Add deblending metric and performance plots

burke86 / astro_rcnn

Astro R-CNN: Instance Segmentation in Astronomical Images using Mask R-CNN Deep Learning

MIT License

43 stars 15 forks source link

Open burke86 opened 4 years ago

burke86 commented 4 years ago

The AP score is not a probably not the most sensitive metric for testing deblending. Should try others, for example the purity metric using a ratio of weighted sums: rho_i = sum G_i G_i / Sum G_i (G_1 + G_2 + ...) from https://project.lsst.org/meetings/lsst2018/sites/lsst.org.meetings.lsst2018/files/Blending%20Session%205%20%28Tools%20and%20Data%20Sets%29%20-%20Slides_0.pdf#page=13&zoom=100,0,0

burke86 commented 4 years ago

Also see pixel accuracy metric, section 5.1 of https://arxiv.org/pdf/2001.05566.pdf