Evaluation of 3D Reconstruction for Cultural Heritage Applications

Cristián Llull, Nelson Baloian, Benjamin Bustos, Kornelius Kupczik, Ivan Sipiran, Andrés Baloian; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2023, pp. 1642-1651

In recent years, we have seen the emergence of methods for creating 3D digital reproductions of objects using photos. These techniques, particularly when combined with handheld video devices like smartphones, have significant applications in various fields such as medicine, museology, mechanics, and archaeology. However, previous works often lack an objective assessment of the resulting models' quality. To address this issue, the paper focuses on the systematic evaluation of reconstruction methods. This paper investigates the principles and application of the Chamfer distance, specifically the average, forward, and backward variants, for evaluating reconstructions produced by different methods: Photogrammetry, NeRF, and NVDiffrec. We also explore the impact of background filtering on the reconstructions. The ground truth for comparison is a reconstruction obtained with a structured light scanner, considered the best possible reconstruction with current technology. The results demonstrate that a comprehensive evaluation of reconstruction methods requires considering multiple measures, as they provide information about different aspects of reconstruction quality. By utilizing the Chamfer distance and comparing against the ground truth, we highlight the importance of assessing various aspects when analyzing the performance of different reconstruction methods.