Temporal Transductive Inference for Few-Shot Video Object Segmentation

Few-shot video object segmentation (FS-VOS) aims at segmenting video frames using a few labelled examples of classes not seen during initial training. In this paper, we present a simple but effective temporal transductive inference (TTI) approach that leverages temporal consistency in the unlabelled video frames during few-shot inference without episodic training. Key to our approach is the use of a video-level temporal constraint that augments frame-level constraints. The objective of the video-level constraint is to learn consistent linear classifiers for novel classes across the image sequence. It acts as a spatiotemporal regularizer during the transductive inference to increase temporal coherence and reduce overfitting on the few-shot support set. Empirically, our approach outperforms state-of-the-art meta-learning approaches in terms of mean intersection over union on YouTube-VIS by 2.5%. In addition, we introduce an improved benchmark dataset that is exhaustively labelled (i.e., all object occurrences are labelled, unlike the currently available). Our empirical results and temporal consistency analysis confirm the added benefits of the proposed spatiotemporal regularizer to improve temporal coherence. Our code and benchmark dataset is publicly available at, https://github.com/MSiam/tti_fsvos/.