Active Vision

Project Work


Visual Memory Aggregation : Project

Depiction of working of the model on single task

Meta Task : Several Rooms from different Environments

This projects aims to solve the following problem:

  • Given Data from Tasks T1-Tn in the training set learn to do Task Tn+1 from test/train with a few samples from it.

  • Each Task Ti is as follows : given an environment sampled from distribution Ti (A particular kind of room) learn to create the map of a property(Where all the objects are places/ what are the safe places to land) of the whole room with a few images from the room.

  • Object detection is done using images from Airsim and Safe mapping is done using images from Gazebo.

Pink Boxes are Ground Truth

Grid Visual Memory

Depicted : Update and Decoding

Reconstruction with random sampling

ppo_rendered.mp4

Reconstruction with sampling using PPO policy

The Safe Mapping Task : ROS-Gazebo Environment

Paper Summaries


Copy of On the convergence theory of MAML : Summary and Review