First: about the FPS problem, when I tested the refusion pro, the FPS was normal 28-30, but I found that the FPS output from the DOS window was only 1 or 2 using the SDK, and during the reconstruction process, I found that the FPS delay was indeed very large, and only SDK found this problem. I don’t know whether it is the wrong output or the FPS is really low.
Second: about volume, I use the SDK sample test. After I calibrate, the default volume is tilted.
But this problem does not appear in the recfusion pro. I found that volume:transformation is the cause. But I didn’t see the SDK’s description of this function. Although I see a volume rotation in the SDK, it is a 3x3 matrix. It should be wrong. I think the SDK should have the ability to obtain transformation and pass in transformation as parameters.
There is nothing that can be done regarding occasional drops in fps. The new version of SDK just prints this information, while previous didn’t, but there was no change in stream handling, so the fps itself was the same. Please ignore these warnings, they will be removed from future versions.
Final volume might be tilted because of the reconstruction volume not being set up properly. Volume transformation is represented by position vector and a 3x3 rotation matrix. If you want to use angles, you can convert these to a rotation matrix
What I want to know is the volume position. Can I use 4x4 matrix information? Because I see that recfusion Pro uses 4x4 matrix information to process. I don’t see how volume “param name=” transformation is obtained? Is it available in SDK??? I directly use the volume “param name=” transformation generated by recfusion pro. Is it OK? Because the parameter generated by it is perfect, I can directly set it in the SDK without calculation. Please tell me how to do it.
The structure of the 4x4 transformation matrix from RecFusion is as follows:
r_xx r_xy r_xz t_x
r_yx r_yy r_yz t_y
r_zx r_zy r_zz t_z
0 0 0 1
where
r_xx r_xy r_xz
r_yx r_yy r_yz
r_zx r_zy r_zz
is the rotation matrix
and
t_x
t_y
t_z
is the translation vector.
You can use the latter two in RecFusion SDK
I have solved this problem. I executed the rotation matrix when saving the file.
The last problem is that when two people with different heights are reconstructed, the shorter one will have size error, while the higher one will have normal size. Please tell me what causes the problem. Such as the location of the sensor? What is the inclination of the camera? I don’t think it should be a calibration problem. I’m a single camera. If the calibration is a problem, how about the size of the high height?