Optimizing the Accuracy and Compactness of Multi-view Reconstructions

Current evaluation metrics and benchmarks for multi-view stereo reconstruction methods mainly focus on measuring the accuracy and completeness and they do not explicitly measure the compactness, and especially the compactness-accuracy trade-off of the reconstructed models. To answer this issue, we present an evaluation method that completes and improves the existing benchmarks. The proposed method is capable of jointly evaluating the accuracy, completeness and compactness of a three-dimensional reconstruction which is represented as a triangle mesh. The evaluation enables the optimization of both the whole reconstruction pipeline from multi-view stereo data to a compact mesh and the mesh simplification. The method takes the ground truth model and the reconstruction as input and outputs an accuracy and completeness value as well as the compactness measure for the reconstructed model. The values of the evaluation measures are independent of the scale of the scene, and therefore easy to interpret.