Logo Questions Linux Laravel Mysql Ubuntu Git Menu
 

THREE.JS & Reality Capture - Rotation issue photogrammetry reference camera's in a 3D space

Tags:

Thanks for taking the time to review my post. I hope that this post will not only yield results for myself but perhaps helps others too!

Introduction

Currently I am working on a project involving pointclouds generated with photogrammetry. It consists of photos combined with laser scans. The software used in making the pointcloud is Reality Capture. Besides the pointcloud export one can export "Internal/External camera parameters" providing the ability of retrieving photos that are used to make up a certain 3D point in the pointcloud. Reality Capture isn't that well documented online and I have also posted in their forum regarding camera variables, perhaps it can be of use in solving the issue at hand?

Only a few variables listed in the camera parameters file are relevant (for now) in referencing camera positioning such as filename, x,y,alt for location, heading, pitch and roll as its rotation.

Internal/External camera parameters file

Currently the generated pointcloud is loaded into the browser compatible THREE.JS viewer after which the camera parameters .csv file is loaded and for each known photo a 'PerspectiveCamera' is spawned with a green cube. An example is shown below:

3D view of the pointcloud + placed camera references with its frustum

The challenge

As a matter of fact you might already know what the issue might be based on the previous image (or the title of this post of course ;P) Just in case you might not have spotted it, the direction of the cameras is all wrong. Let me visualize it for you with shabby self-drawn vectors that rudimentary show in what direction it should be facing (Marked in red) and how it is currently vectored (green).

3D view of the pointcloud displaying the invalid camera orientation vector vs the correct vectors Overview of the in red marked camera refs

Row 37, DJI_0176.jpg is the most right camera with a red reference line row 38 is 177 etc. The last picture (Row 48 is DJI_189.jpg) and corresponds with the most left image of the clustured images (as I didn't draw the other two camera references within the image above I did not include the others).

