Huawei Releases ‘SODA10M’, A Large-Scale 2D Dataset For Object Detection In Autonomous Driving With 10M Unlabeled Images And 20K Labeled Images Over 6 Classes

783
Source: https://arxiv.org/pdf/2106.11118.pdf

Autonomous driving technology has the potential to save lives, but it must be engineered with safety in mind. The visual perception system is critical because object detection plays such a great role for autonomous cars on roads and highways.

The currently available datasets may limit the performance of current object detection approaches. For example, the largest self-driving dataset in existence is Waymo Open, and it was collected from only three cities which cover a few scenarios or circumstances. Models trained on these datasets may overfit specific features of a scenario. Another limitation includes the limited size of existing datasets is a hindrance to autonomous driving research. With so much information that needs labeling, data annotation becomes an expensive endeavor. 

To make autonomous driving systems more realistic, Huawei researchers develop one of the most extensive Scale Object Detection benchmarks for Autonomous Driving (SODA10M), containing 10 million road images. This SODA10M dataset can be distinguished from existing datasets in three ways: scale, diversity, generalization.

Scale

In terms of ‘scale‘, SODA10M is ten times larger than existing autonomous driving datasets like BDD100K and Waymo, containing 10 million images of road scenes. With 20 thousand tightly fitting high-quality 2D bounding boxes, SODA10M has significantly more data for self-driving cars to learn from.

Diversity

In terms of ‘diversity‘, The pictures in the SODA10M dataset cover four different seasons and 32 cities. The photos were taken under a variety of conditions, such as during rainstorms or snowfalls, to see how self-driving cars would respond to various circumstances.

Generalization

In terms of ‘generalization,’ The SODA10M dataset provides a superior set of data that will be used for pre-training autonomous driving algorithms. This is due to the diversity and size of the dataset, which results in better generalization ability than other existing datasets like Waymo or Cityscapes when using MoCov1.d discounts.

https://arxiv.org/pdf/2106.11118.pdf

The SODA10M dataset is 10x larger than the largest available self-driving dataset and was collected from cities across diverse weather conditions, periods, and scenes. With this data set, you can get high-quality labeled data with a large amount of unlabeled driving footage in much more diversity.

The research paper shows that SODA10M can serve as a promising dataset for training and evaluating different self/semi-supervised learning methods. This could promote the exploration of advanced techniques to help move autonomous driving systems forward, in addition to being standardized evaluations.

Paper: https://arxiv.org/pdf/2106.11118.pdf

Dataset: https://soda-2d.github.io/index.html

Project: https://sslad2021.github.io/index.html

Challenges: https://sslad2021.github.io/pages/challenge.html