• Project
  • PyTorch Replication of Deep Seeded Region Growing (DSRG)

PyTorch Replication of Deep Seeded Region Growing (DSRG)

This is the coder replcation project for DSRG. The DSRG is a weakly-supervised semantic segmentation method, where the original publication has been accepted by CVPR in 2018. [Codes]

project1_overview

Details

  • A Semantic segmentaion is a task of per-pixel classification of an image.
  • A weakly-supervsied learning means to give indirect supervision, e.g., Bounding boxes, Scribbles, Points, Image tags.
  • A cam method is used in this method to generate "seed cues" as initial supervision for a deep segmentaion model.
  • In DSRG, the seeded region growing mechanism is utilized to let the network generate new seeds in training process.

    •  



  • This PyTorch replication enables user to train DSRG network by using multiple GPUs.
  • It also provides a flexibility for using any segmentation dataset, other than VOC 2012.
  • Users can track and visualize few metrics, e.g., training loss, and few sample images, by connecting to your local tensorboard
  • Users can also use our scripts file for evaluating the network's performance, and obtain the mIoU result directly.


    • Results

    • Portfolio name

      Home Page

    • Portfolio name

      Face Detection

    References

  • DSRG conference Paper (CVPR2018) [1]
  • DSRG offical implementation (Caffe framework) [2]
  • DSRG reimplementation (TensorFlow) [3]
  • SEC reimplementation (PyTorch) [4]
  • Semantic segmentaion [5]

    • See other projects