ViCE: Self-Supervised Visual Concept Embeddings as Contextual and Pixel Appearance Invariant Semantic Representations

This work presents a self-supervised method to learn dense semantically rich visual concept embeddings for images inspired by methods for learning word embeddings in NLP. Our method improves on prior work by generating more expressive embeddings and by being applicable for high-resolution images. Viewing the generation of natural images as a stochastic process where a set of latent visual concepts give rise to observable pixel appearances, our method is formulated to learn the inverse mapping from pixels to concepts. Our method greatly improves the effectiveness of self-supervised learning for dense embedding maps by introducing superpixelization as a natural hierarchical step up from pixels to a small set of visually coherent regions. Additional contributions are regional contextual masking with nonuniform shapes matching visually coherent patches and complexity-based view sampling inspired by masked language models. The enhanced expressiveness of our dense embeddings is demonstrated by significantly improving the state-of-the-art representation quality benchmarks on COCO (+12.94 mIoU, +87.6\%) and Cityscapes (+16.52 mIoU, +134.2\%). Results show favorable scaling and domain generalization properties not demonstrated by prior work.