IROS 2016: Simultaneous Place Learning and Recognition

This video illustrates the strength of our proposed scene learning and recognition method for real-time appearance-based mapping [1]. The experiment is performed on KITTI test sequence 02.
It demonstrates a growing network of self-organizing neurons (top right panel) which adapts itself to the topological representation of the input space. For online place recognition, i.e. loop-closure detection, the maximum a posteriori is estimated (bottom right panel) over the learned network.
The approach has several advantages:
a) feature space is learned online rather than offline phase of training
b) no a priori information of the environment is used
c) the places are associated to the clusters of neurons in a learned network rather than place-to-place association. This allows generalization such that a neuron represents a group of places having similar perceptual properties.
See the video contents for details on experimental setup.
[1] S.M.A.M. Kazmi and B. Mertsching.
Simultaneous Place Learning and Recognition for Real-time Appearance-based Mapping.
Accepted for: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2016), 2016.