The GIST and RIST of Iterative Self-Training for Semi-Supervised Segmentation

TL;DR

This paper investigates iterative self-training for semi-supervised segmentation, identifies degradation issues with naive approaches, and proposes GIST and RIST strategies that improve performance by alternating training data sources.

Contribution

The paper introduces GIST and RIST, novel strategies for iterative self-training that prevent performance degradation and enhance semi-supervised segmentation.

Findings

Naive iterative self-training causes performance degradation.

GIST and RIST strategies improve segmentation accuracy.

Combining GIST and RIST with existing methods boosts results.

Abstract

We consider the task of semi-supervised semantic segmentation, where we aim to produce pixel-wise semantic object masks given only a small number of human-labeled training examples. We focus on iterative self-training methods in which we explore the behavior of self-training over multiple refinement stages. We show that iterative self-training leads to performance degradation if done na\"ively with a fixed ratio of human-labeled to pseudo-labeled training examples. We propose Greedy Iterative Self-Training (GIST) and Random Iterative Self-Training (RIST) strategies that alternate between training on either human-labeled data or pseudo-labeled data at each refinement stage, resulting in a performance boost rather than degradation. We further show that GIST and RIST can be combined with existing semi-supervised learning methods to boost performance.