Towards Science in Chile with LSST Viña del Mar, December 2016

AstroCV: A Computer Vision Library for Astronomy

Alejandro Sazo (UTFSM, Engineer) Roberto González (PUC, Astronomer) Felipe San Martín (Universidad de Chile, Engineer)

Roberto Muñoz Instituto de Astrofísica

Pontificia Universidad Católica de Chile

Surveys before CCDsBefore CCDs became the standard in Astronomy, the major surveys were conducted using photographic plates.

Hubble galaxy classification and Sersic profiles were derived at that epoch.

Roberto Muñoz

Advent of large and rich surveys

Roberto Muñoz

Time span: 15 yearsArea: 14,555 deg2

# sources: 469,053,874

Time span: 10 yearsArea: 25,000 deg2

# sources: 37,000,000,000

CCDs became more sensitive and mosaics were affordable

Photometric Spectroscopic UV-IR coverage High cadence

Galaxies in the UniverseGalaxies come in many flavors

• Giants and Dwarfs • Spirals, Ellipticals, Irregulars • LSBs, UCDs, UFDs, UDGs

Roberto Muñoz

M87

Malin 1

NGC 1300

Galaxy Zoo and KaggleGalaxyZoo: One of the most successful citizen science project in Astronomy. Started in 2007.

Visual classification of about 900.000 galaxies

Roberto Muñoz

SDSS Hubble ACS and WFC3

Kaggle competition37 classes from the GalaxyZoo decision tree

Why Computer Vision?

Interdisciplinary field: CS, Physics, Math, AI

Computers gain high-level understanding

Roberto Muñoz

Ongoing (DES, VVV) and coming surveys (LSST) will produce enormous datasets.

Compute features for later classification

Automatic detection, segmentation and classification

Features and classificationAstronomers have been computing features for years: luminosity, color, µr, reff, Sersic index, etc

Major improvements for feature descriptor: SIFT, HOG, GISTand classification: SVM, Random forest, Neural network

Roberto Muñoz

Missrate and FPS

Roberto Muñoz

2003

2014

FPS

Region-based FullyConvolutional Networks = 6 FPS

Dai et al. 2016Benenson et al. 2014

Dollár et al. 2014

AstroCV libraryWe are developing a computer vision library optimised for astronomical datasets. Our goal is doing on-the-fly data processing using GPUs.

The library is being developed in C++ and CUDA. The feature descriptor is based on HOG and the classification in Random forest.

We process 30,000 x 30,000 pixel image in 25 seconds, single GPU. Fast Deep Learning will be 8 minutes.

Roberto Muñoz

github.com/rpmunoz/astrocv

FITS or JPEG?The computer vision community prefers working with 8-bit and raster images (computing power and memory).

AstroCV creates a RGB image using 3 filters. Asinh scaling.

Roberto Muñoz

Dwarf galaxies

New datasetsSupervised learning requires large datasets

Collaborators are currently observing the Virgo, Fornax and Coma clusters. Study GCs, UCDs, giant and dwarf galaxies.

Better resolution and SNR in comparison to SDSS

Roberto Muñoz

NGFS surveyMuñoz et al. 2015

NGVS surveyFerrarese et al. 2012

Ongoing workThe datasets and code will be available in http://astrocv.org

Interface similar to ImageNet. Search by telescope, survey, morphological type, etc.

Roberto Muñoz