Closed jwindhager closed 2 years ago
Just to keep track of this. I have now compared the effect of downscaling the probabilities before segmentation. For this I have compared the segmentation masks that were generated (i) after downscaling of the probabilities and (ii) before downscaling and resizing of objects. Most cells (90%) show an overlap of 80%. 30 out of 45000 cells were lost due to first downscaling. So I think this is fine and can be integrated into the pipeline. @jwindhager is there an easy way to speed up the writing of PNG files in the preprocessing steps?
I have now reduced the compression level when writing png files using imageio
. This results in reasonably well compressed PNGs (~40mb per panorama) and speeds things up from around 6-7 minutes per MCD to a few seconds. @jwindhager if that's alright for you I think we can release v3.
I have now reduced the compression level when writing png files using
imageio
. This results in reasonably well compressed PNGs (~40mb per panorama) and speeds things up from around 6-7 minutes per MCD to a few seconds. @jwindhager if that's alright for you I think we can release v3.
Thanks! Any reason you didn't disable compression at all (i.e., 0
)?
Thanks! Any reason you didn't disable compression at all (i.e., 0)?
Yes, no compression (0) would mean a file size of 250mb instead of 40mb (compression level 1). Increasing the compression level further has little effect on the file size
Great, happy to release!
Perfect! I'll do it tomorrow to directly check if everything is usable and to upload the new results to zenodo
See #85