By: Peter Ondruska, Director, Research; and Guido Zuidhof, Software Engineer
Our conference digest series continues! This week, we’re sharing our summary of two conferences — one focused on computer vision, and one on robotics. Read on for a rundown of our favorite papers from BMVC 2019 and RSS 2019.
BMVC is a small but high-quality computer vision conference that takes place every year in the United Kingdom. We’ve selected the papers that caught our attention from this year’s edition, which happened in Cardiff.
Visuomotor Understanding for Representation Learning of Driving Scenes
Paper from KAIST University, South Korea — Unsupervised pre-training can unlock the potential of the large amount of unlabelled data self-driving cars collect every day. In this paper, the authors outline how to pre-train neural networks on video data that helps to improve performance in semantic segmentation and other tasks.
Orthographic Feature Transform for Monocular 3D Object Detection
Paper from Cambridge University — This paper introduces a new neural network layer that maps image space with a 2D occupancy grid around the vehicle. This transformation helps to achieve a higher level of accuracy in determining the 3D position of visible cars using only one camera.
Pedestrian Action Anticipation Using Contextual Feature Fusion in Stacked RNNs [code]
Paper from Toronto University — Predicting behaviour of pedestrians is one of the most challenging prediction cases for a self-driving car. This paper outlines a method that predicts the likelihood of pedestrian crossing the road using deep learning.
Attentional demand estimation with attentive driving models
Paper from MindVisionLabs, London — Although driving involves interaction with many objects, only some of them are worth paying attention to. The presented system creates heat maps of what to pay attention to. This can then help to decrease computational resources by focusing only on these objects.
Hybrid Deep Network for Anomaly Detection
Paper from University of Montreal — Anomaly detection for self-driving cars is important in order for the car to recognize unusual situations it might not be ready for. In this paper, the authors describe a system for detecting anomalies in static scenes where the network learns statistics of the input using auto-encoder and a classifier.
Triangulation: Why Optimize?
Paper from University of Zaragoza, Spain — In this paper, the authors present a fast closed-form solution to one of the most basic structure-from-motion tasks: Triangulation of features based on multiview observations.
Adversarial Examples for Handcrafted Features
Paper from National University of Science and Technology, Pakistan — Neural networks are susceptible to adversarial attacks . Seemingly correct inputs can produce incorrect output with high confidence. This paper shows that classical hand-engineered systems, such as a feature detection matching pipeline, is equally susceptible to such attacks.
RSS 2019 Digest
RSS is a robotics conference with a long tradition of showcasing the latest algorithms. This year, Level 5 sponsored the conference and sent eight members of the Level 5 team to attend in Freiburg.
ChauffeurNet: Learning to Drive by Imitating the Best and Synthesizing the Worst [website]
Paper from Waymo — Mid-to-mid deep learning for controlling a vehicle is the latest trend in self-driving research. This paper explores how to train a neural network to drive a self-driving car capable of avoiding problems associated with imitation learning — specifically the poor performance in never-before-seen scenarios. The system was evaluated both in simulation and real-world.
Segment2Regress: Monocular 3D Vehicle Localization in Two Stages
Paper from KAIST, South Korea — Knowledge of the environment can simplify and improve performance of perception systems. This new method shows how to use the knowledge of the ground plane to estimate the 6dof pose of a car from camera data, demonstrating state-of-the-art performance on the KITTI dataset.
Game Theoretic Planning for Self-Driving Cars in Competitive Scenarios
Paper from Stanford University — Optimal planning in reactive environment is difficult, as the planner needs to take into account the likely actions of other agents responding to this plan. The paper presents a system that plans a trajectory for a car while specifically anticipating responses of other cars on a racetrack.
OIL: Observational Imitation Learning [website]
Paper from KAUST University, Saudi Arabia — One of the ways to teach a self-driving car to drive is to provide driving examples for system to imitate. Authors demonstrate a framework that can learn basic maneuvers from imperfect examples and evaluate its simulated performance.
Harnessing Reinforcement Learning for Neural Motion Planning
Paper from Technion Institute, Israel — Reinforcement learning is a promising approach for robotics problems; however, it is hard to make this work in practice. This paper shows how to merge Reinforcement learning and neural planning for robotics manipulation.
If this post caught your interest, we’re hiring! Check out our open roles here, and be sure to follow our blog for more technical content.
Self-driving Research in Review: BMVC 2019 and RSS 2019 was originally published in Lyft Level 5 on Medium, where people are continuing the conversation by highlighting and responding to this story.