Low Contrast Panoramic Image Stitching

This project addresses the problem of constructing large-scale panoramic mosaics from low-contrast underwater imagery, where standard stitching pipelines frequently fail due to weak features and accumulated alignment drift. The objective was to achieve a globally consistent panorama rather than relying on locally accurate but globally inconsistent pairwise alignments.

I began by extracting SIFT features and estimating pairwise homographies between overlapping images. However, small alignment errors quickly accumulated, resulting in severe drift as more images were added. To mitigate this, I modeled the panorama as a pose graph and introduced loop closure constraints between non-consecutive but overlapping image pairs.

Using GTSAM, I performed global optimization over homography transformations, correcting drift and enforcing consistency across the entire image set. This approach enabled the construction of a seamless high-resolution mosaic from 29 low-contrast images. I further quantified reconstruction reliability by computing marginal covariance matrices, which provided insight into state uncertainty and ensured robust image placement in feature-poor environments.