05.07. Meeting Notes

[sudo-Boris]

3 Stage training

Train in parallel gt, sl, occ

• implement saving and loading of individual unet decoder branches
• When using SSIM for loss: 1 - SSIM_value (to avoid negative loss, because we want to maximise SSIM)
• implement losses
  • GT:
    • MSE(gt_original, gt_prediction)
  • SL:
    • MAE(light_generated, light_prediction)
    • MAE(shadow_generated, shadow_prediction)
  • OCC:
    • 1 - SSIM(gt_original * shadow_generated + light_generated XOR occ_predicted, input_image)
• prepare configs
  • GT:
    • MSE
  • SL:
    • MAE
  • OCC:
    • SSIM