高的需要注释掉config里面最后一行内容：

Makefile.config

数据文件准备

预训练模型（VGG）：VGG_ILSVRC_16_layers_fc_reduced.caffemodel
（下载地址：密码: t9ub）

下载完毕后将VGG模型放到caffe主目录下 models\VGGNet 下面(如果没有的话，models 下面没有的话mkdir VGGNet)

VOC2007 和 VOC2012 数据集

进入caffe主目录下的data目录：

1
2
3

wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar

如果安装失败，请转到：（VOC 2007 & 2012 Dataset 密码：j3in ）

紧接着解压：

1
2
3

tar -xvf VOCtrainval_11-May-2012.tar 
tar -xvf VOCtrainval_06-Nov-2007.tar 
tar -xvf VOCtest_06-Nov-2007.tar

将数据转换为caffe处理的数据类型（LMDB）：

cd caffe主目录，执行：
1
2
./data/VOC0712/create_list.sh
./data/VOC0712/create_data.sh

注意在执行 create_data.sh 如果提示no module caffe 的话，用如下指令：
export PYTHONPATH=$PYTHONPATH:/home/xxx/.../caffe主目录/python（自行修改中间路径）

训练示例
在caffe主目录下面运行：
python examples/ssd/ssd_pascal.py

评估 & 检测

1 2	python examples/ssd/score_ssd_pascal.py python examples/ssd/ssd_detect.py

这里注意指定使用的快照模型的路径 & 在caffe主目录下面运行程序
ssd_detect.py

训练

训练准备
创建自己的数据目录myData:
1
2
cd /data
mkdir myData
将/data/VOC0712下面的create_list.sh,create_data.sh,labelmap_voc.protoxt这三个文件拷贝到’data/myData’:
1
2
cp data/create* ./myData
cp data/label* ./myData
在/data/VOCdevkit目录按照VOC数据框架下面创建myData,用来存放自己的数据集
1
2
3
4
5
6
7
8
9
10
cd data/VOCdevikit
mkdir myData
cd myData
mkdir Annotations
mkdir ImageSets
mkdir JPEGImages
cd ImageSets
mkdir Layout
mkdir Main
mkdir Segmentation
一般地，我们只需要关注：
Annotations：XML描述文件
ImageSets: Main目录下面放 train.txt, val.txt, trainval.txt, test.txt
JPEGImages:存放所有图片
制作VOC数据集
按照VOC Dataset要求整理好数据集后，将之转换为caffe的输入数据。首先，根据自己数据集特点，修改labelmap_voc.protxt,注意保留item中background类，其余的类别可以按照自己的需要照葫芦画瓢，给一个简单的示例：

然后，依次运行create_list.sh,create_data.sh.注意修改sh中的路径到你自定义的数据集路径。
需要注意的参数有：
create_data.sh: data_root_dir, data_name,mapfile
create_list.sh: root_dir,
1
2
3
4
5
# create_list.sh 中应该注释掉
# if [[ $dataset == "test" && $name == "VOC2012" ]]
# then
# continue
# fi

训练
在python主目录下运行命令：
python examples/ssd/ssd_pascal.py

#需要指定的路径与参数
...
train_data = xxx/xxx_test_lmdb
test_data = xxx/xxx_trainval_lmdb
...
model_name =
save_dir =
snapshot_dir =
job_dir = 
output_result_dir =
...
name_size_file =
pretrain_model =
label_map_file =
...
num_classes =
...
num_test_image =

Note: solver parameters中GPU的指定,个数不要超过可用个数，可以用nvidia-smi来查看可用GPU情况;另外，也可以调整solver_param参数，比如：iter_size, max_iter, etc.

检测

score_pascal.py
Note: 注意修改参数与ssd_pascal.py中的路径相同
ssd_detect.py**
Note: 检测单张图片，指定‘–gpu_id’, ‘–model_def’, ‘–model_weights’, ‘–image_file’.

