SODA10M: A Large-Scale 2D Self/Semi-Supervised Object Detection Dataset for Autonomous Driving

Abstract

Aiming at facilitating a real-world, ever-evolving and scalable autonomous driving system, we present a large-scale benchmark for standardizing the evaluation of different self-supervised and semi-supervised approaches by learning from raw data, which is the first and largest benchmark to date. Existing autonomous driving systems heavily rely on 'perfect' visual perception models (e.g., detection) trained using extensive annotated data to ensure the safety. However, it is unrealistic to elaborately label instances of all scenarios and circumstances (e.g., night, extreme weather, cities) when deploying a robust autonomous driving system. Motivated by recent powerful advances of self-supervised and semi-supervised learning, a promising direction is to learn a robust detection model by collaboratively exploiting large-scale unlabeled data and few labeled data. Existing dataset (e.g., KITTI, Waymo) either provides only a small amount of data or covers limited domains with full annotation, hindering the exploration of large-scale pre-trained models. Here, we release a Large-Scale Object Detection benchmark for Autonomous driving, named as SODA10M, containing 10 million unlabeled images and 20K images labeled with 6 representative object categories. To improve diversity, the images are collected every ten seconds per frame within 32 different cities under different weather conditions, periods and location scenes. We provide extensive experiments and deep analyses of existing supervised state-of-the-art detection models, popular self-supervised and semi-supervised approaches, and some insights about how to develop future models. We show that SODA10M can serve as a promising pretraining dataset for different self-supervised learning methods, which gives superior performance when finetuning autonomous driving downstream tasks. This benchmark will be used to hold the ICCV2021 SSLAD challenge. The data and more up-to-date information have been released at https://soda-2d.github.io.

A Roadmap for Big Model

“…Robustness. Robustness indicates that the stability of big model system can deliver failure-free services [1067,1068,1069,1070,1071,1072,1073,1074], which mainly contains two aspects: the reliability of its internal module and its system coupling, and the generalization ability of big model system for untested data. Therefore, the goal of big model governance's reliability aims to identify the potential problems in its system, scientifically assess the generalization ability of big models, pursue process management, technology migration, and model representation effectively, and guarantee each stakeholder can afford the stable operation level of the model.…”

Memory Replay with Data Compression for Continual Learning

“…For memory replay, we follow Wang et al (2021a) to randomly select the old labeled data and save them in the memory buffer, which indeed achieves competitive performance as analyzed by Chaudhry et al (2018). We follow the implementation of Han et al (2021) for semi-supervised object detection. Specifically, we use Faster- RCNN (Ren et al, 2015) with FPN (Lin et al, 2017) and ResNet-50 backbone as the object detection network for Pseudo Labeling (Han et al, 2021) and Unbiased Teacher (Liu et al, 2021b).…”

“…We follow the implementation of Han et al (2021) for semi-supervised object detection. Specifically, we use Faster- RCNN (Ren et al, 2015) with FPN (Lin et al, 2017) and ResNet-50 backbone as the object detection network for Pseudo Labeling (Han et al, 2021) and Unbiased Teacher (Liu et al, 2021b). For each incremental phase, we train the network for 10K iterations using the SGD optimizer with initial learning rate of 0.01, momentum of 0.9, and constant learning rate scheduler.…”

“…We run each baseline with 8 Tesla V100. The performance is evaluated by the metrics (AP, AP 50 , AP 75 ) used in the prior works (Liu et al, 2021b;Han et al, 2021).…”

“…The advantages of memory replay with data compression are more significant in realistic scenarios such as autonomous driving, where the incremental data are extremely large-scale with huge storage cost. Here we evaluate continual learning on SODA10M(Han et al, 2021), a large-scale object detection benchmark for autonomous driving. SODA10M contains 10M unlabeled images and 20K labeled images of size 1920 × 1080 with default format of JPEG, which is much larger than the scale of ImageNet.…”