Skip to content

Releases: MARDAScience/XBD-hurricanes

"Noisy" Retinanet for building and damaged building detection - best model checkpoints, July 2021

22 Jul 21:52
Compare
Choose a tag to compare

"Noisy" Retinanet for building and damaged building detection - best model checkpoints, July 2021

Two zipped folders containing model checkpoints (trained weights) for two instances of a custom "noisy" Retinanet for 1) building detection, and 2) damaged building detection

The building damage model is damage-scratch-val0.9-gamma5-alpha0.025-sigma0.3-modelcheckpoint.zip: it used 90% of the data for validation, a gamma parameter of 5 (note this is high compared to the original Retinanet paper recommendations), an alpha of 0.025, and a 'sigma' (noise parameter, not in the original RetinaNet implementation) of 0.3

The building detector is scratch_val0.7_gamma4_alpha0.2_modelweights.zip

Both models trained on all 4 datasets (hurricanes Harvey, Michael, Matthew, and Florence)

  • Florence: 3060 images and labels
  • Harvey: 2880 images and labels
  • Matthew: 2180 images and labels
  • Michael: 3180 images and labels
  • 11,300 images and labels in total