Resolution Control for Size Bias Elimination in Multi-resolution Visual Matching

Abstract

Visual matching for tracking and recognition, for example in video indexing, often uses image features measured at multiple resolutions. As a tracked object moves away from the camera, appearing progressively smaller, the higher resolutions consecutively become unavailable for matching, causing step changes in the similarity or “match score” of the tracked object. If several candidate matches (hypotheses) are maintained for a tracked region, this effect causes a bias toward larger region hypotheses that match at one extra resolution relative to even slightly smaller hypotheses. The effect is subtle and appears intermittent because it occurs only around a specific discrete set of object sizes. We describe the problem and the class of visual matching methods that it affects, and propose a solution. We present experimental results from a real video indexing system to illustrate both the problem and the effectiveness of the proposed solution.