3D tracking with one file per timepoint

Hello! I’m trying to perform a 3D tracking of a lightsheet dataset of a developing Xenopus embryo. I have the data in hdf5 format, in separate z-stacks for each timepoint. I’m running the pixel classification pipeline first, loading all the files in the data input applet, and this works well. The problem is that when I try to load the files in the tracking pipeline, I get an error saying the following: “For tracking, the dataset must have a time axis with at least 2 images. Please load a time-series data instead.”. I was wondering if loading a t-z-x-y stack is the only way to perform tracking in 3D, or if there’s any way to load the timepoints separately. I have big (> 500 Gb) datasets, and I think that scrolling through a single file with all the data will probably kill my computer. What is the correct way to approach this problem in ilastik when working with big datasets?

Thanks a lot for the help!

Hi @bmoretti,

for tracking to work there is no way around loading a time series.
in principle you can stack your individual files (those are 3D already, right?) across time if you choose add single 3D/4D volume from stack, as described here in our docs.

Let us know if the tracking works then, or if you run into other problems. Then we might be able to help with some workarounds.


Hi @k-dominik, thank you for your reply! I tried stacking the individual z-stacks across time but it takes a few hours just to load the raw data and prediction probabilities. And when I try to apply the threshold and size filters I get an out of memory error. The final shape of the stacked array is 60x116x2560x2160 (t-z-y-x). Maybe a better way to approach this would be to use a subset of the data (cropping in all four dimensions) to train the object and division classifiers, then create a single t-z-x-y hdf5 file with the whole timelapse and running the batch processing on it. Do you think this could work?



Hi @bmoretti,

sorry for responding so late.

sure this could work.

Another thing to consider would be spacial resolution. Do you think that the full resolution is really necessary for the tracking? Subsampling here would speed up all parts of the pipeline.

1 Like

+1. We had this exact issue doing 4D tracking (albeit with Trackmate / Imaris) with cells in spheroids (~750GB datasets) and as @k-dominik suggests, downsampling was the solution (and didn’t affect spot localisation precision significatly). Details in the preprint and repo if you’re interested.

1 Like