AutoPano

Panomoric image stitching

Traditional approach and deep learning approach to estimate homography between 2 sets of images. Implemented supervised and unsupervised deep learning to estimate homography

Input Images

Corner Detection

Adaptive Non-Maximal Suppression

Feature Encoding

• We encode each feature point for matching. To obtain encodings, we take a patch of size 41x41 around the feature point, then apply gaussian blur and sub-sample a 8x8 patch from it. This patch is then flattened to a feature vector of size 64x1 and standardized to set mean=0 and variance=1

Feature Matching

RANSAC - outlier rejection

Warping and Blending

Results on some other data

• Image Stitching using traditional approach

• Homography estimation by deep learning approach

• MCE - Mean Corner Error