Temporally stable video segmentation without video annotations

Temporally consistent dense video annotations are scarce and hard to collect. In contrast, image segmentation datasets (and pre-trained models) are ubiquitous, and easier to label for any novel task. In this paper, we introduce a method to adapt still image segmentation models to video in an unsupervised manner, by using an optical flow-based consistency measure. To ensure that the inferred segmented videos appear more stable in practice, we verify that the consistency measure is well correlated with human judgement via a user study. Training a new multi-input multi-output decoder using this measure as a loss, together with a technique for refining current image segmentation datasets and a temporal weighted-guided filter, we observe stability improvements in the generated segmented videos with minimal loss of accuracy.

1. Introduction

There are two main challenges in semantic video segmentation: accuracy and temporal consistency. While the accuracy challenge is essentially shared with the corresponding task for still images, temporal consistency/stability is a challenge unique to video footage. These two challenges are interconnected; for example, perfectly accurate segmentation masks are, by definition, also perfectly consistent. The opposite, however, doesn’t always hold; e.g. a sequence of empty masks is perfectly consistent, but usually far from accurate.