The project is intended to simplify the process of generation and correction of indoor localization datasets. Also included is a support tool for estimating relative pose using DeMoN using RGB image pairs from a given dataset.
Part_1: Floorplan tools - openCV based code to mark lables on the floorplan. Output is an edited png file with the marked labels and a text file with the correspinding coordinated on the map as well as the relative distances betwenn consecutive points in the ground frame. Also included is a tool for plotting camera centres on the floorplan image for better visualization.
Part_2: Relative Pose estimation - Using DeMoN: Depth and Motion Network for Learning Monocular Stereo (https://arxiv.org/abs/1612.02401) to estimate relative pose between pairs of RGB (png) images only. The output is a binary file containing the estimated rotation and translation vectors. Format <rotation-quaternion (real x y z)> <translation 3D>
Part_3: Automated Dataset Generation - This tool is meant to simplify the process of creating indoor localization datasets using a Google Tango device for odometry and Point Clouds and iPhone for hi-res RGB images mounted together.
Part_4: Odometry Correction - For correcting the raw Tango odometry using Pose Graph optimization for loop closure. Anchor points are chosen from the floorplan (Part_1) which serve as absolute Ground truths for correcting the complete path.
Deatiled installation and usage instructions are given separetely for each part.
- Surya Kalia - Summer Project Intern
- Zakaria Laskar - Supervisor
- Juho Kannala - Professor
Computer Vision Lab, Aalto Univesity, Espoo, Finland.