Probabilistic Segmentation and Targeted Exploration of Objects in Cluttered Environments

[JisuHann]

Probabilistic Segmentation and Targeted Exploration of Objects in Cluttered Environments

Probabilistic Approach: by quantifying the expected informativeness of its own actions in information-theoretic terms Using physical interaction with objects through pushing, grasping, or lifting + ICP, Point Cloud etc.

Defining a probability distribution over the complete set of possible segmentation

Repeat

1. Scene
2. Part models
   • by point clouds, track using local key points descriptors
3. Detect movement
   • quantify the informativeness of possible actions
     • the probability that the proposed object is the result of over- or under- segmentation (in this theory)
     • perceptual parameters
     • expected information gain based on discrete sets of possible actions and models based on a physics simulator
     • Q-learning
4. Segmentation based on probabilistic inference
5. Object models