search

mubeipeng / objectSLAM

75 stars 27 forks source link

About the posterior calculation #4

Closed XunshanMan closed 4 years ago

XunshanMan commented 4 years ago

Hi, I have a small question about the calculation of the posterior. In the paper, it contains a prior of Dirichlet Process, and the likelihood of object detection. In the code,

dp_prior = p(obj.measurements.label(k),:); sq_dist = sum((obj_poses - repmat(pos,1,N)).^2,1); posterior = exp(-4sq_dist).dp_prior; [prob, I]=max(posterior);

  1. I'm wondering why the dp_prior is an integer, beacause after several observations, the posterior will be far more than 1.
  2. the coefficient -4, is that mean -1/(0.5)^2? an assumption of the gaussian distribution of the object position, with a sigma of 0.5m?

I just read your paper. I'm wondering if I understood it right. Thanks for your help!

mubeipeng commented 4 years ago
  1. the original DP, prior is proportional to the number of observations. p(n) ~ n/(N+alpha), where N is the total number of observations so far, and alpha is a factor. Since (N+alpha) is a constant multiplier for all classes, it does not affect the max-posterior calculation, so the calculation could be simplied.
  2. that's probably right as far as I can recall
XunshanMan commented 4 years ago

Happy to hear your response so fast. Thanks, that really helps.