Label-Efficient Segmentation for Diverse Scenarios

Abstract

Segmentation, a fundamentally important task in computer vision, aims to partition an image into multiple distinct and meaningful regions or segments. In this thesis, we analyze the importance label-efficient segmentation techniques and provide a series of methods to address segmentation tasks in different scenarios. First, we propose a Deep Reasoning Network for few-shot semantic segmentation, termed as DRNet, which is a novel approach that relies on dynamic convolutions to segment objects of new categories. Unlike previous works that directly apply convolutional layers to integrated features to predict segmentation masks, our DRNet generates learnable parameters for predicting layers based on query features, allowing for greater flexibility and adaptability. Second, we conduct further experiments and propose mining both dynamic and regional context, termed as DRCNet, for few-shot semantic segmentation. Specifically, we introduce a Dynamic Context Module to capture spatial details in the query images, and a Regional Context Module to model the prototypes for ambiguous regions while excluding background and ambiguous objects in query images. The superior performance of our method is demonstrated on various benchmarks. Third, we address the unsupervised video object segmentation task by learning both motion and temporal cues, in a method termed as MTNet. The proposed MTNet integrates appearance and motion information through a Bi-modal Feature Fusion Module and models the relations between adjacent frames using a Mixed Temporal Transformer. Achieving state-of-the-art results on multiple datasets while maintaining a much faster inference speed. Finally, we propose a semi-supervised learning method for bird’s-eye-view (BEV) semantic segmentation, which represents the first attempt at performing label-efficient learning in this field.Without any whistles-and-bells, our proposed BEV-S4 can achieve results on par with fully-supervised methods while requiring significantly fewer labels. We hope that our approach could serve as a strong baseline and potentially attract more attention to learning BEV perception with fewer labels.