viso2_ros stereo odometry performance

edit

I am using your ROS wrapper for libviso2. I made several tests and I am not really satisfied with the results. I recorded a video that is a quick indicator of the performance I am experiencing. Most of the relevant data is in the description below the video. Also in vidoe description there are results of the stereo_odometer in form of path trajectories in 3D space.

Currently, I am trying to decide "who to blame" for these results:

1. Myself, bad parameters, camera calibration, stereo calibration
2. The stereo camera I am using, I feel the left camera is slightly blurred
3. LibViso algorithm and its insufficient capabilities, bad scene, bad lighting, lack of distant points...

Please, share your opinion. It would mean a great deal to me.

EDIT:

The resolution is (2*1280) x 720. The video is recorded at 60 fps. When I publish the video frames to ROS, I publish them at 1Hz. Although, I am not publishing every frame but every fifth frame instead.

As I said, a lot of extra info is in video description on YouTube.

Also, I added some keyframes with epipolar lines draw on top of them. The name of the image indicates which keypoint produced the epiline.

https://dl.dropboxusercontent.com/u/5... https://dl.dropboxusercontent.com/u/5... https://dl.dropboxusercontent.com/u/5... https://dl.dropboxusercontent.com/u/5... https://dl.dropboxusercontent.com/u/5... https://dl.dropboxusercontent.com/u/5... https://dl.dropboxusercontent.com/u/5... https://dl.dropboxusercontent.com/u/5...