arXiv subjects

Jonathan T. Barron

Publications and source records attributed to Jonathan T. Barron.

97 records · Page 6Linked to original sources

arXiv2017

Continuously Differentiable Exponential Linear Units

Exponential Linear Units (ELUs) are a useful rectifier for constructing deep learning architectures, as they may speed up and otherwise improve learning by virtue of not have vanishing gradients and by having mean activations near zero. However, the ELU activation as parametrized in [1] is not continuously differentiable with respect to its input when the shape parameter alpha is not equal to 1. We present an alternative parametrization which is C1 continuous for all values of alpha, making the rectifier easier to reason about and making alpha easier to tune. This alternative parametrization has several other useful properties that the original parametrization of ELU does not: 1) its derivative with respect to x is bounded, 2) it contains both the linear transfer function and ReLU as special cases, and 3) it is scale-similar with respect to alpha.

cs.LG↗

arXiv2016

The Fast Bilateral Solver

We present the bilateral solver, a novel algorithm for edge-aware smoothing that combines the flexibility and speed of simple filtering approaches with the accuracy of domain-specific optimization algorithms. Our technique is capable of matching or improving upon state-of-the-art results on several different computer vision tasks (stereo, depth superresolution, colorization, and semantic segmentation) while being 10-1000 times faster than competing approaches. The bilateral solver is fast, robust, straightforward to generalize to new domains, and simple to integrate into deep learning pipelines.

cs.CV↗

arXiv2016

Semantic Image Segmentation with Task-Specific Edge Detection Using CNNs and a Discriminatively Trained Domain Transform

Deep convolutional neural networks (CNNs) are the backbone of state-of-art semantic image segmentation systems. Recent work has shown that complementing CNNs with fully-connected conditional random fields (CRFs) can significantly enhance their object localization accuracy, yet dense CRF inference is computationally expensive. We propose replacing the fully-connected CRF with domain transform (DT), a modern edge-preserving filtering method in which the amount of smoothing is controlled by a reference edge map. Domain transform filtering is several times faster than dense CRF inference and we show that it yields comparable semantic segmentation results, accurately capturing object boundaries. Importantly, our formulation allows learning the reference edge map from intermediate CNN features instead of using the image gradient magnitude as in standard DT filtering. This produces task-specific edges in an end-to-end trainable system optimizing the target semantic segmentation quality.

cs.CV↗

arXiv2016

Multiscale Combinatorial Grouping for Image Segmentation and Object Proposal Generation

We propose a unified approach for bottom-up hierarchical image segmentation and object proposal generation for recognition, called Multiscale Combinatorial Grouping (MCG). For this purpose, we first develop a fast normalized cuts algorithm. We then propose a high-performance hierarchical segmenter that makes effective use of multiscale information. Finally, we propose a grouping strategy that combines our multiscale regions into highly-accurate object proposals by exploring efficiently their combinatorial space. We also present Single-scale Combinatorial Grouping (SCG), a faster version of MCG that produces competitive proposals in under five second per image. We conduct an extensive and comprehensive empirical validation on the BSDS500, SegVOC12, SBD, and COCO datasets, showing that MCG produces state-of-the-art contours, hierarchical regions, and object proposals.

cs.CV↗

arXiv2015

Convolutional Color Constancy

Color constancy is the problem of inferring the color of the light that illuminated a scene, usually so that the illumination color can be removed. Because this problem is underconstrained, it is often solved by modeling the statistical regularities of the colors of natural objects and illumination. In contrast, in this paper we reformulate the problem of color constancy as a 2D spatial localization task in a log-chrominance space, thereby allowing us to apply techniques from object detection and structured prediction to the color constancy problem. By directly learning how to discriminate between correctly white-balanced images and poorly white-balanced images, our model is able to improve performance on standard benchmarks by nearly 40%.

cs.CV↗

arXiv2008

Blind Date: Using proper motions to determine the ages of historical images

Astrometric calibration is based on patterns of cataloged stars and therefore effectively assumes a particular epoch, which can be substantially incorrect for historical images. With the known proper motions of stars we can "run back the clock" to an approximation of the night sky in any given year, and in principle the year that best fits stellar patterns in any given image is an estimate of the year in which that image was taken. In this paper we use 47 scanned photographic images of M44 spanning years 1910-1975 to demonstrate this technique. We use only the pixel information in the images; we use no prior information or meta-data about image pointing, scale, orientation, or date. Blind Date returns date meta-data for the input images. It also improves the astrometric calibration of the image because the final astrometric calibration is performed at the appropriate epoch. The accuracy and reliability of Blind Date are functions of image size, pointing, angular resolution, and depth; performance is related to the sum of proper-motion signal-to-noise ratios for catalog stars measured in the input image. All of the science-quality images and 85 percent of the low-quality images in our sample of photographic plate images of M44 have their dates reliably determined to within a decade, many to within months.

astro-ph↗

arXiv2008

Cleaning the USNO-B Catalog through automatic detection of optical artifacts

The USNO-B Catalog contains spurious entries that are caused by diffraction spikes and circular reflection halos around bright stars in the original imaging data. These spurious entries appear in the Catalog as if they were real stars; they are confusing for some scientific tasks. The spurious entries can be identified by simple computer vision techniques because they produce repeatable patterns on the sky. Some techniques employed here are variants of the Hough transform, one of which is sensitive to (two-dimensional) overdensities of faint stars in thin right-angle cross patterns centered on bright ($<13 \mag$) stars, and one of which is sensitive to thin annular overdensities centered on very bright ($<7 \mag$) stars. After enforcing conservative statistical requirements on spurious-entry identifications, we find that of the 1,042,618,261 entries in the USNO-B Catalog, 24,148,382 of them ($2.3 \percent$) are identified as spurious by diffraction-spike criteria and 196,133 ($0.02 \percent$) are identified as spurious by reflection-halo criteria. The spurious entries are often detected in more than 2 bands and are not overwhelmingly outliers in any photometric properties; they therefore cannot be rejected easily on other grounds, i.e., without the use of computer vision techniques. We demonstrate our method, and return to the community in electronic form a table of spurious entries in the Catalog.

astro-ph↗