Contrastive Self-supervised Representation Learning Using Synthetic Data

Abstract 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 strongly preferred for its soaring performance visual representation learning. This paper introduces a contrastive framework generalizable the synthetic data can be obtained easily complete controllability. Specifically, we propose optimize task physical property prediction simultaneously. Given scene, first aims maximize agreement between pair of images generated by our proposed view sampling module, while second predict three maps, i.e., depth, instance contour surface normal maps. In addition, feature-level domain adaptation technique adversarial training applied reduce difference realistic data. Experiments demonstrate method achieves state-of-the-art several recognition datasets.