Rebuild my 3D world [open source] [serialization-1]

I want to do something interesting recently ： Rebuild around me 3D The world . This sounds like a cool thing , In fact, many people have done , And did well , I know there is a big wisdom map 、Pix4D、PhotoScan、ContextCapture And other commercial software that has already been done very well . But it doesn't need business to spend a little time License Ready to use software still seems to make sense , Anyway, I want to write some code .

thank Github The open source community , I plan to use it first VS2019+QT+ Open source algorithm library OpenMVG+OpenMVS To do this , Simply take the first step , If there is time later , Maybe there will be more interesting things to do .

As QT beginner , It is bound to appear rough in interface development （ Very slow, too ！）.

Please forgive me .

compile OpenMVG+OpenMVS

First , We went to the Github On Clone OpenMVG as well as OpenMVS Source code , And then use cmake-gui Compile . This is a troublesome process , The process doesn't show .

PS. compiled OpenMVG(vs2019) engineering , I put it on Baidu online disk , link ：https://pan.baidu.com/s/1X9nBeWHXeHR8ETSIqH23XQ Extraction code ：1i42, Interested students download by themselves .

then , download QT, The version number I selected is 5.12.9, Largely because this version is free .

Last , Build a QT UI engineering , Take a pretending name ：TellusBuilder.

Besides , I also use third-party libraries Eigen For matrix operations and OpenCV For images IO, The version numbers are 3.2.8 and 3.1.0.

1. Add an image and display

After building the project , The first function we need to achieve is to add images and display , Add a name "Add Images" The menu item of , And write the slot function in the main framework class , What the slot function needs to do is to open the dialog box and import multiple images .

At the same time, design and implement a QTreeView, It is attached to a DockPane above , Imported image information （ The most intuitive is the image name ） Will be in Tree Control is displayed one by one .

I also implemented a simple interactive function ： double-click Tree Control , It's on the right 2D The image is displayed in the view window . Concrete , After double clicking the command , I will follow the absolute path of the image , Import the image and render .

Um. , Simple and full of temptation to Xiaobai ！

These operations took not a long time , Under the final rendering ：

At present only 2K There is such an interface effect on the display .（ cough , Adaptive window adjustment is not done yet , food ！）

2. link OpenMVG, Realization SFM

SFM, It is well known Structure From Motion（ Exercise restores structure ）, This term sounds very professional , Simply speaking , This step does two things ：

1. Restore the camera of each image at the time of shooting Postures （ Position and posture ）, There are two ways of expression , One is commonly used in the field of photogrammetry $X,Y,Z,\varphi ,\omega ,\kappa$ Expression ; The second is commonly used in the field of computer vision $R,t$ Expression . They are equivalent , It's just in different forms , To describe the position and attitude of the camera , Popular said , It is to determine where and in which direction each image is taken by the camera .
2. Restore the intrinsic internal parameters of the camera , Like focal length 、 Principal point 、 Distortion parameters, etc , This is optional , Because in some cases, the camera has calibrated these internal parameters in advance .

There are many open source libraries that can complete this step , such as OpenMVG、Colmap、MVE etc. , This time I choose OpenMVG, It integrates incremental 、 Global 、 Hybrid SFM, For unordered image sets , I expect it to have a more robust performance .

OpenMVG Of Github Address ：Github : openMVG
OpenMVG Document address ：OpenMVG : The Open Multiple View Geometry Library

stay OpenMVG Introduction , There is a picture , The description is SFM Basic steps ： First, extract the feature descriptor , Then match the feature points between the images , And then through SFM Complete the recovery of the whole structure . Of course, the actual algorithm implementation will be far more complex than the description , There is no extension here .

( The picture is from OpenMVF Of Github Introduce )

Okay , That's all for today , The first step of the long march , I feel more like a software development exercise at present , Even the whole cycle, I think, may be like this , But I think it doesn't matter , Build the software framework , We can do more . The visual interface is still more comfortable than the console , It's not excluding the console , It may be that there are too many algorithms developed to find something new , Some people must think the console is cooler ！

Next, let's see how to use OpenMVG Library to complete SFM The process , So as to get the position and posture of each figure , At the same time, there will be some sparse points in 3D space and camera rendering .

I will put the code in my Github Warehouse ：TellusBuilder in , Interested students , It is suggested to click star and watch, Of course, based on my work, what I want to do in the future is also welcome fork ah .

If you think this series is good , Please let me know in the comments section , Maybe I will update more diligently ！