Kristofer and I found a paper Learning Implicit Fields for Generative Shape Modeling [https://arxiv.org/pdf/1812.02822.pdf], which is about using the implicit function to do 3D reconstruction. They use a DenseNet-like structure. Unlike DeepSDF, they use a binary output. Given a latent vector and a 3D coordinate, the network will output 0/1 which expresses outside/inside. Its structure is in the following graph.
They apply their decoder to many previous works, like 3DGAN, CNNGAN, etc. And get good results.
Their work is only a decoder, so I tried it on my previous auto-encoder NN for implicit function. And trained it using my own data. It works well on single data but does not enhance a lot. It still cannot solve my encode issues.
Actually, when I was doing the implicit function part, I also tried binary classifier instead of a continuous function. That means, I use a one-hot output, which has 3 entry presenting negative, zero and positive. However, I found it does not work better than continuous output.
