Monte Carlo Scene Search for 3D Scene Understanding

MonteBoxFinder: Detecting and Filtering Primitives to Fit a Noisy Point Cloud

“…While exploring the solution tree with MCTS as done in [13] is efficient, we show we can still speed up the search for a solution significantly more. The tree structure imposes an ordering of the possible 3D models to pick from.…”

“…For example two primitives should not intersect. To tackle this problem, we take inspiration from a recent work on 3D scene understanding [13]. [13] proposes to rely on the Monte Carlo Tree Search (MCTS) algorithm to handle a similar combinatorial problem to select objects' 3D models: The MCTS algorithm is probably best known as the algorithm used by AlphaGo [29].…”

“…To tackle this problem, we take inspiration from a recent work on 3D scene understanding [13]. [13] proposes to rely on the Monte Carlo Tree Search (MCTS) algorithm to handle a similar combinatorial problem to select objects' 3D models: The MCTS algorithm is probably best known as the algorithm used by AlphaGo [29]. It is typically used to explore the tree of possible moves in the game Go because it scales particularly well to high combinatorics.…”

“…It is typically used to explore the tree of possible moves in the game Go because it scales particularly well to high combinatorics. [13] adapts it to 3D models selection by considering a move as the selection of a 3D model for one object, and showed it performs significantly better than the simple hill-climbing algorithm that is sometimes used for similar problems [34]. Another advantage of this approach is that it does not impose assumptions on the form of the objective function, unlike other approaches based on graphs, for example [26].…”

“…Comparative overview The hill-climbing algorithm-simply taking the primitive that improves the most the objective function-can terminate × quickly as it gets stuck into a local minimum because of the constraints between primitives. MCTS as used in [13] explores iteratively the solution tree by traversing blue paths, updating which primitives are the most promising ones, but keeping the tree structure fixed. At each iteration, our approach also updates (→) which primitives are the most promising ones, and starts with them.…”