During the training of DeepLS, the code reshapes the points sampled from different object instances into a single sample (here) and builds the KD tree accordingly (here). Thus, when we find near_sample_indices (here), for each voxel grid of an object instance, the code actually searches for points sampled on all object instances in the batch. This is not correct, because different instances are not correlated.
During the training of DeepLS, the code reshapes the points sampled from different object instances into a single sample (here) and builds the KD tree accordingly (here). Thus, when we find near_sample_indices (here), for each voxel grid of an object instance, the code actually searches for points sampled on all object instances in the batch. This is not correct, because different instances are not correlated.