DP1 Deblended Object Footprints

Support
,
#1

Hello,

Are DP1 object footprints that are created by the deblender accessible? I would like to access individual object footprints to use as segmentation masks. I am using a neural network that will utilize the masks to help with other tasks. Previously, I could access these in DP0.2 with the deepCoadd_deblendedFluxcatalogs, or the meas catalogs. It appears that DP1 does not have these, but I did find object_scarlet_models accessible through the butler. Can these be used to recreate scarlet models, or at least get footprints?

Thanks,
Grant

#2

Hi Grant – thanks a lot for the question. For DP1, the footprints are not included in the data release due to storage volume considerations. I’m double checking internally about using object_scarlet_models to recreate the DP1 object footprints. Thanks again…

1 Like
#3

Hi Grant,
Yes, you can reconstruct the full scarlet model for a blend from the object_scarlet_models structure. Unfortunately for DP1 you have to load all of the models into memory and this takes up a lot of time (and memory). In DP2 and future data releases you will be able to load a single model as the code is already implemented in main.

I don’t remember the exact syntax for the version of the code that was used in DP1 but it will be something like

scarlet_model_data = butler.get("object_scarlet_models", **dataId)
model_psf = scarlet_model_data.psf
blend_data = scarlet_model_data.blends[blendId]
blend = blend_data.minimal_data_to_blend(model_psf, dtype: np.float32)

Just be warned that the exact syntax will change slightly in future data releases, but the blend given is the scarlet lite model.

#4

Hi Fred, thanks for your help. Trying the code snippet you sent, I get this error after the last line:

"ValueError: Both images must have the same number of axes, got 3 and 2"

I think this is due to the scarlet_model_data.psf being a 2D array, whereas the model data is 3D. I tried using the blend_data.psf object instead, which does work, but I’m not sure this is the right thing to do. An example image I get from running

import lsst.scarlet.lite as sl
norm = sl.display.AsinhMapping(minimum=0, stretch=1, Q=20)
model = blend.get_model(convolve=True)
rgb = sl.display.img_to_rgb(model, norm=norm)
plt.imshow(rgb, origin=‘lower’)
plt.axis(‘off’)
plt.show()

appears a bit odd (not properly convolved?)
Capture

Thanks for your help on this!