To study the The authors of the original article realize that using transformers in D on a full voxel grid would lead to a multiple increase in the number of values for attention. Their total number will be calculate according to the following formula where _zis the number of voxels along the and z axes respectively. In the case of we find Nthat at Nwill give about one petabyte of data The solution to the problem considere is that we will calculate the attention value for a given voxel using only nonempty voxels locate within a fixe window as demonstrate in the figure.
At the same time we also consider
attention for the neighbors of fille voxels dialate voxels shade in yellow. This is necessary to preserve the completeness of the geometric structure of the reconstructe surface which is usually lost after downsample blocks with D convolutions. Now we are ready to look USA Phone Number List at the complete diagram of one D SDF Former and complete the analysis of the model Whats going on here Lets get a look First we take a volumetric feature map as input V run it through Sparse D CNN thereby obtaining a reuce feature map Down. Then we concatenate it with the calculate.
Dilate Attention and transfer
It all together to D Sparse Window Attention. We repeat the proceure describe above times in a row. It would seem that lets now expand the selecte features to USA Email List TSDF but not everything is so simple. A problem arises due to the fact that all our selecte features are quite local and limite to a given window. There is a socalle loss of global relationships within the scene or small receptive field.
