Cristián Llull's Doctoral Qualifying Exam

Topic: A Feature-Aware Chamfer Distance for the Evaluation of Geometric Detail in 3D Reconstruction

Abstract: The geometric evaluation of 3D reconstruction techniques, ranging from traditional photogrammetry to neural implicit representations, remains a significant challenge. Cultural heritage applications require metrics that prioritize histor­ ically significant geometric features over global shape approximation. Standard metrics, such as the Chamfer Distance (CD), are often inadequate because they rely on global error averaging, which effectively masks the loss of high-frequency surface details (carvings, reliefs). To ad­dress this limitation, we introduce a novel benchmarking framework. We propose (1) a controlled, procedurally gen­erated synthetic dataset of Khachkar-inspired stone reliefs, derived from one 2D real-world image of real-world, with per-vertex semantic annotations, enabling controlled eval­uation; and (2) the Weighted Chamfer distance (WCD), a feature-aware metric designed to prioritize structurally sig­nificant geometric features. Evaluating four reconstruction methods (COLMAP, NeuS, Neuralangelo and VGGT), our experiments reveal that traditional CD frequently misranks reconstructions, favoring methods with good global align­ment over those that preserve critical details, while WCD provides archaeologically more plausible assessments. Our framework offers both a methodological advance in evalu­ation and practical guidance for selecting reconstruction approaches in heritage preservation workflows.

Keywords: 3D-reconstruction, Cultural-Heritage, 3D-deep-learning, Benchmarking, Quality Evaluation.

Details

• Advisors: Ivan Sipiran, Nelson Baloian
• Room: Ada Lovelance
• Date: Monday, March 16th at 11:00 AM