When you copy the data below into an Excel sheet it should display correctly ^^

#name   x   y   alt heading pitch   roll    f   px  py  k1  k2  k3  k4  t1  t2
DJI_0174.JPG    3.116820957 -44.25690188    14.05258109 -26.86297007    66.43104338 1.912026354 30.35179628 7.25E-03    1.45E-03    -4.02E-03   -2.04E-02   3.94E-02    0   0   0
DJI_0175.JPG    -5.22E-02   -46.97266554    14.18056658 -16.2033133 66.11532302 3.552072396 30.28063771 4.93E-03    4.21E-04    1.38E-02    -0.108013599    0.183136287 0   0   0
DJI_0176.JPG    -3.056586953    -49.00754998    14.3474763  4.270483155 65.35247679 5.816970677 30.50596933 -5.05E-03   -3.53E-03   -4.94E-03   3.24E-02    -3.84E-02   0   0   0
DJI_0177.JPG    -6.909437337    -50.15910066    14.38391206 19.4459053  64.26828897 6.685020944 30.6994734  -1.40E-02   4.72E-03    -5.33E-04   1.90E-02    -1.74E-02   0   0   0
DJI_0178.JPG    -11.23696688    -50.36025313    14.56924433 19.19192622 64.40188316 6.265995184 30.7665397  -1.26E-02   2.41E-03    1.24E-04    -4.63E-03   2.84E-02    0   0   0
DJI_0179.JPG    -16.04060554    -49.92320365    14.69721478 19.39979452 64.85507307 6.224929846 30.93772566 -1.19E-02   -4.31E-03   -1.27E-02   4.62E-02    -4.48E-02   0   0   0
DJI_0180.JPG    -20.95614556    -49.22915437    14.92273203 20.39327092 65.02028543 6.164031482 30.99807237 -1.02E-02   -7.70E-03   1.44E-03    -2.22E-02   3.94E-02    0   0   0
DJI_0181.JPG    -25.9335097 -48.45330177    15.37330388 34.24388008 64.82707628 6.979877709 31.3534556  -1.06E-02   -1.19E-02   -5.44E-03   2.39E-02    -2.38E-02   0   0   0
DJI_0182.JPG    -30.40507957    -47.21269946    15.67804925 49.98858409 64.29238807 7.449650513 31.6699868  -8.75E-03   -1.31E-02   -4.57E-03   2.31E-02    2.68E-03    0   0   0
DJI_0183.JPG    -34.64277285    -44.84034207    15.89229254 65.84203906 62.9109777  7.065942792 31.78292476 -8.39E-03   -2.94E-03   -1.40E-02   8.96E-02    -0.11801932 0   0   0
DJI_0184.JPG    -39.17179024    -40.22577764    16.28164396 65.53938063 63.2592604  6.676581293 31.79546988 -9.81E-03   -8.13E-03   1.01E-02    -8.44E-02   0.179931606 0   0   0
DJI_0185.JPG    -43.549378  -33.09364534    16.64130671 68.61427166 63.15205908 6.258411625 31.75339036 -9.78E-03   -7.12E-03   4.75E-03    -6.25E-02   0.1541638   0   0   0
DJI_0186.JPG    -46.5381556 -24.2992233 17.2286956  74.42382577 63.75110346 6.279208736 31.88862443 -1.01E-02   -1.73E-02   1.02E-02    -6.15E-02   4.89E-02    0   0   0
DJI_0187.JPG    -48.18737751    -14.67333218    17.85446854 79.54477952 63.0503902  5.980759013 31.69602914 -8.83E-03   -1.01E-02   -7.63E-03   -7.49E-03   2.71E-02    0   0   0
DJI_0188.JPG    -48.48581505    -13.79840485    17.84756621 93.43316271 61.87561678 5.110113503 31.6671977  1.99E-03    -9.40E-04   2.40E-02    -0.180515731    0.32814456  0   0   0
DJI_0189.JPG    -48.32815991    -13.88055437    17.77818573 106.3277582 60.87171036 4.039469869 31.50757712 2.84E-03    4.12E-03    8.54E-03    -1.32E-02   3.89E-02    0   0   0

Things tried so far

Something we discovered was that the exported model was mirrored from reality however this did not affect the placement of the camera references as they aligned perfectly. We attempted to mirror the referenced cameras, pointcloud and viewport camera but this did not seem to fix the issue at hand. (hence the camera.applyMatrix4(new THREE.Matrix4().makeScale(-1, 1, 1));)

So far we attempted to load Euler angles, set angles directly or convert and apply a Quaternion sadly without any good results. The camera reference file is being parsed with the following logic:

// Await the .csv file being parsed from the server
    await new Promise((resolve) => {
      (file as Blob).text().then((csvStr) => {
        const rows = csvStr.split('\n');
        for (const row of rows) {
          const col = row.split(',');
          if (col.length > 1) {
            const suffixes = col[0].split('.');
            const extension = suffixes[suffixes.length - 1].toLowerCase();
            const validExtensions = ['jpeg', 'jpg', 'png'];
            if (!validExtensions.includes(extension)) {
              continue;
            }
            // == Parameter index by .csv column names ==
            // 0: #name; 1: x; 2: y; 3: alt; 4: heading; 5: pitch; 6: roll; 7:f (focal);
            // == Non .csv param ==
            // 8: bool isRadianFormat default false
            this.createCamera(col[0], parseFloat(col[1]), parseFloat(col[2]), parseFloat(col[3]), parseFloat(col[4]), parseFloat(col[5]), parseFloat(col[6]), parseFloat(col[7]));
          }
        }
        resolve(true);
      });
    });
  }

