Complementary-Stream Sum Fusion for the Semantic Segmentation of the Borobudur Temple Bas Relief Panels Using RGBD and Edge-Depth

Abstract

Semantic segmentation of the Borobudur temple bas-relief panels is difficult due to stone weathering, class imbalance, complex iconography, and limited annotated data. In this paper, we propose a dual-stream fusion framework, which includes an RGBD branch (RGB+depth, 4-channel) and an Edge-Depth branch (softedge+depth, 2-channel), both based on a DeepLabV3+-inspired encoder-decoder with a ResNet-50 backbone. We compare the proposed scheme with state of the art decision-level fusion strategies on the BRSD dataset (248 images, 7 foreground classes), under a strict held-out evaluation protocol. The Complementary-Stream Sum Fusion (CSSF, β = 0.5) achieves the highest test mean Intersection over Union (mIoU) of 0.5220, outperforming the RGBD branch (0.5034), the Edge-Depth branch (0.4251), and all data-calibrated counterparts. The zero-parameter prior beats val-optimized EM (0.5143) and Weighted Fusion (0.5215), consistent with calibration overfitting at small validation sizes (N = 40). Five-fold cross-validation was utilized to exhibit the stability of each branch.