Regarding the method of tracking points, the problem I encountered was that I couldn’t fully track the movement of the upper face point, so I had to manually pull back to the correct position in many places. This increased the workload considerably.
Regarding the denois and samrtvector approach, the problem I encountered was that the face had undergone a large deformation and I was unable to restore it to its original position, thus causing the tracking to fail.
Using denoise first, and note that it is better to choose the blackest part of the analysis point, so that the most information effect comes out. Export exr to keep more information.
Then use smartvectorb to export the noise-reduced video to exr (you can export directly without changing the channel settings)