vinairesearch / geoformer Goto Github PK

Hi Tuan Ngo, thanks for your work very much. There are a few questions that confuse me about Geodesic-Former as follows:

1. Do you have the AP@50 performance of Geodesic-Former, PointGroup and DyCo3D on the base classes of ScanNet? The modules in Geodesic-Former are inspirational for general instance segmentation tasks. I am curious how many points can the proposed modules improve on base classes.

2. Will the training on few-shot samples affect the performance on the base classes? I think that the transformer decoder is both used in the base classes and few-shot classes. The training on few-shot classes may change the parameters of transformer decoder dramastically and thus makes it fail on the base classes.

3. From my understanding, the geodesic distance (a float value) is a distance measure between two points. For one point out of N points, which point should the distance be calculated from? How can we embed a float value?

Looking forward to your reply!