Issues in feature extractions in OpenVSLAM

[surfii3z]

Hi all

I have been working on Visual SLAM for quite a while and I have the idea to use Neural Network, packnet-sfm in particular, to infer the depth and use the SLAM as RGB-D system instead of just monocular system.

By doing so, I hope to solve a few challenges of the monocular SLAM system, namely scale ambiguity, pure-rotation and map initialization.

I pose this Youtube link to show some result of my experiment on KITTI odometry sequence 07. I found that OpenVSLAM has some issues.

In the video, OpenVSLAM has a lot of features flickering on the road, in contrast to ORB_SLAM2. After that, it also failed to track the features after the car was turning, whereas ORB_SLAM2 works just fine.

I wonder if we cloud identify/ solve these problems of OpenVSLAM.

Best

Surf

[ymd-stella]

It's related to #73.