NFS dataset

[alexr]

Dear Mathiue,

Can you describe how exactly did you use the NFS dataset in your original paper on RTABMAP?
I wish to use similar dataset for training and machine learning applications. How did you perform the calibration and saved all the images from the drive and used as ROS dataset? Does recent NFS editions be used as well with weather information.

Thanks
Alex

[matlabbe]

Hi Alex,

The dataset has been done with a screen recorder, I don't remember the name (fraps?), it was on Windows, one that is widely used for recording video games. I think I could configure it to record 1 frame each second and save the resulting images in a folder. There was no calibration, that dataset was used only for appearance-based loop closure detection evaluation, there was no metric mapping.

https://www.youtube.com/watch?v=kghs6XM8Yzw

It could be indeed nice to recover the opengl/directx camera intrinsics and set the game in stereo mode to record left/right images. With that data, we could do SLAM with RTAB-Map. Note for the stereo camera calibration, it could be guessed (set focal distance and baseline arbitrarily), as there should not have any distortions (the map could be rescaled afterwards).

EDIT: if you are looking for realistic simulated environment with a car, take a look at https://microsoft.github.io/AirSim/ simulator.
https://www.youtube.com/watch?v=gnz1X3UNM5Y

EDIT 2: just saw Carla simulator.
https://www.youtube.com/watch?v=7jej46ALVRE

cheers,
Mathieu