Thanks for sharing the codes and it is really helpful! I'm confused about the evaluation metrics. It seems the arccos((Trace(R_pred^T R_gt) - 1) / 2) is used to measure the viewpoint estimation error. However, a more commonly used metric is the geodesic distance |log R_pred^T R_gt|_F / \sqrt{2}. The first one seems to have higher values (i.e. rotation error) than the second one, so the scores of your model might be better if you use the geodesic distance. I'm not sure if I misunderstood anything here. What are your thoughts on this issue?
Thanks for sharing the codes and it is really helpful! I'm confused about the evaluation metrics. It seems the arccos((Trace(R_pred^T R_gt) - 1) / 2) is used to measure the viewpoint estimation error. However, a more commonly used metric is the geodesic distance |log R_pred^T R_gt|_F / \sqrt{2}. The first one seems to have higher values (i.e. rotation error) than the second one, so the scores of your model might be better if you use the geodesic distance. I'm not sure if I misunderstood anything here. What are your thoughts on this issue?