Documente Academic
Documente Profesional
Documente Cultură
Ivan Petrovic
To look up
QR Decomposition
Blur Lens
Sobel Operator
Jacobian Transposition
Actual Notes
𝑥 𝑥
Apparently equation (*) = 𝑢 = 𝑥1 , 𝑣 = 𝑥2.
3 3
Fun fact: Colour images are represented by a vector of values for each pixel~
If the intensity change is smooth we can expand the image intensity function to a Taylor series.
𝑓(𝑥 + 𝑑𝑥, 𝑦 + 𝑑𝑦, 𝑡 + 𝑑𝑡) = 𝑓(𝑥, 𝑦, 𝑡). <-Brightness constancy constraint; we assume the
brightness is constant and hence that it doesn’t modify colour values.
𝛿𝑓 𝛿𝑓 𝑑𝑥 𝛿𝑓 𝑑𝑦
Eventually through partial derivatives, you get − 𝛿𝑡 = 𝑥 𝑑𝑡
+ 𝑑𝑦 𝑑𝑡 (*). This is the optical flow
equation.
𝑑𝑥 𝑑𝑦
This gives constraints on the velocities (𝑢, 𝑣) = ( 𝑑𝑡 , 𝑑𝑡 ) dependent on the local derivations of the
𝑑𝑓 𝑑𝑓 𝑑𝑓
image intensity function: (𝐸𝑥 , 𝐸𝑦 , 𝐸𝑡 ) = (𝑑𝑥 , 𝑑𝑦 , 𝑑𝑡 ).
Equation (*) is not sufficient to compute optical flow (𝑢, 𝑣), but only the linear constraint (𝐸𝑥 𝑢 +
𝐸𝑦 𝑣 + 𝐸𝑡 = 0).
To compute distances using a single camera ‘orbiting’ around a scene, you need to have 5 point
correspondences. Slide states “possible to compute unknown R and T (up to a scale)”. (R,T) is the
annotation for the curved arrow indicating the ‘orbit’. Slide ‘section’ statement is “MR ego motion”,
‘subject’ is “Distance computation”.
Fun fact: Kinect has an RGB camera in the middle, and the two other apertures around it are an IR
projector and an IR camera; the projector emit the radiation and the camera picks up the reflected
light. I believe this is the first one, and after Apple bought the technology involved (or its provider?),
they had to change the design. Or this is the new one, I forget.