Closed sam575 closed 4 years ago
If you require instantaneous velocity (not a frame-to-frame transform), I advise you to fit the splines to motion data and take derivatives (numpy has this functionality) - otherwise you will suffer from noise (even Vicon has some!). That C++ code was not reliable in this sense, so I would not recommend you to use it.
For transforms some filtering might be required, depending on the amount of actual motion between frames - I suggest you use the pyquaternion and .npz format to compute inverses of transformations and then - transforms between frames in the coordinate frame of your choice.
Keep in mind, that for objects you can compute their velocities in their own instantaneous frame, or in the camera frame, or even in the world fixed frame. For camera, you can compute its velocity in the camera instantaneous frame or in the world frame. Which one to choose depends on the application, so we did not release velocity values with the dataset.
The velocity of objects was missing in the dataset provided, so I looked into the code and found that the following lines 164 and 269 are commented. As I notice a TODO there, I want to know if accurate velocity of objects can be generated by uncommenting these lines?