Beyond Gradient Descent for Regularized Segmentation Losses

Dmitrii Marin, Meng Tang, Ismail Ben Ayed and Yuri Boykov

Appears in IEEE conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, CA, 2019

If you find our work useful in your research please consider citing our paper:

@InProceedings{ADM:cvpr19,
  author = {Dmitrii Marin and Meng Tang and Ismail Ben Ayed and Yuri Boykov},
  title = {Beyond Gradient Descent for Regularized Segmentation Losses},
  booktitle = {IEEE conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2019},
  address = {Long Beach, California}
}

[pdf]

ScribbleSup Dataset

Download original PASCAL VOC 2012 dataset: http://host.robots.ox.ac.uk/pascal/VOC/

Download Scribble annotations: http://cs.uwaterloo.ca/~m62tang/rloss/pascal_2012_scribble.zip, the original dataset can be found here

Compilation

In deeplab/code/ rename Makefile.config.example into Makefile.config. Edit Makefile.config to set up the compilation. In particular, set USE_CUDNN := 1 to use CUDA and set CUDA_DIR to point to your CUDA instalation; adjust INCLUDE_DIRS and LIBRARY_DIRS to include libraries BOOST, BLAS, etc. See the dependecy list here. Run make.

Training and Testing

Update variable ROOT in deeplab/exper/run_pascal_scribble.sh to point to the ScribbleSup dataset.

First, train and test a base model (with partial cross entropy only):

cd deeplab/exper
DEV_ID=0 bash -x ./run_pascal_scribble.sh

The model will be saved in pascal_scribble/model/deeplab_largeFOV/train_iter_9000.caffemodel. The mIoU should be approximately 55.8%.

Then, add regularization to the loss and train/test the model using ADM:

TRAIN_CONFIG=GRID-ADM MODEL=pascal_scribble/model/deeplab_largeFOV/train_iter_9000.caffemodel DEV_ID=0 ./run_pascal_scribble.sh

This should give mIoU of approximately 61.7%