@GoNikky thank you, and very valid question.
I checked it out quickly and understood what happened.
First, the algorithm was initially trying to determine cuts in the video to reinitialize the tracker state, but that was counterproductive and generated some signal loss. So I removed it.
Second, in the doggy part, the butt size compared to the frame was beyond a threshold triggering some specific conditions, valid in VR (close up scene), but not optimized for 2D POV, so I tweaked it for in case of 2D POV.
Based on those changes, the algo retrieved way more info to use.
Here is the before (red), after (blue):
And compared to the human-made script (red), the generated one is in blue (currently investigating what happened in section 4):
