Event Location: NSH 1507
Bio: Katerina Fragkiadaki is a Phd student in University of Pennsylvania, interested in video segmentation and tracking. She is currently exploring ways to resolve local in time ambiguities in segmentation and recognition by letting objects move long enough. She has a Bachelor degree in Electrical and Computer Engineering from National Technical University of Athens and a Masters in Computer and Information Science from University of Pennsylvania.

Abstract: We will explore ways of exploiting large temporal context in video sequences
for resolving local in time ambiguities for video segmentation. Point trajec-
tories as multi-frame point correspondences, carry useful motion information,
accumulated over long time horizon. However, under close object interaction
and deformation, motion is not sufficient to distinguish object separation versus
single object deformation. We will introduce object connectedness constraints
that complement motion information for segmentation. We leverage connect-
edness constraints from per frame figure-ground maps, built based on a notion
of ’trajectory saliency’. We cast video segmentation as partitioning of point
trajectories, with attractive weights between similarly moving trajectories and
repulsive weights between trajectories violating object connectedness. We will
introduce a density discontinuity detector acting on the normalized cut tra-
jectory embedding for detecting sudden drops of embedding affinities, strong
indication of motion boundaries.

Finally, we will combine trajectory clustering based on motion and 3D dis-
parity similarity, with object detection. We will formulate pedestrian detection,
tracking and segmentation as a grouping problem in the joint space of detec-
tions and point trajectories, seeking non- accidental grouping alignment in the
two spaces. We introduce a graph mediate process to resolve contradictions
between model driven and bottom-up cues in a confidence prioritisation way.
We show our model can track accurately through sparse, inaccurate detections
and under persistent partial occlusions, outperforming detection based trackers,
while discovering accurate segmentation support for the targets.