On equation 8 of the paper.

[Eightfingers]

Hello!

Thank you for providing the code, I am in the midst of trying to recreate the algorithm on matlab to test it since I do not have an NVIDIA gpu. My question is what is the expected range of value of the map function M? is it a probability function where it is 0 to 1 ?
$M(v) = \frac{O(v)}{N(v)}$
I observed that the values of $M(v)$ that I obtained from recreating the algorithm are not in the range of 0 to 1. Instead it has an erratic range of 0 to ~ 3 depending on the current event packet. Is this correct? since $O(v)$ is just keep tracks the number of events in that pixel and $N(v)$ is simply a normalizing function thus $O(v)$ divide by $N(v)$ should not give a range of 0 to 1 .

[reini1305]

Hi! Thanks for trying to port this to the CPU, I'd be very interested in the end result!

In theory you're correct, but it seems that due to numerical inaccuracies this doesn't hold for all pixels. I see that I clamped the value to 1 in the cuda code, so it should be fine to do the same in your implementation.

[Eightfingers]

Okay thank you for the quick reply! I will just clamp the value to 1 so that M will have a range from 0 to 1. Right now i am in the midst of the figuring out and implementing the tracking portion (I was using the ground truth as orientation input for mapping). If I have any additional question do I just post it under this issue or do i open a new one?

[reini1305]

I'm fine with keeping this conversation open here