CVPR2025

MaDCoW: Marginal Distortion Correction for Wide-Angle Photography with Arbitrary Objects

Kevin Zhang, Jia-Bin Huang, Jose Echevarria, Stephen DiVerdi, Aaron Hertzmann

Abstract

Linear Perspective Current state-of-the-art Our results (a) (b) (c) Linear Perspective Current state-of-the-art Our results (d) (e) (f) Figure 1. Correcting marginal distortion in arbitrary objects. (a) Linear perspective projection of a wide field-of-view (here, 150 • ) produces extreme marginal distortions, such as the buildings' distorted appearances. (b) The method of Carroll et al. [6] with line annotations (top left inset) preserves some straight lines, but the building on the right seems to lean left. (c) For MaDCoW, we supply a single region annotation for the building on the right, and MaDCoW successfully avoids distortion for the whole building. (d) A linear perspective photograph (90 • field-of-view), from Shih et al. [26]. The man on the left appears widened. (e) Shih et al. [26] correct distortion on faces, but not on bodies. The leftmost person's head is corrected but not his body, making his body look even larger, relative to his head. (f) MaDCoW makes people on the margins appear more realistic.