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).
Follow the style used in dataset_loaders/robotcar.py and
dataset_loaders/seven_scenes.py. They handle the reading of dataset-specific images and camera poses.
Camera poses are given to utils.process_poses() for MapNet-specific pre-processing. utils.process_poses()
accepts poses in the for form of a row-major flattned 12 element vector formed from the first 3 rows of
the world-to-camera 4x4 matrix.
Importantly, the dataloader will need to
define the self.gt_idx attribute for use with the higher-level datasets in
dataset_loaders/composite.py.
The gt_idx attribute is used when real == True
(i.e. some SLAM/integration of VO is used to get poses for the images) and the
VO library skips some images. In this case, length of the dataset will be
smaller than if real == False (i.e. using the dataset-provided ground truth
poses). gt_idx is a list whose i-th element is the index of the i-th image
if it were part of an instance of this class with real == False.