Revisiting GrabCut

Abhijit Sarkar
Spring 2014 ECE 6504 Probabilistic Graphical Models: Class Project
Virginia Tech

 

Goal

Grabcut [1] is a well-established background foreground segmentation tool. In its original form, grabcut uses the RGB color mixture model for both background and foreground. We want to test the algorithm with additional image qualities, say color in Luv space and using Textons.

Approach

We follow the original method mentioned in the Grabcut paper, where, the data terms are modeled using a Gaussian mixture model of the RGB cues, each for background and foreground. Each GMM is taken to be a full covariance Gaussian mixture with K components. We take k=6. The pairwise energy term is a function of the Euclidian color distance and a regularization term . We have used the original 50 image dataset from Grabcut paper and varied  from 0.1 to 1.1 in a step of 0.25 to calculate the recall and precision for each image. An average precision and recall is reported for each case. The same technique is followed three different cases.

1.     In the Luv color space

2.     Using RGB + Texton

3.     Using Luv  + Texton

The textons are calculated from the matlab implementation of [2]. We generate a independent texton vocabulary for each image and use them to generate a texton histogram to assign in the data term of the model. A similar experiment is performed for all the three cases as mentioned above. We have used the grabcut code from [1].

Results

We have reported the average precision and average recall for each of the cases.

It can be noted that, The recall is not very much affected by different image cues. Although highest recall is achieved for  For precision, The original model performs better than other models. If we look at the precision recall values together, then we might conclude that, Use of RGB and Texton gives the best results.

Here are some of the results:

RGB + Textons

RGB + Textons

 

For the face, Luv gives best result where using only RGB gives artifacts