Custom CDGNet Architecture for Precise Human Part Semantic Segmentation

Human body part segmentation is a task that involves labelling the pixels in an image with their corresponding body part classes. Since the human body has a hierarchical structure, and each part has a unique location, a technique called sample class distribution was developed to improve accuracy. This technique involves collecting and applying primary human parsing labels in both vertical and horizontal dimensions to exploit the position distribution of the classes. Combining these guided features generates a spatial guidance map and adds to the backbone network. To evaluate the effectiveness of this approach, extensive experiments were conducted on a large dataset called CIHP, and the mean IOU, pixel accuracy, and mean accuracy metrics were used for evaluation. The proposed model (custom CDGNet) outperformed the baseline model and adjacent state-of-the-art techniques, achieving a 1.95% increase in pixel accuracy, 0.20% in mean IOU, and a 1.37% increase in weighted mean accuracy. This suggests that the sample class distribution technique may be valuable for improving human body part segmentation accuracy in images.