Object Recognition Using Distinctive Image Feature From Scale-Invariant Key point D. Lowe, IJCV 2004 Presenting – Anat Kaspi

The problem Reliable object recognition in the presence of clutter and occlusion Find a reliable matching between different views of an object or scene

Approach The paper combines several robust approaches to create a powerful recognition system. The basic stages include: Key point detection SIFT – Scale Invariant Feature Transform Clustering matching with Hough Transform

Previous Approaches The related research Harris corner detector (1992) (compare with key point detection) Schmid and Mohr (1997) (compare with SIFT) Disadvantage very sensitive to changes in scale

The SIFT algorithm Scale space extrema detection - Identify potential interest points that are invariant to scale and orientation using Gaussian function Key point localization – perform a detailed fit to the nearby data of each key point for location, scale and curvature. Some initial key points are rejected

The SIFT Algorithm Orientation assignment – One or more orientation are assign to each key point location based on local image gradient direction Key point descriptor – compute a descriptor for the local image region that is highly distinctive

Advantage of SIFT Distinctiveness Key points which enable correct matching from a large database Large number of key points with near real time performance on standard PC Invariant to image rotation, scale, affine distortion, noise, illumination

Applications Place recognition Robot localization and mapping in unknown environment

Plan Replace key point detection with some available interest point detection, e.g., Harris corner detection (1 week) Implement the heart of the algorithm – the key point descriptor (2 weeks) Thanksgiving Use graph matching algorithm to match to images (1 week) Testing and improving (2 weeks)