Learning Correspondence from the Cycle-Consistency of Time
It addresses the problem of reducing annotation costs for visual correspondence tasks in computer vision, offering a competitive self-supervised approach.
The paper tackles learning visual correspondence from unlabeled video by using cycle-consistency in time as a self-supervised signal, resulting in a representation that generalizes across tasks like video object segmentation and keypoint tracking without finetuning, outperforming previous self-supervised methods and competing with supervised ones.
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.