Training-Free Dataset Pruning for Instance Segmentation

2025-03-02

AI Summary

Existing dataset pruning methods are primarily designed for image classification and struggle to adapt to instance segmentation—due to its pixel-level annotations, large inter-instance scale variations, and severe class imbalance, which collectively exacerbate pruning difficulty; moreover, most rely on model training, resulting in poor efficiency. This paper introduces the first training-free pruning framework tailored for instance segmentation. We propose a Shape Complexity Score (SCS) grounded in geometric shape analysis, and further develop two variants: Scale-Invariant SCS (SI-SCS) and Class-Balanced SCS (CB-SCS), eliminating dependence on model training. Our method computes instance-level scores solely from image-level labels and geometric attributes (e.g., contour compactness, aspect ratio), enabling highly efficient pruning. Evaluated on PASCAL VOC 2012, Cityscapes, and COCO, it achieves state-of-the-art pruning performance while maintaining compatibility with both CNN- and Transformer-based architectures. Pruning speed is accelerated by an average factor of 1349× over training-dependent baselines.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Multi-instance/Multi-view Learning
• Constraint Satisfaction and Optimization: Satisfiability Modulo Theories

Application Category

• Web Mining and Content Analysis: Large pretrained models with web data
• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines

Abstract

Existing dataset pruning techniques primarily focus on classification tasks, limiting their applicability to more complex and practical tasks like instance segmentation. Instance segmentation presents three key challenges: pixel-level annotations, instance area variations, and class imbalances, which significantly complicate dataset pruning efforts. Directly adapting existing classification-based pruning methods proves ineffective due to their reliance on time-consuming model training process. To address this, we propose a novel Training-Free Dataset Pruning (TFDP) method for instance segmentation. Specifically, we leverage shape and class information from image annotations to design a Shape Complexity Score (SCS), refining it into a Scale-Invariant (SI-SCS) and Class-Balanced (CB-SCS) versions to address instance area variations and class imbalances, all without requiring model training. We achieve state-of-the-art results on VOC 2012, Cityscapes, and COCO datasets, generalizing well across CNN and Transformer architectures. Remarkably, our approach accelerates the pruning process by an average of 1349$ imes$ on COCO compared to the adapted baselines. Source code is available at: https://github.com/he-y/dataset-pruning-for-instance-segmentation

Problem

• Address dataset pruning for instance segmentation tasks
• Overcome challenges like pixel-level annotations and class imbalances
• Develop training-free method to accelerate pruning process

Innovation

• Training-Free Dataset Pruning (TFDP) method
• Shape Complexity Score (SCS) for instance segmentation
• Scale-Invariant and Class-Balanced SCS versions