# Improved data preprocessing

| Idea | Implemented | Tested |
| -------- | -------- | -------- |
| Interpolate frames | Yes | ? |
| Fix missing keypoints | Yes | ? |
| Fix shoulder rotation | Yes | No |
| Unified shoulder to shoulder length | No | No |

### Interpolate frames
Better ways to do this?

### Fix missing keypoints
Better ways to do this? -> mirror hand and place on wrist?

### Fix shoulder rotation
How much variance do we see in the shoulder rotations?

Medians are pretty similar but often outliers are detected aswell, it might be a good idea to fix shoulder rotation in general

After rotation, all angles of first frames for all sequences have been corrected and have been rotated. By doing this, all persons are sitting 'straight' and noise by different positions can be reduced.

### Unified shoulder to shoulder length
How much variance do we see in the shoulder to shoulder length?
We notice quite a lot of variation in shoulder to shoulder distances. We know that different signers have signed and not everybody is the same size (duh).
An idea could be to rescale the shoulder to shoulder distance of everybody to 1 such that on all sequences, the same scale is used.

We also notice that there is a lot of variation in shoulder to shoulder lenght per signer. This can be explained by different camera positions. A camera further away might give a smaller shoulder to shoulder distance than a camera closer for the same person.

Unifying the shoulder lenght might solve this problem and bring every sample of one person to the same scale.

After preprocessing