HomeRobotic Database - Robotic platform | TERRINet

The SMILE project aimed at producing a new dataset that facilitates development of indoor localization methods through machine learning and evaluation of such methods. The core idea is that availability of multisensor and accurate data describing highly-structured indoor scenes would allow us to develop new, learning-based methods for scene description with structured geometric features, such as planar segments and edges. The scene description should support the task of agent localization. Obtaining such a description is challenging for passive cameras often applied in personal indoor localization or augmented reality setups. To address the difficulties, we automatically supervise the learning of scene representation with ground truth trajectories from a motion capture system, and depth data obtained from a 3-D laser scanner. Moreover, we embed in the dataset additional localization cues coming from Inertial Measurements Unit (IMU). The facility available at LAAS-CNRS within the TERRINet TNA program made it possible to collect the necessary data.

The result of the SMILE experiment is a large dataset of indoor data that should be useful in our research on data-driven methods for structure-level scene description. The fact that the main outcome of the experiment is a dataset, not direct scientific results, allows us to re-use the gained data many times for different, localization-related research. Moreover, within the TERRINet TNA program, we have gained experience on collecting large datasets with data analysis and generation of optimal experiment scenarios. Our long-term goal is to improve the PlaneLoc system (Wietrzykowski & Skrzypczynski, 2019) by enabling operation with passive vision systems which will make it much more practical for personal localization, large-scale augmented reality applications, and service robots. A proper dataset for learning and evaluation plays an essential role in achieving this goal, and participation in the TERRINet made these plans closer to reality.