批量完成test images的可视化

利用caffe主目录下build/examples/ssd/ssd_detect.bin对test结果进行文本输出，输出的格式为 ( path,label,confidence,xmin,ymin,xmax,ymax )

caffe root 下执行：

build/examples/ssd/ssd_detect.bin models/VGGNet/mydataset/SSD_300x300/deploy.prototxt 	models/VGGNet/mydataset/SSD_300x300/mydataset_SSD_300x300_iter_100236.caffemodel 	data/VOCdevkit/mydataset/test_img_path.txt --confidence_threshold 0.5 --out_file output.txt`

利用plot_detections.py进行检测结果的可视化

output.txt 是 ssd_detect.bin 生成的检测结果的txt文档

1	python examples/ssd/plot_detections.py output.txt /home/wxb/caffe-ssd --labelmap-file data/mydataset/labelmap_voc.prototxt --save-dir results/bbox_results/SSD_300x300/Main/img/

html_experiment

Posted on 2018-02-11 | In learning , html

Words count in article: 1,116

A html example:

you can check the results of following code here

<!DOCTYPE HTML>
<html>
<head>
<meta name="Xingbo WANG"
content="andywangxb.github.io">
<meta http-quiv="Content-Type" content="text/html"; charset=gb2312" />
<meta http-equiv="Refresh" content="5;url=https://andywangxb.github.io" /> 
<!--  title of web  -->
<title>html experiments</title>
<!-- style of web -->
<style type="text/css">
h1 {color: green}
p {color: black}
span.red {color:red;}
#header {
	background-color:black;
	color:white;
	text-align:center;
	padding:5px;
}
#nav{
	line-height:30px;
	background-color:#eeeeee;
	height:300px;
	width:100px;
	float:left;
	padding:5px;
}
#section{
	width:350px;
	float:left;
	padding:10px;
}
#footer{
	background-color:black;
	color:white;
	clear:both;
	text-align:center;
	padding:5px;
}
</style>
<!-- outer style -->
<link rel ="stylesheet" type="text/css" href="/html/csstest1.css">
</head>
<!-- visible part -->
<body bgcolor="lightgrey">
<!-- this is experiment -->
<!-- heading-->
<h1 align="center">h1 heading</h1>
<h2 style="background-color:red">h2 heading</h2>
<h3 style="text-align:right">h3 heading</h3>
<h4>h4 heading</h4>
<h5>h5 heading</h5>
<h6>h6 heading</h6>
<!-- paragraph -->
<p>one paragraph</p>
<p>another paragraph</p>
<hr /><!-- split -->
<!-- link -->
<a href="https://andywangxb.github.io"> my personal homepage</a>
<p>
<a href ="/index.html">This </a >is directed to a link of this website.</p>
<p>
<a href ="http://www.qq.com">This </a>is directed to a link outside this website</p>
<a href ="http://www.qq.com" target="_blank">This</a> will open a new page directed to <i>qq.com</i>
<p> you can mail me at<a href ="mailto: wangxbzb@hotmail?subject=Hello%20again">subject: hello again</a>
</p>
<hr />
<!-- insert image -->
<map>
<p>Image
<img src="/images/photo.png" align="center" alt="photo.png" width="100" height="100"/>
among the texts</p>
</map>
<hr />
<!-- word style -->
<b> this text is bold </b>
<br />
<strong> this text is strong </strong>
<br />
<big> this text is big</big>
<br />
<em> this text is emphasized</em>
<br />
<i> this text is italic</i>
<br/>
<small>this text is small</small>
this text contains <sub>subscript</sub>
<br />
this text contains <sup>superscript</sup>
<hr />
<pre>
	This is pre tag.
	it can be used to demonstarte codes:
	for i in range(0,10):
		print(i)
</pre>
<code> Computer Code </code>
<br />
<kbd>keyboard input</kbd>
<br />
<tt>teletype text</tt>
<br />
<samp>sample text</samp>
