Tags: #notesonai Topics: [[Computer Vision]] --- # Scale-Invariant Feature Transform (SIFT) SIFT is the solution to the question: Can we detect the same interest points regardless of image scale changes? Basic idea: - Divide the $16 \times 16$ window into a $4 \times 4$ grid of cells $(2 \times 2$ case shown below) - Compute an orientation histogram for each cell (with 8 orientations) - 16 cells (from $4 \times 4$ grids) \* 8 orientations (per cell) =128 dimensional descriptor ![[Sift-laplacian.jpg]] Two popular scale invariant detectors: ![[sift-detectors.jpg]] ## Applications Feature points are used for: - Image alignment - 3D reconstruction - Motion tracking - Robot navigation Indexing and database retrieval - Object recognition --- ## References