Evaluation metrics scaling

[bhack]

From the paper:

To evaluate our approach, we use four commonly used
metrics: Sum of Absolute Differences (SAD), Mean Square
Error (MSE), Gradient loss (Grad), and Connectivity loss
(Conn). Lower values of these metrics indicate higher qual-
ity of alpha mattes. Note that we scale the value of MSE to
1e-3 for ease of reading.

But in the implementation I see instead only the SAD scaled

ViTMatte/evaluation.py

Line 28 in a70fcb3

return loss / 1000, np.sum(trimap == 128) / 1000