Doubt regarding soft wing loss

[ajkailash]

I am training a deep learning model for landmark detection (in-house dataset). The ground truth landmark coordinates are normalised (range (-1, 1)). When I take delta between ground truth and model predictions for a batch of tensors (first epoch), all of the values are below the softwing loss parameter (w1 = 2). Hence the soft wing loss, becomes a L1 loss. What parameters can I use for w1, w2 and epsilon?