Contrastive Self-supervised Representation Learning Using Synthetic Data

Learning discriminative representations with deep neural networks often relies on massive labeled data, which is expensive and difficult to obtain in many real scenarios. As an alternative, self-supervised learning that leverages input itself as supervision is strongly preferred for its soaring performance on visual representation learning. This paper introduces a contrastive self-supervised framework for learning generalizable representations on the synthetic data that can be obtained easily with complete controllability. Specifically, we propose to optimize a contrastive learning task and a physical property prediction task simultaneously. Given the synthetic scene, the first task aims to maximize agreement between a pair of synthetic images generated by our proposed view sampling module, while the second task aims to predict three physical property maps, i.e., depth, instance contour maps, and surface normal maps. In addition, a feature-level domain adaptation technique with adversarial training is applied to reduce the domain difference between the realistic and the synthetic data. Experiments demonstrate that our proposed method achieves state-of-the-art performance on several visual recognition datasets.