SimNSD: Similar neighborhood-feature constraint for semi-supervised object detection

Abstract

In recent years, semi-supervised learning has been widely used for object detection to enhance model generalization by leveraging information from a large number of unlabeled samples. Semi-supervised object detection methods are generally categorized into two types: consistency-constrained and pseudo-labeled. While consistency-constrained methods improve performance by ensuring consistency between original and augmented images, they often overlook feature relationships within unlabeled images. To address this, we introduce SimNSD, a plugin implementing a neighborhood-feature constraint method. Based on the smoothing assumption of semi-supervised learning, SimNSD applies constraints when similarity thresholds are met. This approach facilitates smooth learning between central domain samples and their neighbors, enhancing network generalization. Our experiments show that SimNSD compensates for the limitations of traditional consistency-constrained methods and significantly improves performance compared to other semi-supervised object detection approaches.