Deeplabv3+matte_exp

Deeplabv3+matte experiments：

• I transfered the semantic segmentation deeplabv3+ to matting task for high accuracy.
• Main modification:
  • Data processing including pre-processing and post-processing
  • Loss function which includes four parts: L_d loss, CE_loss, Gradients loss and Composition loss. References: SHM-ACMM2018 and Late-fusion-matting-CVPR2019
  • Optimizer data related
• Based projrct: pytorch-deeplab-xception.
• Training protocol:
  • Datasets: Deep Automatic Portrait Matting as portrait datasets and aisegment as aisegmet datasets.
  • Pre-trained on aisegment datasets for 50 epochs and refine on portrait datasets for 100 epochs. Note aisegment (34426 images) is more larger than portrait datasets (2000 images).
  • Others setting follow the based project with a little modification.

---

Pre-trained on aisegment (dev-look-ahead):

• Training checkpoint: epoch0-pth

• [todo] Quantitative metrics overview

• TensorBoard overview

  Loss:

  

  

  

  Example: