Probabilistic Plant Modeling via Multi-View Image-to-Image Translation

task: infer 3D branch structures from multi-view observations

• infer prob. of branch using Bayesian i2i, then generates a prob. plant 3D model
  • why prob? access to credibility of estimates, consolidate view dependent inferences
• previous methods
  • branching model of trees: low deviation from presumed models
  • BNN: in this work, MC via dropout
• method
  • Pix2Pix, with dropout layer inserted
  • take the mean of stochastic samples as estimation of 2D prob of being a branch, multipliing into 3D prob.
  • branch generation using particle flows
    • generates particles propto log prob. of 3D prob, with root set to bottonmost point
    • particle pos. updated by three forces
      • F_root
      • F_c: toward to branch prob.
      • F_d: parallel to branch prob.
    • unification of particles
  • smoothing, refinement, simplification