Adaptive combination of local motion, appearance, and shape for video segmentation.

Authors: Woo-Sung Shim, Se-Hoon Kim, Soochahn Lee

TL;DR: An adaptive method that combines local motion, appearance, and shape cues for robust video object segmentation.

What this paper is about

Segmenting objects in video requires handling moving objects, changing appearances, and deformable shapes all at once. Relying on any single cue (motion, appearance, or shape) alone is fragile, because each cue fails in different scenarios.

Key idea

Overview of the proposed method. We combine image gradient based geodesics, color, shape and motion segmentation cues to accurately propagate object and background segment probabilities to adjacent video frames.

The paper proposes an adaptive combination scheme that dynamically weights local motion, appearance, and shape features for video segmentation. Rather than using fixed weights, the method adjusts how much it relies on each cue based on local context, so it can lean on motion in areas with clear movement and on appearance or shape where motion is ambiguous.

Why it matters

This adaptive multi-cue approach produces more robust video segmentation across diverse and challenging scenarios than methods relying on any single cue.