I am trying to understand the LidarBoost algorithm as explained in this paper (PDF warning). My apologies if this isn't the appropriate exchange; I've posted in DSP, but I didn't get a response there, and the FAQ said that some image processing algorithms are acceptable here.

I don't understand how they take the original depth-images $Y_k$ and transform them into the up-sampled images $D_k$. I get how optical flow is used to align the $Y_k$ into a chosen reference frame, but is the transform between $Y_k$ and $D_k$ just a standard image up-sampling with nearest neighbors for the "interpolation" step? If that's the case, can someone explain how they get the term $W_k$ in their data term of the energy function? They say this about it, but I don't understand how to construct $W_k$ for each $k$:

$W_k \in \mathbb{R}^{\beta n x \beta m}$ is a banded matrix that encodes the positions of $D_k$ which one samples from during resampling on the high-resolution target grid

I'm trying (so far frustratingly ineffectively) to get enough of a sense of the algorithm to try to implement it.

  • $\begingroup$ I'm redoing some similar experiments recently and have met with a lot of questions too. To my understanding, the author applied the NN interpolation to x,y,z matrix respectively (you can find their raw data in this link: gvv.mpi-inf.mpg.de/files/old_site_files/tof/#upsampling). But I'm still confused about how to configure Wk. If you have some new ideas. I really appreciate to discuss with you. $\endgroup$ – ape May 22 '13 at 23:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.