[question] Question about camera.sfm poses

[simoneBrancati]

Describe the problem
After the completion of StructureFromMotion node a camera.sfm file is produced.
Inside this file all the views' instrics and extrinsics are listed. In particular, under the "poses" keyword we find the views' poses.
For example:

poses": [
        {
            "poseId": "300639444",
            "pose": {
                "transform": {
                    "rotation": [
                        "0.99995348941947282",
                        "-0.00057462984792917393",
                        "0.0096275021868713097",
                        "0.00061704033357770128",
                        "0.99999011748701916",
                        "-0.0044027479515035455",
                        "-0.0096248770925701308",
                        "0.004408483734301312",
                        "0.99994396193592638"
                    ],
                    "center": [
                        "-0.011947754013339834",
                        "0.010463514098481298",
                        "0.023576133686596487"
                    ]
                },
                "locked": "0"
            }
        }

Questions:

• Does rotation refers to a 3 x 3 rotation matrix and center to the translation vector.
• in what order i should put the elements of this rotation to create a rotation matrix (i.e. the OpenGL ModelView matrix?)
• what is the scale of this pose calculation (does meshroom approximates the scale of the real world object somehow)?

[simogasp]

Does rotation refers to a 3 x 3 rotation matrix and center to the translation vector.

The rotation is the 3x3 rotation and center is the camera center C in world coordinates. So a point Q in world coordinates projects into the camera as

q ~ K [R -RC] Q

in what order i should put the elements of this rotation to create a rotation matrix (i.e. the OpenGL ModelView matrix?)

That's a column-major Eigen matrix, so the first 3 elements are the first column of the matrix and so on. Incidentally, both OpenGL and Eigen use column-major format for storing matrices, so if you read and store the values in the that order in a raw buffer you can use it as a matrix in opengl.

what is the scale of this pose calculation (does meshroom approximates the scale of the real world object somehow)?

The scale is always arbitrary, reconstruction is always up to a scale factor. In order to bring it in the real world scale, you need to know some relationship (e.g. the distance between two points) and rescale the scene (points and cameras) accordingly.

[simoneBrancati]

Thanks a lot!
If i know the real object dimensions, is there any input parameter in the pipeline i can put in to adjust the scale?

[fabiencastan]

In the SfMTransform, you can use "auto_from_landmark" which will fit a box around your 3D points and rescale it to -scale to +scale. But if you have points reconstructed around your object, it will become completely wrong.

Another solution, is to put markers, and rescale your scene based on these markers, but that's in development here: #652.

[yzslab]

Does rotation refers to a 3 x 3 rotation matrix and center to the translation vector.

The rotation is the 3x3 rotation and center is the camera center C in world coordinates. So a point Q in world coordinates projects into the camera as

q ~ K [R -RC] Q

in what order i should put the elements of this rotation to create a rotation matrix (i.e. the OpenGL ModelView matrix?)

That's a column-major Eigen matrix, so the first 3 elements are the first column of the matrix and so on. Incidentally, both OpenGL and Eigen use column-major format for storing matrices, so if you read and store the values in the that order in a raw buffer you can use it as a matrix in opengl.

what is the scale of this pose calculation (does meshroom approximates the scale of the real world object somehow)?

The scale is always arbitrary, reconstruction is always up to a scale factor. In order to bring it in the real world scale, you need to know some relationship (e.g. the distance between two points) and rescale the scene (points and cameras) accordingly.

@simogasp Hi, would you mind telling me what is the meaning of the K in q ~ K [R -RC] Q ? Thanks.

[simogasp]

K is the 3x3 calibration matrix that contains the intrinsic parameters of the camera, like the focal length and the principal point.
https://en.wikipedia.org/wiki/Camera_resectioning#Intrinsic_parameters

[yzslab]

K is the 3x3 calibration matrix that contains the intrinsic parameters of the camera, like the focal length and the principal point. https://en.wikipedia.org/wiki/Camera_resectioning#Intrinsic_parameters

wow, thanks for your detailed explaination!

[rinki447]

is there any unit for the camera centre under poses? I mean the centre can be located 3mm or 3 meter away from world origin. How to know?