is there optimizing after loop_close?

[csufangyu]

thanks very much for your work!I am very interested in your work!I am reading your code,and I can't understand your code in loop_close,and I can find the functions :
checkNearbyLoops( keyframes, currFrame, &poses, &constraints ); checkRandomLoops( keyframes, currFrame, &poses, &constraints ); but i can't find the results.and why not optimize after loop_close??

[yilei0620]

Thanks for your suggestion. I should let it do global optimization after loop closed instead of local optimization.

Originally, I wrote the code for no DeepLab version which is for real time purpose. Thus, it is doing locally optimization. However, when I moved forward the project and found that it can't be a real-time slam because DeepLab costs too much time. But the baseline is still in the style for real time based slam.

It is still an experimental release.I will try to figure out the best baseline in the future since recently I'm trying to test some real time semantic segmentation methods. And for learning based slam, perhaps you ll be interested in this paper, http://campar.in.tum.de/Chair/ProjectCNNSLAM

[csufangyu]

thanks very much!Are you from china?I find chinese in your code!

[feiran-l]

Thanks for your work! I have tested the similar methods with PSPNet and ICNet, also I added some filters for better result. Please find the code in the link for your interests. https://github.com/SILI1994/static-SLAM-based-on-PSPnet

[yilei0620]

@csufangyu yes, I'm from China.

@SILI1994 I saw your result, it is much better. I think you didn't try to train neural networks so that, based on my knowledge, you used PSPNet not ICNet for segmentation, right?

In the future, I think it perhaps is a good idea that we should train a specific NN model which is just to remove pixels of people, cars and some common animals like cats, dogs. Right now those semantic segmentation models are too heavy because they are able to segment too many categories.

Thanks for your attention on my work.