Replication Data for: A comparison of deep learning-based visual odometry algorithms in challenging scenarios

This is a dataset used in the experiments for the paper "A comparison of deep learning-based visual odometry algorithms in challenging scenarios." The experiments were conducted using data collected by a Quanser QCAR equipped with a Realsense D435 camera. We only provide the RGB information with a resolution of 1920x1080 pixels. The robot's ground-truth pose information was gathered through an external Vicon motion capture system.

There are three sequences, sequence 1 is very simple as it involves no curves. The robot traveled approximately 3.13 meters in 15.5 seconds (103 frames) in this sequence. Sequence 2 contains a curve and the robot traveled for a distance of 7.73 meters in 40 seconds (242 frames). Sequence 3 is the most challenging one as it includes three curves. In this sequence, the robot traveled for a distance of 17.40 meters, taking about 66.6 seconds to complete the path (408 frames).

Additionally, for each sequence, there are 5 types of emulated failures: camera overexposure, camera underexposure, lens condensation, lens breakage, and lens dirt.