Tsukada Lab. Computer Network Group

Reinforcement Learning Based Optimal Camera Placement for Depth Observation of Indoor Scenes

Yichuan Chen

Master Thesis, Graduate School of Information Science and Technology, The University of Tokyo (Esaki Lab), Sep 30, 2021

Summary in three lines

  • The thesis defines RL-OCP, a system that moves indoor depth cameras to improve its observations as it learns about a scene.
  • It develops a simulator that uses shadow maps to estimate visibility, coverage and depth error, and learns camera movements with Soft Actor-Critic.
  • In ten test scenes, the method lowered error in seven and kept total depth error below 90% of the offline baselines.

Figures from the paper

Figure 4.1: Overview of the training routine
Figure 4.1. Overview of the training routineSource: Figure 4.1 from Master's thesis, The University of Tokyo, 2021 (author's version)The agent and the simulation environment exchange camera movements, observations of known point clouds and rewards during reinforcement learning.
Figure 6.1: Depth observation error of each approach in each evaluation scene
Figure 6.1. Depth observation error of each approach in each evaluation sceneSource: Figure 6.1 from Master's thesis, The University of Tokyo, 2021 (author's version)The bars compare the proposed method with two baselines (TDSA and BPSO) across ten scenes; the proposed method has the smallest total error.

How to cite

Y. Chen, “Reinforcement Learning Based Optimal Camera Placement for Depth Observation of Indoor Scenes,” in Master Thesis, Graduate School of Information Science and Technology, The University of Tokyo (Esaki Lab), 2021.

BibTeX

@mastersthesis{Chen2021b,
  title         = {Reinforcement Learning Based Optimal Camera Placement for Depth Observation of Indoor Scenes},
  author        = {Yichuan Chen},
  year          = {2021},
  date          = {2021-09-30},
  urldate       = {2021-09-30},
  school        = {Master Thesis, Graduate School of Information Science and Technology, The University of Tokyo (Esaki Lab)},
  keywords      = {SDM, Thesis},
  pubstate      = {published},
  tppubtype     = {mastersthesis}
}