Classical ANMS+RANSAC pipeline alongside a deep HomographyNet — packaged into a unified library for high-quality panorama stitching achieving ~1.5px RMSE.
This project implements and benchmarks two approaches to homography estimation for panoramic image stitching: a classical feature-based pipeline and a deep learning approach (HomographyNet). Both are packaged into a single library for flexible deployment.
Classical Pipeline: ANMS (Adaptive Non-Maximal Suppression) keypoint detection is followed by descriptor computation, ratio-test matching, and RANSAC-based homography estimation. Multi-image stitching uses sequential homography composition with blending.
HomographyNet: A CNN trained on MSCOCO-derived patch pairs learns to predict the homography directly from pixel intensity differences. The network is trained with an L2 loss on the four-point parameterization and augmented with synthetic perspective distortions.
Unified Library: Both approaches are wrapped in a common API that accepts image pairs or sequences and returns stitched panoramas, enabling easy comparison and deployment.
The classical ANMS+RANSAC pipeline achieved ~1.5px RMSE on standard test pairs. HomographyNet achieved ~2px error on real-world scene pairs, with more graceful failure modes in textureless regions where classical feature matching struggles.