<br />
<var>computer variable</var>
<br />
<hr />
<address>
written by <a href ="wangxbzb@hotmail.com">Xingbo WANG</a>.<br>
Visit us at:<br>
andywangxb.github.io<br>
Wuhan, China<br>
</address>
<hr />
<abbr title="etcetera">etc.</abbr>
<br />
<acronym title="world wide web">www</acronym>
<hr />
<bdo dir="rtl">
here is some Hebrew text
</bdo>
<hr />
<blockquote>
this is for long quotes.this is for long quotes.this is for long quotes.
</blockquote>
<br />
<q>this is for short quotes</q>
<br />
For example:
<q> this is an example from WWF website</q>
<blockquote cite="https://www.worldwildlife.org/who/index.html">
五十年来，WWF 一直致力于保护自然界的未来。世界领先的环保组织，WWF 工作于 100 个国家，并得到美国一百二十万会员及全球近五百万会员的支持。
</blockquote>
<hr />
<p> a "dozen" is not<del> twenty </del> <ins> twelve </ins>
</p>
<hr />
<p><cite>The Scream</cite> by Edward Munch. Painted in 1893.</p>
<hr />
<!-- table -->
<p><h4 align="center">table example</h4>
<table border="1">
<caption>name</caption>
<tr>
<th>heading A</th>
<th>heading B</th>
</tr>
<tr>
<td> row 1 , cell 1 </td>
<td> row 1 , cell 2 </td>
</tr>
<tr>
<td> row 2 , cell 1 </td>
<td> row 2 , cell 2 </td>
</tr>
</table>
</p>
<hr />
<p>unorganized list
<ul>
<li>coffe</li>
<ul>
<li>black coffe</li>
<li>latte</li>
</ul>
<li>milk</li>
</ul>
</p>
<p>organized list
<ol>
<li>coffe</li>
<li>milk</li>
</ol>
</p>
<p>DIY table
<dl>
<dt>computer</dt>
<dd>device</dd>
<dt>monitor</dt>
<dd>device</dd>
</dl>
</p>
<hr />
<!-- span -->
<p><h1>My <span class="red">Important</span> Heading</h1>
</p>
<hr />
<!-- div style -->
<div id="header">
<h1>city gallery</h1>
</div>
<div id="nav">
London<br>
Paris<br>
Tokyo<br>
</div>
<div id="section">
<h1>London</h1>
<p>London is the capital city of England.
</p>
</div>
<div id="footer">
CopyRight
</div>
<hr />
<!-- frame -->
<iframe <!--src="http://www.qq.com"--> name="tencent" width="100%" height="200" frameborder="0"></iframe>
<p>
<a href="http://www.baidu.com" target="tencent">baidu.com</a>
</p>
<hr />
<!-- insert script-->
<script type="text/javascript">
document.write("hello world!")
</script>
<hr />
<!-- special symbols -->
<p>
&nbsp space<q>&nbsp</q><br>
5 &lt 10<br>
5 &amp 10<br>
&pound 5<br>
&yen 10<br>
&cent 10<br>
&reg<br>
&trade<br>
5 &times 10<br>
10 &divide 5<br>
<br>
</p>
<hr />
<!-- form -->
<form>
 First name:<br>
<input type="text" name="firstname">
<br>
 Last name:<br>
<input type="text" name="lastname">
<br>
<input type="radio" name="sex" value="male" checked>Male
<br>
<input type="radio" name="sex" value="female">Female
<br>
</form>
<hr />
</body>
</html>

anaconda

Posted on 2018-02-07 | In learning , python

Words count in article: 340

Use Anaconda to manage your python envs

Anaconda is very convenient tool to manage multiple virtual envs of python in your PC.
So you can enjoy different version of python, avoid conflictions between different projects.
This blog I will introduce how to install anaconda & install envs for tensorflow(CPU)、Scrapy

Install anaconda

Go to the official website to download latest Anaconda.
Install Anaconda with default configurations.
- Of course you can change the location wherever you like to install it;
- Don’t recommend you to add Anaconda to the system PATH env var.

Manage envs with Anaconda

common commands:

# check current env
conda info -e
# create new env with specific version of python 
conda create -n env_name python=2.7
# switch envs
activate env_name
# exit env
deacticate env_name
# remove env
conda remove -n env_name --all

Manage packages with Anaconda

# install new pkgs
conda search pkg_name
conda install pkg_name
# check installed pkg list
conda list
# update pkgs
conda update pkg_name
conda update anaconda
# remove pkgs
conda removw pkg_name

Use domestic mirrors

In order to install and upgrade easily in China. We should specifically allocate domestic mirrors to it.

find .condarcin C:\Users\user_name\.condarc
remove - default
Open Anaconda Prompt,input:

conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/msys2/

conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/conda-forge/

conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/

conda config --set show_channel_urls yes

To ensure that pkgs you use does not interfere with each other. Better to create new envs for them.

Install tenflow(CPU)

1
2
3

conda create -n tensorflow python=3.5
activate tensorflow
conda install tensorflow

Install Scrapy

1
2
3

conda create -n py27 python=2.7
activate py27
conda install -c conda-forge scrapy

Wait a moment & have a cup of tea~

Github-Blog

Posted on 2018-02-05 | In learning , website

Words count in article: 225

Use Github to create your own blog

Register for Github.io

register for Github account
create a new repository:
username.github.io

Install Hexo

Env: Ubuntu 16.04

Dependencies:
- Git
- NodeJs
- Hexo
Methods:
- Git: sudo apt-get install git
- NodeJs: NodeJs+NPM
  - if you wan to upgrade your nodejs, you could install a module named “n”
    1
    2
    sudo npm install -g -n
    sudo n stable
- Hexo: sudo npm install hexo-cli -g

Write &Publish

initialization
hexo init username.github.io
configuration
- install the theme
  cd username.github.io
  use default and popular theme: next
  git clone https://github.com/iissnan/hexo-theme-next themes/next
  for more themes, you can check the links: HEXO THEMEs
- basic configuration
  edit username.github.io/_config.yml
  Hexo Basic Confs: More Confs
  Notice:
  1
  2
  3
  4
  5
  6
  7
  title: [blog name]
  author: [your name]
  language: [zh-Hans、en]
  theme: __next__
  deploy:
  type: git
  repo: git@github.com:username/username.github.io
wirte
hexo new [layout] "essay name"
test
hexo s
deploy
- install hexo-deployer-git hexo-deployer-git tool:
  npm install hexo-deployer-git --save
publish
- hexo clean
- hexo g
- exo d
browse

useful links to learn to build your blog

Hexo: https://hexo.io/
NexT: http://theme-next.iissnan.com/
Individual Settings: 1, 2

ORB-SLAM: a Versatile and Accurate Monocular SLAM System

Introduction

System Overview

Feature Choice

Three Threads: Tracking, Local Mapping & Loop Closing

MapPoints, KeyFrames, Selection

Covisibility Graph & Essential Graph**

BoW Place Recognition

Automatic Map Initialization

Tracking

Local Mapping

Loop Closing

Running ORB-SLAM2

ORB-SLAM2

GITHUB PROJECT: orbslam2

PreRequisites:

1. about OpenCV

2. about Building

3. Tips

Stereo Datasets

ORBSLAM2 (Stereo Camera)

Related Work

Structure

ORB-SLAM2

ORB-SLAM2 code

structure

stereo_EuRoC

code

system

Tracking

Frame

Tracking

ORBmatcher

KeyFrameDatabase

SLAM

核心

运动方程

观测方程

相机与图像

非线性优化

非线性最小二乘

视觉里程计1

特征提取

根据匹配点估计相机运动

论文

配置Caffe-SSD

训练

检测

A html example:

Use Anaconda to manage your python envs

Install anaconda

Manage envs with Anaconda

Manage packages with Anaconda

Use domestic mirrors

Install tenflow(CPU)

Install Scrapy

Use Github to create your own blog

Register for Github.io

Install Hexo

Write &Publish

useful links to learn to build your blog