Stable Single-Pixel Contrastive Learning for Semantic and Geometric Tasks
This addresses the need for joint semantic and geometric representation learning in computer vision, though it appears incremental as it builds on contrastive learning methods.
The authors tackled the problem of learning pixel-level representations that capture both semantic and geometric information by introducing a family of stable contrastive losses, which enable precise point-correspondence across images without momentum-based teacher-student training, as demonstrated in synthetic 2D and 3D environments.
We pilot a family of stable contrastive losses for learning pixel-level representations that jointly capture semantic and geometric information. Our approach maps each pixel of an image to an overcomplete descriptor that is both view-invariant and semantically meaningful. It enables precise point-correspondence across images without requiring momentum-based teacher-student training. Two experiments in synthetic 2D and 3D environments demonstrate the properties of our loss and the resulting overcomplete representations.