Learning Correspondence from the Cycle-Consistency of Time
read the original abstract
We introduce a self-supervised method for learning visual correspondence from unlabeled video. The main idea is to use cycle-consistency in time as free supervisory signal for learning visual representations from scratch. At training time, our model learns a feature map representation to be useful for performing cycle-consistent tracking. At test time, we use the acquired representation to find nearest neighbors across space and time. We demonstrate the generalizability of the representation -- without finetuning -- across a range of visual correspondence tasks, including video object segmentation, keypoint tracking, and optical flow. Our approach outperforms previous self-supervised methods and performs competitively with strongly supervised methods.
This paper has not been read by Pith yet.
Forward citations
Cited by 1 Pith paper
-
Non-Parallel Voice Conversion with Cyclic Variational Autoencoder
CycleVAE optimizes non-parallel voice conversion indirectly via cyclic reconstructed spectra, yielding higher spectral accuracy, latent feature correlation, and improved converted speech quality.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.