The rapid advancement of deep learning (DL) methods for point cloud processing has significantly increased the demand for large, diverse, and annotated datasets to improve model performance across various applications. In the cultural heritage domain, the availability of such datasets is crucial for driving innovation in algorithm development. However, a notable gap exists due to the limited availability of large-scale, labelled point cloud datasets specific to heritage structures especially within the Indian context. This study introduces 3DITA—the 3D Indian Temple Architecture Dataset—India’s first benchmark point cloud dataset tailored for semantic segmentation of Nagara-style temple architecture. The dataset comprises over 325 million points, captured from 47 temple structures using Close-Range Photogrammetry (CRP) and Terrestrial Laser Scanning (TLS). A total of 22,370 photographs were used to reconstruct 46 temples via CRP, while TLS was employed for one site, ensuring high-resolution and geometrically rich data. To evaluate the dataset, a deep learning-based segmentation framework was implemented using PointNet and DGCNN architectures. The models were trained to segment culturally specific classes including Walls, Mandapa, Shikhara, Amalaka, and Garbhagriha. The DGCNN model achieved a peak accuracy of 80%, outperforming the PointNet and demonstrating the dataset's robustness in handling the geometric complexity of Indian heritage structures. Beyond semantic segmentation, 3DITA serves as a foundational resource for a range of applications, including heritage reconstruction, digital twin development, Historic Building Information (HBIM) modelling, and large-scale heritage preservation. By making the dataset available upon request, this study aims to support future research and foster interdisciplinary collaboration in AI-driven cultural heritage documentation and analysis.