Robustness of sparse instance segmentation models under synthetic distribution shifts

Deep learning-based instance segmentation achieves high accuracy in controlled settings but often fails under synthetic visual degradations, particularly in industrial applications characterized by clutter, variable lighting, and limited computational resources. Data augmentation during training and model pruning are two well-known techniques for improving the robustness and generalizability of a segmentation model. While these techniques have been studied individually, they are evaluated independently in this work to assess their respective benefits. This work presents a systematic evaluation of pruning strategies for state-of-the-art instance segmentation models under seven industry-relevant corruption types, including motion blur, fog, and contrast shifts. Using a steel scrap recycling dataset, we evaluate performance across object materials, scales, and corruption severities. Our evaluation uncovers various degrees of segmentation performance degradation for different corruption types. More importantly, we show that pruning compresses the model by 90% and preserves, or in specific cases moderately improves, segmentation performance on degraded image data, though vulnerabilities to specific shifts like contrast loss remain. Based on these findings, this work suggests specific data augmentation and model strategies for industrial use cases.