search

yiakwy / SEMANTIC_VISUAL_SUPPORTED_ODEMETRY

semantic visual slam for monocular and stereo camera devices
Apache License 2.0
246 stars 55 forks source link

Discuss Semantic visual supported odometry #1

Open AlejandroSilvestri opened 4 years ago

AlejandroSilvestri commented 4 years ago

Hi @yiakwy, good job!

It's not an easy task to build a new VO. I can see you are blending deep learning, one step toward Spatial AI (this is a term I read Davidson is installing to name the new visual slam generation with deep learning).

I opened this so called issue, inviting all those who want to join and discuss about this novel VO.

I believe Python is the best choice to deal with deep learning.

yiakwy commented 4 years ago

Thank you @AlejandroSilvestri for your support! I feel to be energised to push things forward. Any suggestions are welcome, the project itself is under evolving, I need more help.

AlejandroSilvestri commented 4 years ago

@yiakwy

I saw many visual slam projects, and it seems to me those with good installation instructions were forked and cited more often than other. It is of great help if you always indicate if this is a work in progress, if the actual version is usable, and what to expect from it. A YouTube video always helps.

yiakwy commented 4 years ago

@AlejandroSilvestri Sorry for the delay of reply. I am currently working on migrating cpp modules and plan to push them to the repository the next month. I understand that the installation is not very neat and I need to improve them.

As suggested by @ausu0917, issue #2 gives me a good hint on strategies improving portability of the system by using docker. I am doing it now.

I also figure out a plausible road map for the next steps.

Proposed Roadmap

1th Priority

Make a dockerfile and improve the installation system. How do you @AlejandroSilvestri think about it?

2nd Priority

Migrating cpp version of the system (I am currently working on it) and replace different models to make the job with a complete report (either good or bad).

Finally, Transplanting to Semi-dense or Dense Projects

We have already observed that semantic features help in stability of matching, recognition and so on (though accuracy are affected by different factors such camera capturing speed, lighting ..., while people who working on ISP are helping us to get rid of them).

Instead of using axis aligned bounding box, maybe more points produce better structure of objects (?). And better structure of points and matching gives us better results though PNP solvers.

This means we can extend the work.

Finally

I have learned too much from ORBSlam. I would like to take this opportunity to invite you help me improving the schedule and keep work in order.

AlejandroSilvestri commented 4 years ago

Yes! I believe containers are perfect encapsulation for this type of work demanding a lot of dependencies. Docker is perfect for this.