Figure 8
An overview of our attenuation field network (AFN). We map a coordinate vector (a) through a fully connected network (b) to the respective attenuation coefficient (c). The training procedure of AFN coincides with the conventional ray-tracing algorithm. We sample points along a ray path to render the projection intensity at the ray end according to the Beer-Lambert’s law. This rendering procedure is fully differentiable, allowing us to optimize our attenuation representations by minimizing the error between the synthesized projection and the acquired projection. The training is done until we iterate the rendering and optimization steps over all the acquired data.
