"point cloud" definition in fiducialreg

[linshaova]

Hi Talley,

Would a calibration 4-channel image like the following work for fiducialreg? In other words, how strict is the "point cloud" requirement? The calibration results I got from this dataset seem to not work very well, which made me think maybe the sample is up to standard. thanks!

-lin

[tlambert03]

hey @linshaova, it "should" work... but I'm sure there's a limit The point cloud registration algorithm is as in https://arxiv.org/abs/0905.2635, which is designed to be robust to missing data points in one or more of the datasets (i.e. you don't need a 1-to-1 match of detected points to get a good result).

It's a bit hard to tell from that image, i think what we really need to see are the pre-registered point clouds extracted from the raw data (based on bead detection). It's been a while since I've looked at this code, would you be able to put together a brief script showing showing how you're using it, along with the dataset somewhere?

[linshaova]

I didn't mean to ask you to debug this for me! Great reminder that CloudSet can be displayed and checked. It looks like if I use the mincount option, the number of found points varies greatly between channels. And strangely, I saw when one channel has >800 points and the other one has <50, count_matching somehow equals to almost 600! I had thought count_matching should be less than the fewer of the two channels?

In any case, I'm going to try specifying threshold when calling CloudSet and see if it improves anything.

[tlambert03]

yeah i do vaguely recall that visualizing the resulting cloudsets was important...

[linshaova]

After re-calibrate with manually setting threshold, instead of using mincount, in calling CloudSet(), the results look much better! Thanks!

[tlambert03]

great! yeah that aligns with my memory as well. if you hit on more robust ways to auto-pick the threshold, i'm sure we could do better there