• Detectron2 with rcnn for foreground detection
  • Passing the bounding box to SAM model
  • Segment mask through controlnet in stable diffusion inpaint to fill the space
  • Used mixed precions training
  • Replaced detectron with yolo