ustczhouyu / contour

Learning Panoptic Segmentation from Instance Contours

Learning Panoptic Segmentation from Instance Contours

TODO:

• Functionalities

  • Add output directory in configs. [Done]
  • Add res5 supervision to hed. [Done]
  • Add Contour Head branching from FPN Block. [Done]
  • Add Dice to contour losses. [Done]
  • Add Cityscapes panoptic evaluator. [Done]
  • Refine instance using center regression. [Done]
  • Add demo scripts that includes visualization for : [TBD]
    • semantic segmentation
    • contour estimation
    • instance segmentation
    • panoptic segmentation
  • Add loss mask. Dilated with 5 pixels [Sep-15]

• Nice to have

  • Add NMS loss.
  • Add Contour performance evaluator.
  • Add visualization for contours.

• Tests

  • Verify contour predictions. [Done] [Binary-only]
  • Verify Instance segmentation output. [Done]
  • Generate Videos. [Sep-20]

• Experiments

  • Cityscapes
    • Loss_fn Ablation. (4 experiments)
      • Binary Contours R_50_1x
        • Weighted BCE [Done]
        • Weighted BCE + Huber [Done]
        • Weighted BCE + Robust Loss [Sep-20]
        • Weighted BCE + NMS Loss [Sep-20]
    • GT Dilation Ablation with best loss_fn. (4 experiments)
      • Binay Contours R_50_1x
        • Dilation factor 1 [Sep-14]
        • Dilation factor 2 [Done]
        • Dilation factor 3 [Sepe-15]
    • Network Ablation with best dilation factor and best loss_fn. (3 experiments)
      • Binary Contours R_50_1x [Sep-10]
        • Seperate heads [Done]
        • Seperate Necks
      • Multi-class Contours R_50_1x [Sep-10]
        • Seperate heads [Done]
        • Seperate Necks
    • Postproc ignore pixel area Ablation. [Eval-only] [Sep-11]
      • Pixel Area 0
      • Pixel Area 100
      • Pixel Area 300
      • Pixel Area 500
    • Best Network with R_101_3x against SOTA [Sep-11]