Inside the 2018 AI Challenger: Datasets, Tracks, and Real‑World Impact

Competition Overview

On August 29, 2018, the AI Challenger 2018 Global AI Competition was officially launched by Meituan, Innovation Works, Sogou, and Meitu. The event was inaugurated by senior executives from the four sponsors and offered a total prize pool exceeding 3 million RMB. Registration closed on October 13, 2018.

The competition invested tens of millions of RMB to build more than ten high‑quality datasets, becoming the largest domestic research‑data platform and non‑commercial competition venue. In 2017, 8,892 teams from 65 countries participated; the 2018 edition aimed to surpass those records.

Key Tracks

The 2018 theme was “Using AI to Tackle Real‑World Problems,” comprising five main tracks and five experimental tracks. Meituan was responsible for two challenging main‑track problems:

Fine‑grained sentiment analysis of user comments in the restaurant domain (NLP track).

Autonomous‑driving visual perception (self‑driving track).

Fine‑Grained Sentiment Analysis Track

Meituan’s NLP Center provided the largest Chinese restaurant‑review sentiment dataset to date, containing 150,000 annotated comments sourced from the public Dazhong Dianping platform. The dataset features a two‑level labeling scheme: a coarse layer for the overall target and a fine layer with six categories and 20 sub‑elements, each verified by at least two professional annotators.

Key characteristics:

Addresses the scarcity of Chinese restaurant‑domain sentiment data, enriching resources beyond movie‑review and shopping datasets.

Rich, high‑quality data covering diverse cuisines and regions, reflecting real‑world user opinions.

Scientifically designed sampling and annotation methodology, ensuring reliability for research and industrial use.

The dataset supports multiple downstream applications, such as improving search ranking on Dazhong Dianping, helping merchants identify quality issues across dimensions (food, service, environment), and powering intelligent customer‑service bots capable of recognizing emotions like happiness, praise, dissatisfaction, and anger.

Autonomous‑Driving Visual Perception Track

The self‑driving track leveraged the Berkeley DeepDrive (BDD) dataset, one of the world’s largest autonomous‑driving collections, with 1.2 billion raw images and 100,000 annotated frames covering diverse weather, lighting, and geographic conditions across four U.S. regions.

Participants focused on two perception tasks:

Object detection : locating and classifying pedestrians, traffic lights, signs, and various vehicles.

Drivable‑area segmentation : distinguishing priority and non‑priority traversable zones.

The competition encouraged multi‑task learning, asking teams to solve detection and segmentation jointly while keeping models lightweight and fast—a rare industry‑level challenge that integrates transfer learning concepts.

Meituan’s autonomous‑delivery vehicles already employ the resulting perception algorithms, combining camera and LiDAR data for path planning, obstacle avoidance, lane keeping, and intelligent following in low‑speed, controlled environments.

Impact and Outlook

Both tracks aim to advance AI research and its practical deployment. The sentiment analysis dataset fuels improvements in natural‑language understanding, knowledge‑graph construction, and personalized recommendation. The autonomous‑driving track pushes forward perception technology essential for safe, efficient delivery robots and future passenger‑level self‑driving services.

Meituan’s participation also includes collaborations with UC Berkeley’s DeepDrive alliance and Tsinghua University, fostering joint research on algorithmic innovation and real‑world validation.

Conclusion

The AI Challenger 2018 provided unprecedented data resources and challenging problem statements that bridge academic research and industrial application, inviting global talent to develop next‑generation AI models that can improve everyday life, from smarter restaurant recommendations to safer autonomous delivery.