domeet webmaster

Focusing on the frontier of machine vision, it aims to give AI a lifelong learning ability through game exploration.

Introduction to IROS Meeting

The IEEE International Conference on Intelligent Robotics and Systems (IROS) is one of the top two international conferences in the field of intelligent robotics and automation. Each IROS conference and its exhibitions have been extremely successful, in related fields. Technological development has played an important role in promoting. IROS 2019 is the 32nd session of IROS, co-sponsored by the world’s largest non-profit professional technology society IEEE, IEEE Robotics and Automation Society, IEEE Industrial Electronics Society, Japan Robotics Society, Institute of Instrument and Control Engineers and New Technology Foundation. At that time, about 4,000 leaders from around the world, such as robots, automation systems and artificial intelligence, top research team representatives and business people will gather in Macau to explore cutting-edge technology in the field of intelligent robots and systems. Share and discuss the latest developments in related fields.

IROS 2019 will include keynotes, technical reports, seminars, competitions, forums and exhibitions. The Lifetime Machine Vision Dataset Global Challenge is part of the IROS 2019 competition.

Event Background

Focusing on the frontiers of machine vision, it aims to give AI a lifelong learning ability through game exploration.

Human: Continue to learn knowledge and skills from the environment and experience

Robots: Require lifelong learning to adapt to changing environments and tasks

Computer Vision: Learn from a pre-built data set

In recent years, new advances in large data sets such as ImageNet and COCO have led to significant improvements in computer vision technology based on deep learning. At present, computer vision applications based on object detection, segmentation and recognition based on a large number of data sets have made outstanding contributions in the fields of smart home, security, and industrial inspection. However, robot vision poses new challenges for the development and landing of visual algorithms. Computer vision algorithms implicitly assume independent and identical distribution of data, such as fixed categories, and a single simple task. It is clear that the semantic concept of the real environment will change dynamically over time. In practical scenarios, robots need to continue to operate in a variable environment for a long time, which requires the robot to have the ability to learn for life to adapt to changes in the environment.

For one of the studies: Lifelong SLAM

Lifelong SLAM -Positioning algorithm adapted to scene changes

SLAM’s full name “simultaneous positioning and mapping” is designed to enable robots to estimate their position and posture independently during the movement process. It is one of the core issues in the field of robotics. Traditional SLAM research often focuses on the positioning accuracy of robots in specific scenes, while ignoring the positioning failure and mismatching problems caused by scene changes. In response to this shortcoming, this competition proposes the positioning success rate, focusing on whether the SLAM algorithm can stably identify its position when the viewing angle, illumination and scene layout change, thus supporting the long-term deployment of the robot.

To tie in with this event, the researchers produced a new SLAM dataset, OpenLORIS-Scene. Compared with the previous SLAM dataset, the scenes contained in OpenLORIS-Scene are closer to life, the sensor configuration is more abundant, and each scene is recorded multiple times, thus including scene changes caused by real life. The OpenLORIS-Scene dataset will be the touchstone for the SLAM algorithm to support the real deployment of robots.

For Study 2: Lifelong Learning

Lifelong Learning – Continuous learning adds new knowledge

Mainstream work tends to get pre-trained models based on large data sets, and then fine-tune or retrain based on the specific application data set. But the resulting model goesThe past has forgotten the patterns that have been learned before (such as the category of the object). This phenomenon is called “catastrophic forgetting” in deep learning. For the robot to face dynamic scenes, it is necessary to retrain the deployed model, which requires the robot to have real memory ability and effectively overcome the defects of catastrophic forgetting.

Lifetime Machine Vision Dataset: OpenLORIS

For SLAM and machine learning research in life-long machine vision, the event organizer produced a new data set, OpenLORIS, for research and testing of the corresponding algorithms. The OpenLORIS dataset is all collected from everyday scenes and is the world’s first large-scale dataset containing changes in indoor scenes.

OpenLORIS-Scene vs. existing scene dataset

OpenLORIS-Object vs. existing object dataset

Top international events in the hottest areas of robotics and artificial intelligence

• The first event was launched at the top international conferences

This event is the official competition unit of IROS Conference;

This event will be directly oriented to more than4,000 experts in robotics and artificial intelligence industry;

• The competition is divided into two major tracks, which are the most popular research topics in the field of robotics and artificial intelligence

Lifetime SLAM (simultaneous positioning and mapping) algorithm track – the ability of the competition robot to continuously self-locate through vision

Final Biometric Tracking Algorithm – The ability of a competing robot to learn new knowledge while not forgetting old knowledge

• Participants are industry practitioners and research institutions of renowned institutions

It is expected that there will be dozens of teams from all over the world in academia and industry participating in each challenge

The currently enrolled research team is from: MIT / Imperial College / Tsinghua University / Hong Kong University of Science and Technology / University of Macau / McGill University / Heriot-Watt U / U Zaragoza / Tel Aviv U / Indian Academy of Sciences / Mercedes Benz / 驭势科技 / …

Contest Process

Dataset publishing 2019/7

• Register for the registration contest

• Download dataset

• Developing algorithms, using dataset enhancement algorithms

Preliminary 2019/7/15 – 2019/9/30

• Download the contest dataset, software tools, upload the results before the deadline

• The high-scoring team advances to the finals and invites to participate in the IROS offline event

Final 2019/10/1 – 2019/10/25

• The final will use the specified new data set

• The entry algorithm will need to run in the specified computing environment

On-site activities 2019/11/4

Location: The Venetian Macao, China

• Competition Reports and Special Reports

• The final game results are announced

• The winning team will be awarded at the IROS 2019 Awards Luncheon

Registration method

Based on the OpenLORIS dataset, researchers from Intel China Research Institute, Tsinghua University, and City University of Hong Kong jointly launched the Lifelong Robotic Vision Challenge. The event was held in conjunction with IROS 2019, the top robotics conference, including Lifelong Object Recognition and Lifelong SLAM. Both challenges are currently undergoing a two-month algorithmic competition. Global researchers can register for the game and download the data and upload the algorithm results for online scoring. Winners will be invited to the finals and report on the IROS 2019 live workshop. . The competition seminar will be held in Macau on November 4th. In addition to the winners of the competition, Professor Chen Baoquan from Peking University, Professor Zhang Jianwei from the University of Hamburg, Germany, and Professor Giorgio Metta from the Italian IIT Research Center will be invited to give on-the-spot reports. Participate in the discussion.

For more details, please visit lifelong-robotic-vision.github.io/competition