Below you will find the code snippet for instantiating a camera with its position and rotation. I left some additional comments to elaborate it somewhat more. I left the commented code lines in as well to see what else we have been trying:

private createCamera(fileName: string, xPos: number, yPos: number, zPos: number, xDeg: number, yDeg: number, zDeg: number, f: number, isRadianFormat = false) : void {
    // Set radials as THREE.JS explicitly only works in radians
    const xRad = isRadianFormat ? xDeg : THREE.MathUtils.degToRad(xDeg);
    const yRad = isRadianFormat ? yDeg : THREE.MathUtils.degToRad(yDeg)
    const zRad = isRadianFormat ? zDeg : THREE.MathUtils.degToRad(zDeg)

    // Create camera reference and extract frustum
    // Statically set the FOV and aspectratio; Near is set to 0,1 by default and Far is dynamically set whenever a point is clicked in a 3D space.
    const camera = new THREE.PerspectiveCamera(67, 5280 / 2970, 0.1, 1); 
    const pos = new THREE.Vector3(xPos, yPos, zPos); // Reality capture z = up; THREE y = up;

    /* ===
    In order to set an Euler angle one must provide the heading (x), pitch (y) and roll(z) as well as the order (variable four 'XYZ') in which the rotations will be applied 
    As a last resort we even tried switching the x,y and zRad variables as well as switching the orientation orders.
    Possible orders:
     XYZ 
     XZY
     YZX
     YXZ
     ZYX
     ZXY
       === */
    const rot = new THREE.Euler(xRad, yRad, zRad, 'XYZ');
    //camera.setRotationFromAxisAngle(new THREE.Vector3(0,))

    //camera.applyMatrix4(new THREE.Matrix4().makeScale(-1, 1, 1));
    // const rot = new THREE.Quaternion();
    // rot.setFromAxisAngle(new THREE.Vector3(1, 0, 0), zRad);
    // rot.setFromAxisAngle(new THREE.Vector3(0, 1, 0), xRad);
    // rot.setFromAxisAngle(new THREE.Vector3(0, 0, 1), yRad);
    // XYZ

    // === Update camera frustum ===
    camera.position.copy(pos);
    // camera.applyQuaternion(rot);
    camera.rotation.copy(rot);
    camera.setRotationFromEuler(rot);
    camera.updateProjectionMatrix(); // TODO: Assert whether projection update is required here
    /* ===
    The camera.applyMatrix listed below was an attempt in rotating several aspects of the 3D viewer.
    An attempt was made to rotate each individual photo camera position, the pointcloud itself aswell as the viewport camera both separately
    as well as solo. It made no difference however.
       === */
    //camera.applyMatrix4(new THREE.Matrix4().makeScale(-1, 1, 1));

    // Instantiate CameraPosition instance and push to array
    const photo: PhotoPosition = {
      file: fileName,
      camera,
      position: pos,
      rotation: rot,
      focal: f,
      width: 5120,  // Statically set for now
      height: 5120, // Statically set for now
    };

    this.photos.push(photo);
  }

The cameras created in the snippet above are then grabbed by the next piece of code which passes the cameras to the camera manager and draws a CameraHelper (displayed in both 3D viewer pictures above). It is written within an async function awaiting the csv file to be loaded before proceeding to initialize the cameras.

private initializeCameraPoses(url: string, csvLoader: CSVLoader) {
    const absoluteUrl = url + '\\references.csv';

    (async (scene, csvLoader, url, renderer) => {
      await csvLoader.init(url);
      const photos = csvLoader.getPhotos(); // The cameras created by the createCamera() method
      this.inspectionRenderer = new InspectionRenderer(scene);  // InspectionRenderer manages all further camera operations
      this.inspectionRenderer.populateCameras(photos);
      for (const photoData of photos) {
        // Draw the green cube
        const geometry = new THREE.BoxGeometry(0.5, 0.5, 0.5);
        const material = new THREE.MeshBasicMaterial({ color: 0x00ff00 });
        const cube = new THREE.Mesh(geometry, material);
        scene.add(cube);

        cube.position.copy(photoData.position);
        photoData.camera.updateProjectionMatrix();
        
