The precision of the optical flow method decreases
with the speed of motion. We then expect that fast
movements will have an impact on the estimation of
noise variance, especially on the edges of a moving object. As a result, we have to suppose that movements
are moderate. Variations in illumination (e.g., a moving
shadow) also cause errors in optical flow computations,
and therefore in our approach.
All of the experiments confirm that our algorithm is
relatively resistant to parameter changes and that it can
also detect altered images even when they have been
compressed or when noise was added to them (up to a
certain point). Interested readers may find more details
about both the influence of the parameters and the robustness to various compression level in [9].