What improvements can we make on the basis of badslam？

[zhulz]

Badslam is a very good method based on surfel that I have seen. After learning, I need to innovate again to publish a paper. Does anyone have any suggestions?

[puzzlepaint]

Here are just a few random suggestions ( not necessarily good ones ;) ):

• Improve the formulation of the photometric residuals somehow (which are not very robust).
• Formulate the optimization residuals differently, such that for every keyframe, a prediction for the color and depth of the pixels is made and compared to the actual color and depth of the pixel (rather than having one optimization residual per surfel). This way, the residuals would directly correspond to the pixel measurements. This might among other things facilitate some level-of-detail handling: For example, if a detailed surface structure is observed by a keyframe from far away, then the comparison between the model and the keyframe should take into account that the observation of this structure in the keyframe will not see the close-up detail but instead a lower-resolution version (possibly also depending on the depth sensing technology used).
• Integrate depth estimation based on the color images in order to reduce the reliance on the depth camera and make the method work outdoors (where many depth cameras don't work) or if the observed surfaces are outside of the depth camera's range.
• Combine the method with techniques such as local / submap based bundle adjustment to make it work for large datasets.
• Combine the method with a real-time surface reconstruction approach to obtain real-time triangle mesh reconstructions.