        // Draws the yellow camera viewport to the scene
        const helper = new CameraHelper(photoData.camera);
        renderer.render(scene, photoData.camera);    
        scene.add(helper);
      }
    })(this.scene, csvLoader, absoluteUrl, this.renderer);
  }

Marquizzo's code snippet

The below posted code snippet of Marquizzo seems to bring us a lot closer towards a solution. The cameras seem to be orientated in the correct direction. However the pitch seems to a little off somehow. Below I will include the source image of DJI_0189.jpg. Note that for this example the FOV is currently not being set as it looks chaotic when for every camera position a camera helper is being rendered. For this example I have rendered only the DJI_0189 camera helper. Source IMG DJI_0189.jpg

The edit @Marquizzo provided inverting the pitch (const rotX = deg2rad(photo.pitch * -1);) would result in the midpoint intersection point always being slightly lower as expected: enter image description here

When the pitch is adjusted to const rotX = deg2rad(photo.pitch * -.5); you'll see that the midpoint intersection is (closer) as that of the source image: Moreover aligned

Somehow I think that a solution is within reach and that in the end it'll come down to some very small detail that has been overlooked. I'm really looking forward into seeing a reply. If something is still unclear please say so and I'll provide the necessary details if required ^^

Thanks for reading this post so far!

like image 444
DiabetiMark Avatar asked Dec 16 '21 14:12

DiabetiMark


1 Answers

At first glance, I see three possibilities:

  • It's hard to see where the issue is without showing how you're using the createCamera() method. You could be swapping pitch with heading or something like that. In Three.js, heading is rotation around the Y-axis, pitch around X-axis, and roll around Z-axis.

  • Secondly, do you know in what order the heading, pitch, roll measurements were taken by your sensor? That will affect the way in which you initiate your THREE.Euler(xRad, yRad, zRad, 'XYZ'), since the order in which to apply rotations could also be 'YZX', 'ZXY', 'XZY', 'YXZ' or 'ZYX'.

  • Finally, you have to think "What does heading: 0 mean to the sensor?" It could mean different things between real-world and Three.js coordinate system. A camera with no rotation in Three.js is looking straight down towards -Z axis, but your sensor might have it pointing towards +Z, or +X, etc.

Edit:

I added a demo below, I think this is what you needed from the screenshots. Notice I multiplied pitch * -1 so the cameras "Look down", and added +180 to the heading so they're pointing in the right... heading.

