No abstract

Much research work in computer vision is being spent on optimizing existing network architectures to obtain a few more percentage points on benchmarks. Recent AutoML approaches promise to relieve us from this effort. However, they are mainly designed for comparatively smallscale classification tasks. In this work, we show how to use and extend existing AutoML techniques to efficiently optimize large-scale U-Net-like encoder-decoder architectures. In particular, we leverage gradient-based neural architecture search and Bayesian optimization for hyperparameter search. The resulting optimization does not require a large company-scale compute cluster. We show results on disparity estimation that clearly outperform the manually optimized baseline and reach state-of-the-art performance.

" This change required some renumbering of subsequent author affiliations. These corrections have been made in the PDF and HTML versions of the article, as well as in any cover sheets for associated Supplementary Information.

Tunable lenses are becoming ubiquitous, in applications including microscopy, optical coherence tomography, computer vision, quality control, and presbyopic corrections. Many applications require an accurate control of the optical power of the lens in response to a time-dependent input waveform. We present a fast focimeter (3.8 KHz) to characterize the dynamic response of tunable lenses, which was demonstrated on different lens models. We found that the temporal response is repetitive and linear, which allowed the development of a robust compensation strategy based on the optimization of the input wave, using a linear time-invariant model. To our knowledge, this work presents the first procedure for a direct characterization of the transient response of tunable lenses and for compensation of their temporal distortions, and broadens the potential of tunable lenses also in high-speed applications.

The Simultaneous Vision simulator (SimVis) is a visual demonstrator of multifocal lens designs for prospective intraocular lens replacement surgery patients and contact lens wearers. This programmable device employs a fast tunable lens and works on the principle of temporal multiplexing. The SimVis input signal is tailored to mimic the optical quality of the multifocal lens using the theoretical SimVis temporal profile, which is evaluated from the through-focus Visual Strehl ratio metric of the multifocal lens. In this paper, for the first time, focimeter-verified on-bench validations of multifocal simulations using SimVis are presented. Two steps are identified as being critical to accurate SimVis simulations. Firstly, a new iterative approach is presented that improves the accuracy of the theoretical SimVis temporal profile for three different multifocal intraocular lens designs -diffractive trifocal, refractive segmented bifocal, and refractive extended depth of focus, while retaining a low sampling. Secondly, a fast focimeter is used to measure the step response of the tunable lens, and the input signal is corrected to include the effects of the transient behavior of the tunable lens. It was found that the root-mean-square of the difference between the estimated through-focus Visual Strehl ratio of the multifocal lens and SimVis is not greater than 0.02 for all the tested multifocal designs.