Temporal Interpolation as an Unsupervised Pretraining Task for Optical Flow Estimation

TL;DR

This paper proposes using unsupervised temporal interpolation as a pretraining task for optical flow estimation, improving performance over supervised training on synthetic data by leveraging real videos.

Contribution

It introduces a novel unsupervised pretraining approach via frame interpolation for optical flow estimation, enhancing accuracy with minimal ground truth data.

Findings

Unsupervised pretraining outperforms supervised training on synthetic data.

Fine-tuning with small ground truth data improves flow estimation in untextured regions.

The method effectively estimates optical flow using real video data.

Abstract

The difficulty of annotating training data is a major obstacle to using CNNs for low-level tasks in video. Synthetic data often does not generalize to real videos, while unsupervised methods require heuristic losses. Proxy tasks can overcome these issues, and start by training a network for a task for which annotation is easier or which can be trained unsupervised. The trained network is then fine-tuned for the original task using small amounts of ground truth data. Here, we investigate frame interpolation as a proxy task for optical flow. Using real movies, we train a CNN unsupervised for temporal interpolation. Such a network implicitly estimates motion, but cannot handle untextured regions. By fine-tuning on small amounts of ground truth flow, the network can learn to fill in homogeneous regions and compute full optical flow fields. Using this unsupervised pre-training, our network outperforms similar architectures that were trained supervised using synthetic optical flow.