Question about Experiments on NOCS-REAL 275 dataset about caspr HOT 1 CLOSED

Hello, thanks for your interest.

Q1/Q3) We did not perform experiments on any real-world data due to the lack of readily-available data. I have not used the datasets in [1], however I think NOCS-REAL is rather small (compared to the synthetic data we used in CaSPR) and intended for fine-tuning to improve the sim to real gap. This means we would likely also need some synthetic data to fully train CaSPR; and I'm not certain but it looks like the CAMERA dataset [1] is just images, not videos. Besides NOCS-REAL, I am not aware of other real-world datasets with dense NOCS labels, which is why we discussed this as a limitation.

Q2) GT-NOCS are the 3D canonical points for a single frame, whereas ground truth T-NOCS (that we use for training) is a sequence of NOCS where each point also has a timestamp, so it is 4D. But if you were to remove the timestamp, then yes you can see GT-TNOCS as the union of a sequence of GT-NOCS.

from caspr.