IEEE Geoscience and Remote Sensing Magazine - March 2019 - 103

Technical committees
BerTrand Le Saux, naoTo YokoYa, ronnY HänScH,
MYron Brown, and GreG HaGer

2019 Data Fusion Contest

he 2019 Data Fusion Contest,
organized by the Image Analysis and Data Fusion Technical Committee (IADF TC) of the IEEE Geoscience and Remote Sensing Society
(GRSS), the Intelligence Advanced
Research Projects Activity (IARPA), and Johns Hopkins University
(JHU), aims to promote research
in semantic 3D reconstruction and
stereo using machine intelligence
and deep-learning techniques applied to satellite images (Figure 1).
The objective is to reconstruct both
a 3D geometric model and a segmentation of semantic classes for an urban scene. Incidental satellite images, airborne lidar data, and semantic labels are provided to the community. As in previous
years [1], [2], the contest will be set up as a benchmark
competition. Four different tasks are evaluated in four
parallel and independent competitions:
◗ track 1: the single-view semantic 3D challenge
◗ track 2: the pairwise semantic stereo challenge
◗ track 3: the multiview semantic stereo challenge
◗ track 4: the 3D point-cloud classification challenge.
We distribute urban semantic 3D (US3D) data, a
large-scale public data set with multiview, multiband satellite images and reference geometric and semantic labels
for two large cities [3]. The US3D data set covers approximately 100 km2 over Jacksonville, Florida, and Omaha,
Nebraska. For the contest, we provide training and test
data sets for each challenge track, including approximately 20% of the US3D data. Incidental satellite images
consist of WorldView-3 panchromatic and eight-band
visible and near-infrared (VNIR) images (provided courDigital Object Identifier 10.1109/MGRS.2019.2893783
Date of publication: 19 March 2019

MARCH 2019

ieee Geoscience and remote sensing magazine

tesy of DigitalGlobe), collected between 2014 and 2016, with a ground
sampling distance of approximately
35 cm and 1.3 m for panchromatic
and VNIR images, respectively (see
Figure 2). Airborne lidar data come
from the Homeland Security Infrastructure Program, and the aggregate nominal pulse spacing is
approximately 80 cm. Training data
derived from lidar include reference
data above-ground-level height images for track 1, disparity images for
track 2, and digital surface models
for track 3. Semantic labels derived
from the lidar products are released as semantic maps for
tracks 1-3 and semantic point clouds for track 4. Classes
comprise buildings, elevated roads and bridges, high vegetation, ground, water, and so on.
In track 1, single-view semantic
3D, an unrectified single-view imIn track 4, 3D poIntage is provided for each geograclouD classIfIcatIon,
phic tile. The objective is to predict
the aIm Is to label
semantic labels and above-ground
poInts from the gIven
heights. In track 2, pairwise semantic stereo, a pair of epipolar
aerIal poInt clouD
rectified images is given, and the
accorDIng to several
objective is to predict semantic
preDetermIneD
labels and stereo disparities. In
semantIc classes.
track 3, multiview semantic stereo, the goal is to predict semantic
labels and a digital surface model
given several multiview, unrectified images associated
with a precomputed geometry model to focus on the data
fusion problem and not on registration.
For tracks 1-3, performance is assessed using the pixel-wise mean intersection over union (mIoU), for which
true positives must have both the correct semantic label
0274-6638/19©2019IEEE

103

IEEE Geoscience and Remote Sensing Magazine - March 2019

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2019