Computer Vision Assignments - Notes and Results

Course taught by Prof. Leonid Sigal

Jump to: A1 | A2 | A3 | A4 | A5

Image Filtering and Hybrid Images

will add soon

Scaled Representations, Face Detection and Image Blending

will add soon

Detected two faces in a family photo of three

Face detection results using normalized cross-correlation at different scales of a Gaussian pyramid with a template face.

Gaussian pyramids
Reconstructing the Gaussian pyramid from the Laplacian pyramid
Image blending

Texture Synthesis

will add soon

RANSAC and Panorama Stitching

Implemented a complete panorama stitching pipeline using SIFT feature detection and RANSAC-based robust estimation. The system matches keypoints between image pairs using descriptor similarity and ratio testing, then applies geometric consistency filtering to eliminate outliers based on orientation and scale agreements. Robust homography matrices are estimated through iterative RANSAC sampling, followed by image warping and blending to create panoramas.

Parameter Effects: RANSAC iteration count balances homography accuracy with computational cost. Inlier tolerance threshold controls precision for keypoint projection - tighter tolerances improve alignment but may exclude valid matches, especially for lower resolution images. Ratio threshold determines initial match quality by filtering ambiguous correspondences.

Effect of varying ratio threshold on keypoint correspondences. No matches were found with a threshold less than 0.3.

Effect of geometric consistency filtering on keypoint correspondences. Orientation agreement (OA) in degrees, Scale agreement (SA) is a ratio.

Stereo and Optical Flow

will add soon