Shape from Focus with Adaptive Focus Measure and High Order Derivatives

Yuval Frommer, Rami Ben-Ari and Nahum Kiryati

Abstract

Shape From Focus (SFF) methods frequently use a single focus measure to obtain a depth map. Common focus measures are fixed and spatially invariant. In this paper we present a framework to create an adaptive focus measure based on ensemble of basis focus operators. Using the proposed framework we derive a new spatially variant focus measure obtained from linear combination of image derivatives. This approach effectively generalizes some of the existing measures. A new measure emerged from the proposed framework includes high order derivatives and presents a highly reliable focus measure. We rely on the focus curve standard deviation (CSTD) to determine the linear coefficients in our model. The emerged focus measure copes effectively with texture variation, strong intensity edges and depth discontinuities. Using CSTD we further suggest a new approach for aggregation in the focus volume succeeded by reconstruction based on the focus curve centroid. This different approach of aggregation and reconstruction yields improved depth maps, respecting shape smoothness and depth discontinuities for diversity of textured images. We assess the performance of our new approach by extensive experiments with highly realistic synthetic images and real images including two unique cases captured in the wild. In terms of focus measure, we significantly outperform the state-of-the-art, while presenting superior results comparing to two previously published alternatives.

Session

Poster 2

Files

Extended Abstract (PDF, 6M)
Paper (PDF, 11M)

DOI

10.5244/C.29.134
https://dx.doi.org/10.5244/C.29.134

Citation

Bibtex

@inproceedings{BMVC2015_134,
	title={Shape from Focus with Adaptive Focus Measure and High Order Derivatives},
	author={Yuval Frommer and Rami Ben-Ari and Nahum Kiryati},
	year={2015},
	month={September},
	pages={134.1-134.12},
	articleno={134},
	numpages={12},
	booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
	publisher={BMVA Press},
	editor={Xianghua Xie, Mark W. Jones, and Gary K. L. Tam},
	doi={10.5244/C.29.134},
	isbn={1-901725-53-7},
	url={https://dx.doi.org/10.5244/C.29.134}
}