Feature requests for next major release

[laxnpander]

This issue poses as collection of possible feature requests for a new major release. There is some topics that I have in mind, feel free to post your opinion on that (or new ones).

High Priority

• Improve Blending

Blending right now is limited to picking out points that were observed with the best (90°) inclination. However, exposure and lighting changes visually reduce the quality of the global map significantly. With multi-feather blending or similar techniques this could be improved. A good implementation must keep performance in mind, which might be challenging.

• Improve Robustness

There is a lot of metrics to check if our current tracking is good or not. Due to the alignment to a global frame we can easily reset the visual SLAM and just start from zero with out losing the global integrity of the map completely. There will be some local effects due to different georeferences, but overall this is a huge chance to achieve very robust mapping when the pipeline detects faulty poses and re-initializes automatically. Similar things have already been requested in #39

• Integration of Visual-inertial SLAM

Another way to improve robustness of the mapping process is to integrate and test visual-inertial SLAM frameworks. This comes with higher hardware requirements and preparation, but might allow tracking even in featureless regions for a while. We are already investigating this, but there are no results yet.

• Documentation

I do realize the documentation of the code and usability of the framework for custom datasets is mediocre at best. I'd like to improve this in the future with tools or guides on how to make sure your custom dataset is properly designed and processed.

Medium Priority

• Integration of different SLAM

Unofficial interfaces outside the repo for all major SLAM have already been written. Though, right now OpenVSLAM seems to be the most reliable candidate. ov2slam achieves higher accuracies though and might be the way to go in the future.

• Integration of different 3D reconstruction

Tests have shown that PSL works suprisingly well with accurate poses from ov2slam for example. Other frameworks have been implemented but none came close to the robustness of PSL yet. It seems to fit the use case very well in terms of observing largely flat areas with small local deviations.