Balancing Bounding Box and Mask Annotations for Semi-Supervised Instance Segmentation

Instance segmentation models are crucial for precise object detection but often require expensive pixel-wise mask annotations.This paper studies the impact of combining bounding box and mask annotations in semi-supervised segmentation.We propose a method that leverages from both types of labeled data within a unified training framework.Through experiments on YOLO (convolution-based) and DETR (transformer-based) architectures, we demonstrate that balancing these annotation types significantly enhances performance while reducing labeling costs, particularly in terms of manual annotation time.Additionally, we evaluate few-shot and zero-shot scenarios, further highlighting the flexibility and efficiency of our method for budget-constrained segmentation tasks.