Lets return to the original problem and take a look at the diagram from the article In the case of compute tomography we had l data and projections of them. In Neural Recon we have tl feature maps and want to get their threeimensional. Projections so we do everything the same only we use voxels as projecte elements of the D scene. Details of the implementation of both back projection and the entire neural. Network as a whole can be found in the official repository.
Next comes Coarse To
Fine a block for obtaining the TSDF of an existing FBV using the method of sequential improvements Coarse To Fine. Coarse To Fine RNN The final stage looks like this Using the existing volumetric feature fields F tiobtaine after. Back projection we iteratively build the TSDF Belgium Phone Number List of the original FBV using a recurrent neural network S_tl. Subsequently it is integrate into the overall reconstruction S_tg which completes the cycle of the neural network. It remains to deal with the RNN device namely GRU Fusion. GRU Fusion.
Schemes are our everything
Lets start with them again What GRU is in this case is clear it is a block for taking into account the hidden state of the current fragment of the entire scene which is built South Africa Whatsapp Number List and complete during the operation of the neural network. You can read more about GRU and other RNNs here . Lets now move on to Sparse Conv this is another optimization which is a convolution on sparse data which is what a voxel grid is. Details also follow the link . At the output from GRU Fusion we receive an update hidden state of the fragment H_tl which we replace in a similar section of the global hidden state.
