Meta tags:

Headings (most frequently used words):

-

Text of the page (most frequently used words):
the (17), and (13), dataset (10), our (8), learning (7), autofocus (6), depth (5), capture (5), single (5), algorithm (5), focal (5), with (4), then (4), index (4), two (4), code (3), over (3), for (3), different (3), these (3), where (3), receives (3), most (3), focus (3), stack (3), any (2), questions (2), classification (2), ordinal (2), regression (2), loss (2), paper (2), dot (2), have (2), samples (2), rgb (2), charles (2), herrmann (2), richard (2), strong (2), bowen (2), neal (2), wadhwa (2), rahul (2), garg (2), qiurui (2), jonathan (2), barron (2), ramin (2), zabih (2), based (2), approach (2), this (2), problem (2), provide (2), realistic (2), pixel (2), from (2), using (2), dual (2), improvement (2), learned (2), stacks (2), slice (2), random (2), starting (2), point (2), estimates (2), full (2), step (2), sensor (2), data (2), results (2), are, happy, answer, regarding, follows, standard, pipeline, slightly, altered, mobilenetv2, details, included, please, contact, cih, cornell, edu, you, training, set, split, archives, train7, train6, train5, train4, 102, train3, train2, 100, train1, test, readme, cvf, arxiv, cvpr, 2020, papers, important, task, digital, cameras, yet, current, approaches, often, exhibit, poor, performance, propose, sufficient, size, effective, labeled, per, depths, obtained, multi, view, stereo, following, camera, estimation, pixels, apply, modern, deep, models, obtain, efficient, technique, demonstrate, that, provides, significant, compared, previous, non, methods, model, reduces, mean, absolute, error, factor, best, comparable, baseline, publicly, available, abstract, 510, captured, wild, along, computed, sfm, views, large, variation, color, texture, scene, elements, define, three, problems, but, can, pick, next, uses, captures, estimate, each, formulations, run, computations, input, values, conventional, network, substantial, improvements, baselines, leading, right, middle, formulation, left, google, research,

Text of the page (random words):
learning to autofocus learning to autofocus charles herrmann richard strong bowen neal wadhwa rahul garg qiurui he jonathan t barron ramin zabih google research paper dataset capture samples of rgb and depth dataset left our dataset we provide a realistic dataset of 510 focal stacks captured in the wild along with a computed depth from sfm on 5 different views these focal stacks have a large variation in color texture scene elements and depth middle our problem formulation we define autofocus as three different problems single slice where the algorithm receives a single capture at a random starting point and then estimates the most in focus index focal stack where the algorithm receives the full focal stack and then estimates the most in focus index and two step where the algorithm receives a single capture at a random starting point but can then pick the next index to capture then the algorithm uses these two captures to estimate the most in focus index for each of these formulations we run computations on two different input values conventional sensor data and dual pixel sensor data right our results our network results in substantial improvements over the baselines leading to a 4x improvement in single slice 3x in two step and 1 5x in full focal stack abstract autofocus is an important task for digital cameras yet current approaches often exhibit poor performance we propose a learning based approach to this problem and provide a realistic dataset of sufficient size for effective learning our dataset is labeled with per pixel depths obtained from multi view stereo following learning single camera depth estimation using dual pixels using this dataset we apply modern deep classification models and an ordinal regression loss to obtain an efficient learning based autofocus technique we demonstrate that our approach provides a significant improvement compared with previous learned and non learned methods our model reduces the mean absolute error by a factor of 3 6 over the best comparable baseline algorithm our dataset is publicly available papers learning to autofocus charles herrmann richard strong bowen neal wadhwa rahul garg qiurui he jonathan t barron and ramin zabih cvpr 2020 arxiv cvf dataset capture samples of rgb and depth dataset readme test 89 gb training set split over 7 archives train1 95 gb train2 100 gb train3 99 gb train4 102 gb train5 99 gb train6 99 gb train7 87 gb code we are happy to answer any questions regarding the code the code follows the standard classification pipeline with ordinal regression loss and a slightly altered mobilenetv2 details included in the paper please contact cih at cs dot cornell dot edu if you have any questions