Documentation - Energy Function, Selected Points and BA

[jhidalgocarrio]

I am missing some documentation on Extend DSOPP about how the Energy function is implemented. I think documentation about three important points would be helpful:

• The front-end on how and where the image-to-image alignment is implemented: minimizing the photometric error from each new frame to the current (last) keyframe. Which classes are involved in the process?
• Immature points and how the implementation relates to the original DSO. Which classes are involved in this process? Where in the code, the immature points are selected to become mature points for the optimization step.
• I also miss some documentation about the Bundle Adjustment (BA). The selected keyframes to minimize the photometric error in the last N-frames. Same here, which classes are computing the BA.

[Solonets]

Hi @jhidalgocarrio. Thank you for your question. Please check this description. Does this answer your question?

[jhidalgocarrio]

Hi @Solonets, This is great! Thanks a lot for the documentation. Yes, the documentation is handy. I am just missing the keyframe selection strategy; which classes perform such operation?

Could you kindly add a few explanatory lines about the difference between the following classes?

• Frame
• TrackingFrame
• Keyframe
• ActiveKeyframe

[Solonets]

Hi @jhidalgocarrio ,

I am not particularly sure about adding those details to the documentation. You could see the class diagram and description of the classes by compiling code documentation with Doxygen. I think I just answer your question here:

Keyframe selection strategy is performed with KeyframeStrategy interface. There are 2 implementations, you could add yours as well.

About Frames:

Frame is a base class for all frames and just carries common functionality. ActiveKeyframe is a track's keyframes. It's position can be optimized and it carries points attached to it. SlamIntrentalTrackingFrame is a frame that is not a keyframe. It is always attached to a keyframe and only stores its relative position to the keyframe.

Keyframe and TrackingFrame are the siblings of the 2 above classes. The only difference is that it doesn't carry additional information needed for a slam. They are used for storing and loading the track as well as manipulating it outside of the context of SLAM.

[jhidalgocarrio]

As you prefer. I still think a class diagram is a description of all the classes. While this doc is only about the key parts of DSO in order to quickly grasp the core of DSO in your implementation.

Just adding the text from your reply would help:

Frame is a base class for all frames and just carries common functionality.
ActiveKeyframe is a track's keyframes. Its position can be optimized, and it carries points attached to it.
SlamIntrentalTrackingFrame is a frame that is not a keyframe. It is always attached to a keyframe and only stores its relative position to the keyframe.

Keyframe and TrackingFrame are the siblings of the 2 above classes. The only difference is that it doesn't carry additional information needed for a slam. They are used for storing and loading the track as well as manipulating it outside of the context of SLAM.