Created by Alexander Kolesnikov and Christoph Lampert at IST Austria.

We propose a new composite loss function for training convolutional neural networks for the task of weakly-supervised image segmentation. Our approach relies on the following three insights:

1. Image classification neural networks can be used to generate reliable object localization cues (seeds), but fail to predict the exact spatial extent of the objects. We incorporate this aspect by using a seeding loss that encourages a segmentation network to match localization cues but that is agnostic about the rest of the image.

2. To train a segmentation network from per-image annotation, a global pooling layer can be used that aggregates segmentation masks into image-level label scores. The choice of this layer has large impact on the quality of segmenta- tions. For example, max-pooling tends to underestimate the size of objects while average-pooling tends to overestimate it. We propose a global weighted rank pooling that is leveraged by expansion loss to expand the object seeds to regions of a reasonable size. It generalizes max-pooling and average pooling and outperforms them in our empirical study.

3. Networks trained from image-level labels rarely capture the precise bound- aries of objects in an image. Postprocessing by fully-connected conditional random fields (CRF) at test time is often insucient to overcome this ect, because once the networks have been trained they tend to be confident even about misclassified regions. We propose a new constrain-to-boundary loss that alleviates the problem of imprecise boundaries already at train- ing time. It strives to constrain predicted segmentation masks to respect low-level image information, in particular object boundaries.

If you find this code useful in your research, please consider citing us:

    @inproceedings{kolesnikov2016seed,
      title={Seed, Expand and Constrain: Three Principles for Weakly-Supervised Image Segmentation},
      author={Kolesnikov, Alexander and Lampert, Christoph H.},
      booktitle={European Conference on Computer Vision ({ECCV})},
      year={2016},
      organization={Springer}
    }

• Python packages:

      $ pip install -r python-dependencies.txt

• caffe: installation instructions are available at https://caffe.berkeleyvision.org/installation.html. Note, you need to compile caffe with python wrapper and support for python layers.

• Fully connected CRF wrapper (requires the Eigen3 package).

      $ pip install CRF/

In order to add new loss layers to caffe install the pylayers package, provided in this repository:

      $ pip install pylayers/

• Go into the training directory:

      $ cd training

• Download the initial model (~80 MB) pretrained on Imagenet:

      $ wget https://www.cs.jhu.edu/~alanlab/ccvl/init_models/vgg16_20M.caffemodel

• Decompress localization cues:

      $ gzip -kd localization_cues/localization_cues.pickle.gz

• Set root_folder parameter in train.prototxt to the directory with PASCAL VOC 2012 images

• Run caffe:

      $ caffe train --solver solver.prototxt --weights vgg16_20M.caffemodel --gpu <gpu_id>

The trained model will be created in training/models

• Go into the deploy directory:

      $ cd deploy

• Download pretrained SEC model (~80 MB):

      $ wget https://pub.ist.ac.at/~akolesnikov/SEC-data/SEC.caffemodel

• Run the model on any image (smooth options switches on CRF postprocessing):

      $ python demo.py --model SEC.caffemodel --image <PATH TO IMAGE> --smooth

• Some segmentation examples: