Self-Superflow: Self-Supervised Scene Flow Prediction in Stereo Sequences

Most methods in few-shot learning adopt episodic training, where classes for generating episodes are randomly sampled. Here, most of the episodes are easily solvable, i.e., the dataset in terms of the episodes becomes biased towards easy ones. in this paper, we propose a novel sampling method named Episode Difficulty Based Sampling (EDBS) that aims to remove the dataset bias in terms of episode difficulty. We define the episode difficulty to be proportional to the similarity between the classes composing the episode. Then we determine an episode as easy or hard depending on their episode difficulty and design a balanced episode dataset in terms of the difficulty. Through our EDBS, few-shot networks become less biased to the easy episodes and learn detailed features necessary for solving challenging episodes. Experiments demonstrate that our sampling method is widely applicable and achieves state-of-the-art performance in the benchmarks.