Katerina Fragkiadaki, Jianbo Shi Figure-Ground Image Segmentation Helps Semi-Supervised Learning of Objects. ECCV. September 2010.

Abstract : Given a collection of images containing a common object, we seek to learn a model for the object without using bounding boxes or segmentation masks. We separate image topics into foreground and background. Most salient image parts are likely to capture image foreground. We propose a novel probabilistic model, we call shape and figure-ground aware model (sFGmodel) that exploits figure-ground organization in each image separately, as well as feature re-occurrence across images, to find the common object. We use an image dependent topic prior and optimize a conditional likelihood of the image collection given the image bottom-up saliency information. Our framework can tolerate larger intraclass variability with fewer training data. We present results of our approach on diverse datasets showing great improvement over generative probabilistic models.

Katerina Fragkiadaki, Jianbo Shi Detection Free Tracking: Exploiting Motion and Topology for Tracking and Segmenting Under Entanglement CVPR. June 2011.

Abstract : Our goal is to segment multiple interacting and deforming agents in a video. Detectors often fail under large body deformation or agent entanglement. Grouping based on motion similarity fails to separate similarly moving agents or to group dinstinctly moving body parts. Our algorithms exploits the topology of foreground frame maps along with motion information of large temporal context to judge if two parts should be grouped together or not. We set video segmentation as a graph clustering problem. Our nodes are dense pixel trajectories. We have 2 types of edge weights: attractive weights between similarly moving trajectories, and repulsive weights between trajectories that belong to different connected components in the foreground map of any frame. We obtain our segmentation by computing spectral clustering (normalized cut) on the trajectory graph. We introduce FigMent, a challenging dataset containing scenes from basketball games with close agent interaction. We present state of the art segmentation results on Figment as well as previously established motion segmentation datasets.