const DATA = [
{name: "DJI_0174.JPG",  x: 3.116820957,     y: -44.25690188,    alt: 14.05258109,   heading: -26.86297007,  pitch: 66.43104338, roll: 1.912026354},
{name: "DJI_0175.JPG",  x: -5.22E-02,       y: -46.97266554,    alt: 14.18056658,   heading: -16.2033133,   pitch: 66.11532302, roll: 3.552072396},
{name: "DJI_0176.JPG",  x: -3.056586953,    y: -49.00754998,    alt: 14.3474763,    heading: 4.270483155,   pitch: 65.35247679, roll: 5.816970677},
{name: "DJI_0177.JPG",  x: -6.909437337,    y: -50.15910066,    alt: 14.38391206,   heading: 19.4459053,    pitch: 64.26828897, roll: 6.685020944},
{name: "DJI_0178.JPG",  x: -11.23696688,    y: -50.36025313,    alt: 14.56924433,   heading: 19.19192622,   pitch: 64.40188316, roll: 6.265995184},
{name: "DJI_0179.JPG",  x: -16.04060554,    y: -49.92320365,    alt: 14.69721478,   heading: 19.39979452,   pitch: 64.85507307, roll: 6.224929846},
{name: "DJI_0180.JPG",  x: -20.95614556,    y: -49.22915437,    alt: 14.92273203,   heading: 20.39327092,   pitch: 65.02028543, roll: 6.164031482},
{name: "DJI_0181.JPG",  x: -25.9335097,     y: -48.45330177,    alt: 15.37330388,   heading: 34.24388008,   pitch: 64.82707628, roll: 6.979877709},
{name: "DJI_0182.JPG",  x: -30.40507957,    y: -47.21269946,    alt: 15.67804925,   heading: 49.98858409,   pitch: 64.29238807, roll: 7.449650513},
{name: "DJI_0183.JPG",  x: -34.64277285,    y: -44.84034207,    alt: 15.89229254,   heading: 65.84203906,   pitch: 62.9109777,  roll: 7.065942792},
{name: "DJI_0184.JPG",  x: -39.17179024,    y: -40.22577764,    alt: 16.28164396,   heading: 65.53938063,   pitch: 63.2592604,  roll: 6.676581293},
{name: "DJI_0185.JPG",  x: -43.549378,      y: -33.09364534,    alt: 16.64130671,   heading: 68.61427166,   pitch: 63.15205908, roll: 6.258411625},
{name: "DJI_0186.JPG",  x: -46.5381556,     y: -24.2992233,     alt: 17.2286956,    heading: 74.42382577,   pitch: 63.75110346, roll: 6.279208736},
{name: "DJI_0187.JPG",  x: -48.18737751,    y: -14.67333218,    alt: 17.85446854,   heading: 79.54477952,   pitch: 63.0503902,  roll: 5.980759013},
{name: "DJI_0188.JPG",  x: -48.48581505,    y: -13.79840485,    alt: 17.84756621,   heading: 93.43316271,   pitch: 61.87561678, roll: 5.110113503},
{name: "DJI_0189.JPG",  x: -48.32815991,    y: -13.88055437,    alt: 17.77818573,   heading: 106.3277582,   pitch: 60.87171036, roll: 4.039469869},
];

const scene = new THREE.Scene();
const camera = new THREE.PerspectiveCamera(
    45,
    window.innerWidth / window.innerHeight,
    1,
    1000
);
camera.position.z = 100;

const renderer = new THREE.WebGLRenderer({
    antialias: true,
    canvas: document.querySelector("#canvas")
});
renderer.setSize(window.innerWidth, window.innerHeight);
const controls = new THREE.OrbitControls( camera, renderer.domElement );

// Helpers
const axesHelper = new THREE.AxesHelper( 20 );
scene.add(axesHelper);
const plane = new THREE.Plane( new THREE.Vector3( 0, 1, 0 ), 0 );
const planeHelper = new THREE.PlaneHelper( plane, 50, 0xffff00 );
scene.add(planeHelper);

let deg2rad = THREE.MathUtils.degToRad;

function createCam(photo) {
    let tempCam = new THREE.PerspectiveCamera(10, 2.0, 1, 30);
    // Altitude is actually y-axis,
    // "y" is actually z-axis 
    tempCam.position.set(photo.x, photo.alt, photo.y);

  // Modify pitch & heading so it matches Three.js coordinates
    const rotX = deg2rad(photo.pitch * -1);
    const rotY = deg2rad(photo.heading + 180);
    const rotZ = deg2rad(photo.roll);
  
    tempCam.rotation.set(rotX, rotY, rotZ, "YXZ");
    let helper = new THREE.CameraHelper(tempCam);
  scene.add(tempCam);
    scene.add(helper);
}

for(let i = 0; i < DATA.length; i++) {
    createCam(DATA[i]);
}

function animate() {
  renderer.render(scene, camera);
  requestAnimationFrame(animate);
}

animate();
html, body { margin:0; padding:0;}
<script src="https://cdnjs.cloudflare.com/ajax/libs/three.js/r128/three.min.js"></script><script src="https://rawgit.com/mrdoob/three.js/dev/examples/js/controls/OrbitControls.js"></script>

<canvas id="canvas"></canvas>
like image 71
Marquizzo Avatar answered Oct 05 '22 12:10

Marquizzo