Computer Vision News - December 2018

is a penalty parameter; y is a close approximation of x. The above can be minimized by alternating steps updating either x or y iteratively. The two steps are temporal fusion and mask refinement. In addition Iterated Conditional Modes (ICM) is used to find an approximate solution of ( , ) . At each iteration, ICM updated one random variable while fixing the rest of the random variables (see TF section in Algorithm 1 below). Moreover, ( ) is used as an approximation of since solving it directly would be computationally difficult. Algorithm 1 summarizes the inference model developed by the authors together with all the approximation assumptions describe above: Implementation Details ● The gCNN() takes a 4-channel input (RGB image + preliminary mask), and outputs a refined mask. gCNN() is trained in two stages: (1) an offline model is trained using object segmentation data available, and (2) the offline model is fine-tuned using the ground truth mask in the first frame of a given video. Research 7 Research Computer Vision News