Implemented Canny Edge Detection on RGB images with normal and low light environments. The same high and low magnitude thresholds for hysteresis edge linking were used for all images. In the images above, the left image is the original image and the right image is the Canny Edge Detection result.
Performed Gradient Domain Blending on images using the gradient domain processing technique described in the "Seamless Cloning" section of the Poisson Image Editing paper by Patrick Perez, Michel Gangnet, and Andrew Blake.
Example of Face to Face Image Morphing. It is a morphing of Iain Armitage into Jim Parsons. These actors play Sheldon Cooper in Young Sheldon and in The Big Bang Theory, respectively.
Example of Face to Object Image Morphing. It is a morphing of Dr. Kent Fuchs, president of the University of Florida, and Albert E. Gator, University of Florida's mascot.
Example Object to Object Image Morphing. It is a morphing of the Hudson Hornet into Doc Hudson, a character in the movie "Cars".
Conducted Seam Carving on images using the methods described in the Seam Carving for Content-Aware Image Resizing paper by Shai Avidan and Ariel Shamir. In the images below, the left image is the original image and the right image represents the Seam Carving result. The videos below show the seam carving animations.
Performed Image Stitching on images from around University of Pennsylvania's campus. Conducted feature detection using the Harris Corner Detector and the number of feature points was reduced using Adaptive Non-Maximal Suppression (ANMS). Random Sampling Consensus (RANSAC) was used to match features between two images and estimate the homography between these images.
Note: Project was conducted with Brian Barrows and Zachary Fisher.
Franklin Field Mosaic
Levine Hall Bump Space
Applied Object Tracking to videos to track up to three objects at the same time. Performed feature detection using the Shi-Tomasi Corner Detector and feature tracking used the Kanade-Lucas-Tomasi tracking procedure (KLT tracker). In addition, the KLT tracker has iterative refinement, outliers were removed using anomalous flow, and the bounding box for the object scales dynamically. The bounding boxes are in red, the feature points are in blue, and the trajectory path for each bounding box is in yellow.
Note: Project was conducted with Brian Barrows and Zachary Fisher.