Skip to content

Latest commit

 

History

History
66 lines (48 loc) · 3.29 KB

README.md

File metadata and controls

66 lines (48 loc) · 3.29 KB

License CC BY-NC-SA 4.0 Python 2.7

MapNet: Geometry-Aware Learning of Maps for Camera Localization

License

Copyright (C) 2018 NVIDIA Corporation. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode).

Hyperparameters

Most parameters in the config files are self-explanatory. Here are some notes:

  • beta is the initial weight of the absolute rotation. In case of geometric weighting, it is the initial value of Beta in Eq. 3 of the paper. In case of uncertainty weighting, it is used as log(sigma) of the absolute rotation.

  • beta_translation is the initial weight of the absolute translation. It's default value is zero. In case of geometric weighting, it is the initial value of the translation weight similar to Beta for rotation sin Eq. 3 of the paper. In case of uncertainty weighting, it is used as log(sigma) of the absolute translation.

  • gamma is the initial weight of the relative rotation. In case of geometric weighting, is the initial value of Gamma in Eq. 3 of the paper. In case of uncertainty weighting, it is used as log(sigma) of the relative rotation.

  • gamma_translation is the initial weight of the relative translation. It's default value is zero. In case of geometric weighting, is the initial value of the translation weight similar to Gamma for rotation in Eq. 3 of the paper. In case of uncertainty weighting, it is used as log(sigma) of the relative translation.

  • steps is the size of tuples of images to use for training (parameter s described in Section 3.5 of the paper)

  • skip is the spacing between these images (parameter k described in Section 3.5 of the paper)

  • real is a flag indicating whether the poses should be GPS/SLAM/integration of visual odometry (true) or from ground truth (false)

  • color_jitter is the intensity of color jittering (brightness, hue, contrast and saturation) data augmentation. NOTE: Set color_jitter = 0 in mapnet.ini while training it on the 7 Scenes dataset.

  • s_abs_trans, s_abs_rot, s_rel_trans, s_rel_rot are the covariance values for absolute and relative translations and rotations passed to the PGO algorithm (see Appendix in our arXiv paper). To reproduce results from our paper, use the following values:

7 Scenes:

Scene s_abs_trans s_abs_rot s_rel_trans s_rel_rot
chess 1 1 35 35
fire 1 1 10 10
heads 1 1 1 1
office 0.1 0.1 20 10
pumpkin 1 1 500 500
redkitchen 1 1 35 35
stairs 1 1 2 4

RobotCar

Scene s_abs_trans s_abs_rot s_rel_trans s_rel_rot
loop 1 1 20 20
full 1 1 1 10

Which files to use?

Training

Use the files according to their name e.g. use mapnet++_7Scenes.ini if you want to train a MapNet++ model on the 7 Scenes dataset.

Inference

If you want to perform pose-graph optimization (PGO) during inference on any of the trained models, use the pgo_inference_*.ini files. The inference of a MapNet++ model (without PGO) should be done with mapnet.ini.