Learning 3D Shape Fidelity Metric from Real-world Distortions

Figure 1. (a) Chamfer Distance [4]: Small changes in details do not lead to large differences in Chamfer Distance, yet they significantly affect the perceived fidelity. (b) SAUCD [30]: Minor alterations in the frequency domain can also result in notable changes in fidelity, indicating that frequency domain metrics alone cannot fully capture human perception. (c) Our learnable metric: We learn human perception of fidelity using a real-world reconstruction/generation dataset annotated by humans.