Baidu, the Chinese Internet giant, has released ApolloScape, a massive dataset for autonomous vehicle simulation and machine learning.
ApolloScape is an order of magnitude bigger and more complex than existing similar datasets such as Kitti and CityScapes. ApolloScape offers 10 times more high-resolution images with pixel-by-pixel annotations, and includes 26 different recognizable objects such as cars, bicycles, pedestrians and buildings. The dataset offers several levels of scene complexity with increasing number of pedestrians and vehicles, up to 100 vehicles in a given scene, as well as a wider set of challenging environments such as heavy weather or extreme lighting conditions. The ApolloScape dataset is a work in progress, and this release corresponds to the first subset, which contains 144k image frames.
The ApolloScape dataset is part of version 2 of Apollo, Baidu's open autonomous driving platform. The Apollo source code, open sourced under an Apache-2.0 license, includes a 2D/3D simulation driving vehicle environment as well as hardware instructions to set up a vehicle for further data collection. Clear instructions can be found in the Apollo GitHub project to help install the simulation environment within a Docker environment.
This dataset will be used to boost research on automated-learning tasks such as finding the roads (Drivable Area Segmentation), detecting the objects (Road Object Detection), allowing model generalization for different locations or weather conditions (Domain Adaptation of Semantic Segmentation) and tracking moving objects (Instance-level Video Movable Object Segmentation).
These research tasks make up the Workshop on Autonomous Driving (WAD) Challenge sponsored by Baidu and taking place next June during CVPR 2018, the IEEE International Conference on Computer Vision and Pattern Recognition. The WAD challenge regroups researchers and engineers across academia and industries to discuss computer vision applications in autonomous driving.
According to ArsTechnica, Waymo, the self-driving unit of Google parent company Alphabet, is currently leading the global innovation in autonomous vehicles along with GM, while Baidu is for the time being viewed more as a contender in the automated driving sector. Opening up the ApolloScape dataset could be interpreted as a move by Baidu to weaken Google's data advantage and increase its own relative position in the industry.
To that effect, Baidu further announced it has joined the Berkeley DeepDrive (BDD) Industry Consortium, a top-tier research alliance which includes Ford, NVIDIA, Qualcomm, and General Motors. BDD focuses on innovations in deep reinforcement learning, cross-modal transfer learning applied to autonomous driving.
Baidu has also partnered with Udacity, an online data-science education website, to launch on online course titled Intro to Apollo which is part of Udacity’s nano degree on self-driving cars. The course start date has not yet been set.
KPMG’s 2018 Index on Autonomous Vehicles Readiness ranks China at number 16 in terms of the 20 countries preparedness for an autonomous vehicle future. Baidu is one of three major Chinese autonomous driving companies along with JingChi.ai and Pony.ai.