ACM SIGSPATIAL, Nov. 6-9 2018
Academic by nature, but backed by industrial sponsors, the conference is organized around the main problems for the field with an orthogonal dimension highlighting the tools used to solve them. As a relatively newcomer to the field, I was pleased to find more than half of the presentations to be somewhat relevant to what we’re working on here at NAVER LABS Europe.
With a bit more than 400 attendees, three quarters of whom were from academia, and a friendly atmosphere, it’s perfect for knowledge sharing especially in the thematic workshops held the day before the main event and the poster session on the first evening. The mix of hyper-experts in their subdomain thinking about their problems and the technology focused practitioners with their performant but opaque-ish solutions, did however translate into somewhat limited reactions to many talks.
Xin Chen, Director of Engineering for Highly Automated Driving at HERE, gave the first keynote entitled “HD Live Maps for Automated Driving: An AI Approach”. Alas it was not recorded but an earlier version of a similar talk at IDEAS in Chicago is available. Chen told us about HERE’s pipeline to build and maintain HD Maps, making heavy use of deep learning, the differences between urban and rural areas, the difficulty of dealing with parked cars, the use of the “crowd” (of cars), and their quality index (presence, accuracy, classification).
On the 2nd day, (now) Apple’s Daniel Delling, father of RAPTOR, had a keynote entitled “Route planning in transportation networks – From research to practice”. He took us on a fascinating tour of trip planning over the last 20 years, a perfect example of academia working successfully on an industry problem, first in the open case of route planning (e.g. for cars) then in the more constrained case of Public Transit (PT). It’s rather interesting that the former is considered to have been a solved research problem for almost 10 years, with algorithms powerful enough to now enable the suggestion of detours along a main itinerary, e.g. to stop at Points of Interest (POIs), yet there still remains a number of difficult challenges for PT trip planning, not handled as a graph problem e.g. when trying to optimize for multiple criteria.
Just to illustrate the width and depth of SIGSPATIAL program, here are a few selected papers from the presentations I attended.
Topological Signatures For Fast Mobility Analysis by Ghosh et al. at the University of Edinburgh, looks at ways to map trajectories to a low dimensional Euclidean space using topological signatures, removing noise but retaining enough information to apply mining techniques.
A Force-Directed Approach for Offline GPS Trajectory Map Matching by Rappos et al. from the University of Applied Sciences of Western Switzerland, uses the Force to match GPS trajectories to an existing road network.
Efficient Generation of Geographically Accurate Transit Maps by Bast et al. at the University of Freiburg, produces geographically accurate transit maps from schedule data provided in a GTFS and nicely orders lines for esthetics crossing and separation such as the ones found on manually designed maps.
Efficient Astronomical Query Processing using Spark, Brahem et al. from the Université Versailles St Quentin, bridges the expressivity of ADQL the Astronomical Data Query Languages used by scientists in this field and the power of SPARK for familiar and efficient queries on huge datasets.
What Is It Like Down There? Generating Dense Ground-Level Views and Image Features From Overhead Imagery Using Conditional Generative Adversarial Networks by Deng et al. at the University of California, has competing Conditional GANs that use aerial imagery to generate ground level views which capture (some of) the expected natural qualities of the locations.
SIGSpatial is clearly a very interesting conference for the Geospatial communities but it’s also a valid stage for research in technical fields applied to the problems of the domain.
To my personal taste there weren’t enough papers on geospatial semantics: maybe it’s seen as a solved problem and the focus has shifted to HD (cf the first keynote) or Indoor mapping… The effectiveness of deep learning might also be luring practitioners looking for quick results, but the brilliant “gallery of applications for Geosimulation” presented by George Mason University’s Andrew Crooks to keynote the GeoSim workshop tells me otherwise: explanations matter to this community.