In the function get_generator() in file src/conv_onet/config.py, the parameters are initialized as: recep_field=2*(cfg['model']['encoder_kwargs']['unet3d_kwargs']['num_levels'] + 2), grid_reso = 25 + recep_field - 1, query_vol_size = 25 unit_size.
Why the reception field is calculated like this? Where does the 25 comes from? In my understanding, query_vol_size and input_vol_size define the size of space that each prediction will be operated in but their relationship with the reception field and 25 is unclear.
Also, is unit_size's physical meaning removed since the queried result points are sampled on a 3D grid with interval (query_crop_size/nx<--25*unit_size/self.resolution0)? (in config/pointcloud_crop/demo_matterport.yaml, a comment says that unit_size define the size of a voxel, in meter)
In the function get_generator() in file src/conv_onet/config.py, the parameters are initialized as: recep_field=2*(cfg['model']['encoder_kwargs']['unet3d_kwargs']['num_levels'] + 2), grid_reso = 25 + recep_field - 1, query_vol_size = 25 unit_size.
Why the reception field is calculated like this? Where does the 25 comes from? In my understanding, query_vol_size and input_vol_size define the size of space that each prediction will be operated in but their relationship with the reception field and 25 is unclear.
Also, is unit_size's physical meaning removed since the queried result points are sampled on a 3D grid with interval (query_crop_size/nx<--25*unit_size/self.resolution0)? (in config/pointcloud_crop/demo_matterport.yaml, a comment says that unit_size define the size of a voxel, in meter)