| Title | : | Auto-Correction of Perspective: Exploring blind Homography estimation for Image Refinement (eg Rectification application) |
| Speaker | : | Pooja Kumari (IITM) |
| Details | : | Wed, 29 Nov, 2023 11:00 AM @ SSB-334 |
| Abstract: | : | In computer vision, homography is a 2D projective transformation that maps the points in one plane to the corresponding points in another plane. It is commonly used to correct geometric distortions between two images of the same scene or object taken from different perspectives. Given two views of a scene, one can easily compute the homography matrix (H) using Direct Linear Transformation (DLT) or its variants. However, the problem of estimation of H has not been dealt with, when you have only a single image of an inclined planar texture surface. Can one estimate the homography in such blind scenarios? We work on this inverse scenario to solve an ill-posed problem, where given only a single image of an inclined textured planar surface, we attempt to estimate both H and recover the orthogonal (rectified) view of the planar surface. The accurate estimation of homography from a single image is a fundamental task in computer vision with applications ranging from image rectification, image stitching, registration, scene correction, panorama creation, analyzing dynamic scenes, camera calibration to augmented reality. Our research introduces a novel methodology for single-image blind homography estimation, aiming to robustly capture the underlying geometric transformations within the image and subsequent rectification to correct geometric distortions as an application to exhibit its usefulness. Geometric distortions in images, arising from factors such as perspective, tilt, or lens distortions (partly) are rectified using our method. In conclusion, this research introduces a robust and versatile (unsupervised) method for single view (blind) homography estimation and presents a comprehensive solution for image rectification application. One notable strength of our research lies in its adaptability to diverse scenes (such as planar texture images and document images) and imaging conditions. By incorporating techniques to handle challenges such as viewpoint changes and highly oblique views, the proposed approach demonstrates resilience in estimating homographies for images with complex geometric patterns. The effectiveness of the methodology is validated through extensive experiments using our own (curated) and benchmark datasets, exhibiting its accuracy and computational efficiency. Comparative analysis against existing (DLT and ML) methods highlight the superiority of our approach, especially in scenarios with limited data or significant geometric variations. A few failure cases will also be discussed. The talk will conclude with a few observations on a related application of document image correction with composited distortions. |
