Why use Butterworth filter after MediaPipe smoothing (One-Euro) for pose landmarks?

[28.Lê Trung Hiếu]

I'm working on human pose estimation using **MediaPipe** (Pose or Face Mesh), which applies a **LandmarkSmoothing** filter (One-Euro style) to stabilize landmarks in real-time.

However, for **offline post-processing**, I see many studies still apply a **Butterworth filter** (low-pass or band-pass) to the landmark data before further analysis (like computing velocity, acceleration, spectral analysis of tremors, etc.).

---

### What I understand / have done so far

- MediaPipe smoothing (One-Euro) trades a bit of lag for real-time jitter reduction.
- Butterworth filters allow design with clear frequency cutoff and order, and—when used with filtfilt—they can achieve **zero-phase distortion**.

---

### My question

1. **Why do researchers often apply a Butterworth filter even after MediaPipe's smoothing?**  
   Is it primarily for frequency-selective filtering, zero-phase correction, or something else?

2. Does MediaPipe’s smoothing approach **lack control over frequency response**, making it unsuitable for signal analysis purposes?

---

### Example snippet (Python + SciPy)
```python
from scipy.signal import butter, filtfilt

def butter_lowpass(data, fs, cutoff=6.0, order=2):
    b, a = butter(order, cutoff / (0.5 * fs), btype='low')
    return filtfilt(b, a, data)

# Example: apply to MediaPipe landmark time series

  I would greatly appreciate insights from those working with real-time pipelines or conducting research in biomechanics/signal processing. Thank you!