9:25 Workshop Opening
9:30 Invited talk of Eike Rehder (KIT, Germany)
"Neural Networks as Graph Solvers"
10:00 Contributed Talk of Anelia Angelova (Google, US)
Learning with Proxy Supervision for End-To-End Visual Learning
10:15 Contributed Talk of Bert De Brabandere (University of Leuven - KU Leuven, Belgium)
Fast Scene Understanding for Autonomous Driving. Download the extended Abstract: Fast Scene Understanding
10:30 Coffee Break / Poster Session
11:00 Invited talk of Fisher Yu (UC Berkeley, US)
Learning from Large Scale Driving Data with Efficient Networks
11:30 Invited talk of Rudolf Mester (J.W.Goethe-Universitaet Frankfurt, Germany)
"Enhancing Monocular Visual Odometry with Deep Learning"
12:00 Invited talk of Andrew G. Howard (Google, US)
"MobileNets: Efficient Convolutional Neural Networks"
12:30 Lunch Break // Poster Session
13:30 Invited talk of Adrien Gaidon (Toyota Research Institute, US)
"Need for Sim: Procedural Generation of Realistic Driving Environments and Human Actions for Deep Learning"
14:00 Invited talk of Xue Mei (TuSimple, US)
Autonomous Driving on Benchmarks
14:30 Poster Session



Deep learning, and in particular convolutional neural networks, has become the main component of many intelligent vehicle algorithms. Jointly with the explosive growth in the available amount of driving data these data-driven algorithms have certainly enabled the next generation of platforms for reliable autonomous driving as evidenced by the algorithms used by a large number of companies like MobilEye, AutoX, Zoox, Toyota Research, General Motors, Volkswagen, and Daimler among many others. The goal of this workshop on deep learning for vehicle perception (as a second edition of the Deep-Driving workshop held in conjunction with IV2016 for further information) is to foster discussion and to accelerate the study of deep architectures in autonomous driving problems with a focus on the efficiency of the algorithms.

More precisely, the goal of the second edition of the Deep-Driving workshop is two fold. First, to foster discussion and to accelerate the study of the use of deep architectures in autonomous driving problems. Deep learning has become the main component of many computer vision algorithms. However, efficient algorithms working on real-world driving environments have specific challenges (e.g., varying acquisition conditions or hardware constraints) that still need to be addressed. The second goal of the workshop is to foster progress on network efficiency which is a current necessity for practical intelligent vehicles and robotics in general. To achieve this second goal a scene labeling challenge will be organized along the workshop to emphasize this specific. Interested participants will receive video snippets from several European cities for processing. Details will be announced soon.

The workshop program will include invited talks from different areas within industry and academia to highlight address the main challenges within this topic and will also invite the presentation of extended abstracts to encourage discussion related to the last advances in this area. Moreover, following with the experience of the first edition of the workshop, we will provide hands-on-experience towards the use of deep learning algorithms. Finally, during the workshop, the winners of the challenge will be announced.

In summary, with the second edition of this workshop on learning representations for autonomous driving we aim at providing basic theoretical and practical knowledge to understanding the potential of deep learning architectures as well as to encourage discussions on the specific challenges within the field of intelligent vehicles and advance the current status of efficient algorithms by proposing a challenge in this particular topic.

The idea of the poster session is to support discussions between participants, not to provide "yet another publication venue".