[3K4-IS-2a] Machine learning

[3K4-IS-2a-05] Comparative Analysis of Data Augmentation Methods for Semantic Segmentation and Multi-label Classification

〇Yuto Kohata¹, Yuna Park², Masaki Onishi³ (1. Diablo Valley College, 2. University of Tsukuba, 3. National Institute of Advanced Industrial Science and Technology)

[[オンライン]]

キーワード：Data Augmentation, Deep Learning, Semantic Segmentation, Multilabel Classification

Data augmentation is widely recognized as an effective strategy for enhancing the generalization performance of deep neural networks. To date, numerous studies have systematically evaluated various augmentation techniques for image classification, and automatic augmentation methods have been developed. In contrast, the effectiveness of data augmentation for semantic segmentation has not been sufficiently validated compared to classification. Since classification captures global features while semantic segmentation requires fine-grained, pixel-level details, the critical information for each task differs. Therefore, designing data augmentation strategies is essential to tailor to the specific task. This study provides a comprehensive evaluation of 18 diverse augmentation techniques, including both color-based and geometric-based augmentations under comparable conditions using the PASCAL VOC dataset. The evaluation is performed using mean Average Precision (mAP) for multi-label classification and mean Intersection over Union (mIoU) for semantic segmentation.
The results indicate that, when compared to the same dataset, effective augmentation techniques vary depending on the specific task. Moreover, certain methods exhibit dataset-dependent characteristics.