Project Spec: https://inst.eecs.berkeley.edu/~cs194-26/fa18/hw/proj6/
Input
Output
Input
Output
Input
Output: Left is naive overlaying, right is laplacian blending. Unfortunate artifacts due to mask :(
Input
Output
Input
Output
I started with two pictures of my fireplace:
Default arris corners
Harris corners after applying ANMS:
Feature matching using Lowe's technique:
Special thanks to Ajay Ramesh for code to visualize the feature pairs.
Original:
Stitched:
Original:
Stitched:
Original:
Stitched:
Some takeaways:
Original image quality is very important. I originally didn't pay much attention to exposure/shutter speed settings, so the images had different lighting. Not much post-processing can do to fix this.
Automatic feature matching is much cooler after laborious manual selection
I was very impressed at how well the feature descriptors and lowe reranking worked. I was particularly impressed at these feature correspondences: